diff --git a/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/README.md b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/README.md new file mode 100644 index 0000000000..f8c89c20f6 --- /dev/null +++ b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/README.md @@ -0,0 +1,39 @@ +# PR #1855 base + AWQ-lite mixed-precision GPTQ — val_bpb 1.06086 (3-seed mean) + +Applies activation-aware mixed-precision GPTQ (from PR #1908 / romeerp) on top of codemath3000 PR #1855 stack. + +## Results + +| Seed | val_bpb (post-TTT) | artifact bytes | steps | eval time | +|------|--------------------|----------------|-------|-----------| +| 42 | 1.06118 | 15,978,503 | 4989 | 392.8s | +| 314 | 1.06005 | 15,976,469 | 4986 | 395.8s | +| 1234 | 1.06135 | 15,976,673 | 4977 | 395.5s | +| **mean** | **1.06086** | — | — | — | + +3-seed std: 0.00069. Beats codemath3000 PR #1855 (1.06108) by 0.00022 BPB. + +## Technique + +Training is identical to PR #1855. The only change is post-training quantization: + +**AWQ-lite (activation-aware GPTQ):** +1. Collect per-input-channel activation RMS during GPTQ calibration +2. Score column groups: `saliency = act_rms * mean(abs(weight))` +3. Select top-1 most salient 64-column group per matrix +4. Quantize that group at int8 inside the same full-tensor GPTQ solve (rest stays int6) + +Env vars: `AWQ_LITE_ENABLED=1 AWQ_LITE_BITS=8 AWQ_LITE_GROUP_TOP_K=1 AWQ_LITE_GROUP_SIZE=64` + +## Setup +1. `pip install -r requirements.txt` +2. `apt-get install -y lrzip` +3. Install FA3: `pip install --no-deps flash_attn_3 --find-links https://windreamer.github.io/flash-attention3-wheels/cu128_torch291/` +4. Run `prepare_caseops_data.py` to build the dataset +5. `AWQ_LITE_ENABLED=1 AWQ_LITE_BITS=8 AWQ_LITE_GROUP_TOP_K=1 AWQ_LITE_GROUP_SIZE=64 torchrun --standalone --nproc_per_node=8 train_gpt.py` + +## Environment +- 8xH100 80GB SXM (RunPod) +- PyTorch 2.9.1+cu128 +- FlashAttention 3.0.0 +- Triton 3.5.1 diff --git a/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/lossless_caps.py b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/lossless_caps.py new file mode 100644 index 0000000000..98e472f824 --- /dev/null +++ b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/lossless_caps.py @@ -0,0 +1,833 @@ +"""Lossless capitalization pre-encoding helpers. + +This module provides a narrow, reversible transform that only touches +ASCII capital letters `A-Z`. Each uppercase ASCII letter is rewritten as +``, where `sentinel` is a private-use Unicode +character that is escaped by doubling if it appears literally in the +input text. + +Example with the default sentinel `\\uE000`: + + "The NASA Launch" -> "\\uE000the \\uE000n\\uE000a\\uE000s\\uE000a \\uE000launch" + +The transform is intentionally simple for v1: + +- lowercase ASCII letters are unchanged +- uppercase ASCII letters become sentinel + lowercase letter +- non-ASCII characters are left untouched +- literal sentinel characters are escaped as sentinel + sentinel + +This makes the transform exactly invertible while allowing a downstream +tokenizer to reuse lowercase subwords across case variants. +""" + +from __future__ import annotations + +import json +from pathlib import Path +from typing import Callable, Iterable + +LOSSLESS_CAPS_V1 = "lossless_caps_v1" +LOSSLESS_CAPS_V2 = "lossless_caps_v2" +LOSSLESS_CAPS_V3 = "lossless_caps_v3" +LOSSLESS_CAPS_V4 = "lossless_caps_v4" +LOSSLESS_CAPS_V5 = "lossless_caps_v5" +LOSSLESS_CAPS_V6 = "lossless_caps_v6" +LOSSLESS_CAPS_V7 = "lossless_caps_v7" +LOSSLESS_CAPS_CASEOPS_V1 = "lossless_caps_caseops_v1" +IDENTITY = "identity" +DEFAULT_SENTINEL = "\uE000" +DEFAULT_V2_TITLE = "\uE001" +DEFAULT_V2_ALLCAPS = "\uE002" +DEFAULT_V2_CAPNEXT = "\uE003" +DEFAULT_V2_ESC = "\uE004" +DEFAULT_V5_TITLE_MIN_LEN = 7 +DEFAULT_V6_ALLCAPS_MIN_LEN = 3 +DEFAULT_V7_ALLCAPS_MIN_LEN = 4 + + +class LosslessCapsError(ValueError): + """Raised when a transformed string is malformed.""" + + +def _is_ascii_upper(ch: str) -> bool: + return "A" <= ch <= "Z" + + +def _is_ascii_lower(ch: str) -> bool: + return "a" <= ch <= "z" + + +def _is_ascii_alpha(ch: str) -> bool: + return _is_ascii_lower(ch) or _is_ascii_upper(ch) + + +def _validate_distinct_single_chars(*chars: str) -> None: + if any(len(ch) != 1 for ch in chars): + raise ValueError("all control characters must be exactly one character") + if len(set(chars)) != len(chars): + raise ValueError("control characters must be distinct") + + +def encode_lossless_caps_v1(text: str, *, sentinel: str = DEFAULT_SENTINEL) -> str: + """Encode ASCII capitals reversibly using a one-character sentinel.""" + if len(sentinel) != 1: + raise ValueError("sentinel must be exactly one character") + out: list[str] = [] + for ch in text: + if ch == sentinel: + out.append(sentinel) + out.append(sentinel) + elif _is_ascii_upper(ch): + out.append(sentinel) + out.append(ch.lower()) + else: + out.append(ch) + return "".join(out) + + +def decode_lossless_caps_v1(text: str, *, sentinel: str = DEFAULT_SENTINEL) -> str: + """Decode the `lossless_caps_v1` transform back to the original text.""" + if len(sentinel) != 1: + raise ValueError("sentinel must be exactly one character") + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch != sentinel: + out.append(ch) + i += 1 + continue + if i + 1 >= n: + raise LosslessCapsError("dangling capitalization sentinel at end of string") + nxt = text[i + 1] + if nxt == sentinel: + out.append(sentinel) + elif _is_ascii_lower(nxt): + out.append(nxt.upper()) + else: + raise LosslessCapsError( + f"invalid sentinel escape sequence {sentinel + nxt!r}; " + "expected doubled sentinel or sentinel + lowercase ASCII letter" + ) + i += 2 + return "".join(out) + + +def encode_lossless_caps_v2( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + capnext: str = DEFAULT_V2_CAPNEXT, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Encode ASCII word capitalization with cheap word-level markers. + + Rules over maximal ASCII alphabetic runs: + - lowercase words stay unchanged + - TitleCase words become `title + lowercase(word)` + - ALLCAPS words become `allcaps + lowercase(word)` + - mixed-case words use: + - optional `title` when the first letter is uppercase + - `capnext + lowercase(letter)` for subsequent uppercase letters + - literal control characters are escaped as `esc + literal` + """ + _validate_distinct_single_chars(title, allcaps, capnext, esc) + controls = {title, allcaps, capnext, esc} + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch in controls: + out.append(esc) + out.append(ch) + i += 1 + continue + if not _is_ascii_alpha(ch): + out.append(ch) + i += 1 + continue + + j = i + 1 + while j < n and _is_ascii_alpha(text[j]): + j += 1 + word = text[i:j] + lower_word = word.lower() + + if word.islower(): + out.append(word) + elif len(word) >= 2 and word.isupper(): + out.append(allcaps) + out.append(lower_word) + elif _is_ascii_upper(word[0]) and word[1:].islower(): + out.append(title) + out.append(lower_word) + else: + if _is_ascii_upper(word[0]): + out.append(title) + out.append(lower_word[0]) + for orig_ch, lower_ch in zip(word[1:], lower_word[1:], strict=True): + if _is_ascii_upper(orig_ch): + out.append(capnext) + out.append(lower_ch) + i = j + return "".join(out) + + +def decode_lossless_caps_v2( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + capnext: str = DEFAULT_V2_CAPNEXT, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v2` transform back to the original text.""" + _validate_distinct_single_chars(title, allcaps, capnext, esc) + out: list[str] = [] + pending_escape = False + pending_word_mode: str | None = None + active_allcaps = False + pending_capnext = False + in_ascii_word = False + + for ch in text: + if pending_escape: + if pending_word_mode is not None and not _is_ascii_alpha(ch): + raise LosslessCapsError("escaped control char cannot satisfy pending word capitalization mode") + out.append(ch) + pending_escape = False + if _is_ascii_alpha(ch): + in_ascii_word = True + else: + in_ascii_word = False + active_allcaps = False + continue + + if ch == esc: + pending_escape = True + continue + if ch == title: + if pending_word_mode is not None or in_ascii_word or pending_capnext: + raise LosslessCapsError("invalid title marker placement") + pending_word_mode = "title" + continue + if ch == allcaps: + if pending_word_mode is not None or in_ascii_word or pending_capnext: + raise LosslessCapsError("invalid allcaps marker placement") + pending_word_mode = "allcaps" + continue + if ch == capnext: + if pending_capnext: + raise LosslessCapsError("duplicate capnext marker") + pending_capnext = True + continue + + if _is_ascii_alpha(ch): + at_word_start = not in_ascii_word + if at_word_start: + if pending_word_mode == "allcaps": + out.append(ch.upper()) + active_allcaps = True + elif pending_word_mode == "title": + out.append(ch.upper()) + elif pending_capnext: + out.append(ch.upper()) + else: + out.append(ch) + pending_word_mode = None + pending_capnext = False + in_ascii_word = True + continue + + if pending_word_mode is not None: + raise LosslessCapsError("word capitalization marker leaked into the middle of a word") + if active_allcaps: + out.append(ch.upper()) + elif pending_capnext: + out.append(ch.upper()) + else: + out.append(ch) + pending_capnext = False + continue + + if pending_word_mode is not None or pending_capnext: + raise LosslessCapsError("capitalization marker not followed by an ASCII letter") + out.append(ch) + in_ascii_word = False + active_allcaps = False + + if pending_escape: + raise LosslessCapsError("dangling escape marker at end of string") + if pending_word_mode is not None or pending_capnext: + raise LosslessCapsError("dangling capitalization marker at end of string") + return "".join(out) + + +def encode_lossless_caps_v3( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Encode only common word-level capitalization patterns. + + Rules over maximal ASCII alphabetic runs: + - lowercase words stay unchanged + - TitleCase words become `title + lowercase(word)` + - ALLCAPS words become `allcaps + lowercase(word)` + - all other mixed-case words are left unchanged + - literal control characters are escaped as `esc + literal` + """ + _validate_distinct_single_chars(title, allcaps, esc) + controls = {title, allcaps, esc} + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch in controls: + out.append(esc) + out.append(ch) + i += 1 + continue + if not _is_ascii_alpha(ch): + out.append(ch) + i += 1 + continue + + j = i + 1 + while j < n and _is_ascii_alpha(text[j]): + j += 1 + word = text[i:j] + + if word.islower(): + out.append(word) + elif len(word) >= 2 and word.isupper(): + out.append(allcaps) + out.append(word.lower()) + elif _is_ascii_upper(word[0]) and word[1:].islower(): + out.append(title) + out.append(word.lower()) + else: + out.append(word) + i = j + return "".join(out) + + +def decode_lossless_caps_v3( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v3` transform back to the original text.""" + _validate_distinct_single_chars(title, allcaps, esc) + out: list[str] = [] + pending_escape = False + pending_word_mode: str | None = None + active_allcaps = False + in_ascii_word = False + + for ch in text: + if pending_escape: + if pending_word_mode is not None and not _is_ascii_alpha(ch): + raise LosslessCapsError("escaped control char cannot satisfy pending word capitalization mode") + out.append(ch) + pending_escape = False + if _is_ascii_alpha(ch): + in_ascii_word = True + else: + in_ascii_word = False + active_allcaps = False + continue + + if ch == esc: + pending_escape = True + continue + if ch == title: + if pending_word_mode is not None or in_ascii_word: + raise LosslessCapsError("invalid title marker placement") + pending_word_mode = "title" + continue + if ch == allcaps: + if pending_word_mode is not None or in_ascii_word: + raise LosslessCapsError("invalid allcaps marker placement") + pending_word_mode = "allcaps" + continue + + if _is_ascii_alpha(ch): + at_word_start = not in_ascii_word + if at_word_start: + if pending_word_mode == "allcaps": + out.append(ch.upper()) + active_allcaps = True + elif pending_word_mode == "title": + out.append(ch.upper()) + else: + out.append(ch) + pending_word_mode = None + in_ascii_word = True + continue + + if pending_word_mode is not None: + raise LosslessCapsError("word capitalization marker leaked into the middle of a word") + out.append(ch.upper() if active_allcaps else ch) + continue + + if pending_word_mode is not None: + raise LosslessCapsError("capitalization marker not followed by an ASCII letter") + out.append(ch) + in_ascii_word = False + active_allcaps = False + + if pending_escape: + raise LosslessCapsError("dangling escape marker at end of string") + if pending_word_mode is not None: + raise LosslessCapsError("dangling capitalization marker at end of string") + return "".join(out) + + +def encode_lossless_caps_v4( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Encode only ALLCAPS ASCII words, leaving all other case untouched.""" + _validate_distinct_single_chars(allcaps, esc) + controls = {allcaps, esc} + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch in controls: + out.append(esc) + out.append(ch) + i += 1 + continue + if not _is_ascii_alpha(ch): + out.append(ch) + i += 1 + continue + j = i + 1 + while j < n and _is_ascii_alpha(text[j]): + j += 1 + word = text[i:j] + if len(word) >= 2 and word.isupper(): + out.append(allcaps) + out.append(word.lower()) + else: + out.append(word) + i = j + return "".join(out) + + +def decode_lossless_caps_v4( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v4` transform back to the original text.""" + _validate_distinct_single_chars(allcaps, esc) + out: list[str] = [] + pending_escape = False + pending_allcaps = False + in_ascii_word = False + active_allcaps = False + + for ch in text: + if pending_escape: + if pending_allcaps and not _is_ascii_alpha(ch): + raise LosslessCapsError("escaped control char cannot satisfy pending allcaps mode") + out.append(ch) + pending_escape = False + if _is_ascii_alpha(ch): + in_ascii_word = True + else: + in_ascii_word = False + active_allcaps = False + continue + + if ch == esc: + pending_escape = True + continue + if ch == allcaps: + if pending_allcaps or in_ascii_word: + raise LosslessCapsError("invalid allcaps marker placement") + pending_allcaps = True + continue + + if _is_ascii_alpha(ch): + if not in_ascii_word: + active_allcaps = pending_allcaps + pending_allcaps = False + in_ascii_word = True + out.append(ch.upper() if active_allcaps else ch) + continue + + if pending_allcaps: + raise LosslessCapsError("allcaps marker not followed by an ASCII letter") + out.append(ch) + in_ascii_word = False + active_allcaps = False + + if pending_escape: + raise LosslessCapsError("dangling escape marker at end of string") + if pending_allcaps: + raise LosslessCapsError("dangling allcaps marker at end of string") + return "".join(out) + + +def encode_lossless_caps_v5( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, + title_min_len: int = DEFAULT_V5_TITLE_MIN_LEN, +) -> str: + """Encode ALLCAPS words and only sufficiently long TitleCase words.""" + _validate_distinct_single_chars(title, allcaps, esc) + controls = {title, allcaps, esc} + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch in controls: + out.append(esc) + out.append(ch) + i += 1 + continue + if not _is_ascii_alpha(ch): + out.append(ch) + i += 1 + continue + j = i + 1 + while j < n and _is_ascii_alpha(text[j]): + j += 1 + word = text[i:j] + if len(word) >= 2 and word.isupper(): + out.append(allcaps) + out.append(word.lower()) + elif len(word) >= title_min_len and _is_ascii_upper(word[0]) and word[1:].islower(): + out.append(title) + out.append(word.lower()) + else: + out.append(word) + i = j + return "".join(out) + + +def decode_lossless_caps_v5( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v5` transform back to the original text.""" + return decode_lossless_caps_v3(text, title=title, allcaps=allcaps, esc=esc) + + +def encode_lossless_caps_v6( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, + allcaps_min_len: int = DEFAULT_V6_ALLCAPS_MIN_LEN, +) -> str: + """Encode only ALLCAPS words with length >= allcaps_min_len.""" + _validate_distinct_single_chars(allcaps, esc) + controls = {allcaps, esc} + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch in controls: + out.append(esc) + out.append(ch) + i += 1 + continue + if not _is_ascii_alpha(ch): + out.append(ch) + i += 1 + continue + j = i + 1 + while j < n and _is_ascii_alpha(text[j]): + j += 1 + word = text[i:j] + if len(word) >= allcaps_min_len and word.isupper(): + out.append(allcaps) + out.append(word.lower()) + else: + out.append(word) + i = j + return "".join(out) + + +def decode_lossless_caps_v6( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v6` transform back to the original text.""" + return decode_lossless_caps_v4(text, allcaps=allcaps, esc=esc) + + +def encode_lossless_caps_v7( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, + allcaps_min_len: int = DEFAULT_V7_ALLCAPS_MIN_LEN, +) -> str: + """Encode only ALLCAPS words with length >= 4.""" + return encode_lossless_caps_v6( + text, + allcaps=allcaps, + esc=esc, + allcaps_min_len=allcaps_min_len, + ) + + +def decode_lossless_caps_v7( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v7` transform back to the original text.""" + return decode_lossless_caps_v6(text, allcaps=allcaps, esc=esc) + + +def get_text_transform(name: str | None) -> Callable[[str], str]: + """Return the forward text transform for the given config name.""" + normalized = IDENTITY if name in {None, "", IDENTITY} else str(name) + if normalized == IDENTITY: + return lambda text: text + if normalized == LOSSLESS_CAPS_V1: + return encode_lossless_caps_v1 + if normalized == LOSSLESS_CAPS_V2: + return encode_lossless_caps_v2 + if normalized == LOSSLESS_CAPS_V3: + return encode_lossless_caps_v3 + if normalized == LOSSLESS_CAPS_V4: + return encode_lossless_caps_v4 + if normalized == LOSSLESS_CAPS_V5: + return encode_lossless_caps_v5 + if normalized == LOSSLESS_CAPS_V6: + return encode_lossless_caps_v6 + if normalized == LOSSLESS_CAPS_V7: + return encode_lossless_caps_v7 + if normalized == LOSSLESS_CAPS_CASEOPS_V1: + return encode_lossless_caps_v2 + raise ValueError(f"unsupported text_transform={name!r}") + + +def get_text_inverse_transform(name: str | None) -> Callable[[str], str]: + """Return the inverse transform for the given config name.""" + normalized = IDENTITY if name in {None, "", IDENTITY} else str(name) + if normalized == IDENTITY: + return lambda text: text + if normalized == LOSSLESS_CAPS_V1: + return decode_lossless_caps_v1 + if normalized == LOSSLESS_CAPS_V2: + return decode_lossless_caps_v2 + if normalized == LOSSLESS_CAPS_V3: + return decode_lossless_caps_v3 + if normalized == LOSSLESS_CAPS_V4: + return decode_lossless_caps_v4 + if normalized == LOSSLESS_CAPS_V5: + return decode_lossless_caps_v5 + if normalized == LOSSLESS_CAPS_V6: + return decode_lossless_caps_v6 + if normalized == LOSSLESS_CAPS_V7: + return decode_lossless_caps_v7 + if normalized == LOSSLESS_CAPS_CASEOPS_V1: + return decode_lossless_caps_v2 + raise ValueError(f"unsupported text_transform={name!r}") + + +def normalize_text_transform_name(name: str | None) -> str: + """Normalize empty/None transform names to the identity transform.""" + return IDENTITY if name in {None, "", IDENTITY} else str(name) + + +def get_text_transform_control_symbols(name: str | None) -> list[str]: + """Return reserved control symbols used by a transform, if any.""" + normalized = normalize_text_transform_name(name) + if normalized == IDENTITY: + return [] + if normalized == LOSSLESS_CAPS_V1: + return [DEFAULT_SENTINEL] + if normalized == LOSSLESS_CAPS_V2: + return [DEFAULT_V2_TITLE, DEFAULT_V2_ALLCAPS, DEFAULT_V2_CAPNEXT, DEFAULT_V2_ESC] + if normalized == LOSSLESS_CAPS_CASEOPS_V1: + return [DEFAULT_V2_TITLE, DEFAULT_V2_ALLCAPS, DEFAULT_V2_CAPNEXT, DEFAULT_V2_ESC] + if normalized in {LOSSLESS_CAPS_V3, LOSSLESS_CAPS_V5}: + return [DEFAULT_V2_TITLE, DEFAULT_V2_ALLCAPS, DEFAULT_V2_ESC] + if normalized in {LOSSLESS_CAPS_V4, LOSSLESS_CAPS_V6, LOSSLESS_CAPS_V7}: + return [DEFAULT_V2_ALLCAPS, DEFAULT_V2_ESC] + raise ValueError(f"unsupported text_transform={name!r}") + + +def infer_text_transform_from_manifest(tokenizer_path: str | Path) -> str: + """Best-effort lookup of a tokenizer's text transform from a local manifest.""" + tokenizer_path = Path(tokenizer_path).expanduser().resolve() + manifest_candidates = [ + tokenizer_path.parent.parent / "manifest.json", + tokenizer_path.parent / "manifest.json", + ] + for manifest_path in manifest_candidates: + if not manifest_path.is_file(): + continue + try: + payload = json.loads(manifest_path.read_text(encoding="utf-8")) + except (OSError, json.JSONDecodeError): + continue + tokenizers = payload.get("tokenizers") + if not isinstance(tokenizers, list): + continue + for tokenizer_meta in tokenizers: + if not isinstance(tokenizer_meta, dict): + continue + model_path = tokenizer_meta.get("model_path") or tokenizer_meta.get("path") + if not model_path: + continue + candidate = (manifest_path.parent / str(model_path)).resolve() + if candidate == tokenizer_path: + return normalize_text_transform_name(tokenizer_meta.get("text_transform")) + return IDENTITY + + +def surface_piece_original_byte_counts( + surfaces: Iterable[str], + *, + text_transform_name: str | None = None, + sentinel: str = DEFAULT_SENTINEL, +) -> list[int]: + """Return exact original UTF-8 byte counts contributed by each surface piece. + + `surfaces` must be the exact decoded text fragments emitted by SentencePiece + in order, e.g. `piece.surface` from `encode_as_immutable_proto`. + """ + normalized = normalize_text_transform_name(text_transform_name) + if normalized == IDENTITY: + return [len(surface.encode("utf-8")) for surface in surfaces] + if normalized == LOSSLESS_CAPS_V1: + if len(sentinel) != 1: + raise ValueError("sentinel must be exactly one character") + sentinel_bytes = len(sentinel.encode("utf-8")) + pending_sentinel = False + counts: list[int] = [] + for surface in surfaces: + piece_bytes = 0 + for ch in surface: + if pending_sentinel: + if ch == sentinel: + piece_bytes += sentinel_bytes + elif _is_ascii_lower(ch): + piece_bytes += 1 + else: + raise LosslessCapsError( + f"invalid continuation {ch!r} after capitalization sentinel" + ) + pending_sentinel = False + continue + if ch == sentinel: + pending_sentinel = True + else: + piece_bytes += len(ch.encode("utf-8")) + counts.append(piece_bytes) + if pending_sentinel: + raise LosslessCapsError("dangling capitalization sentinel across piece boundary") + return counts + if normalized not in {LOSSLESS_CAPS_V2, LOSSLESS_CAPS_V3, LOSSLESS_CAPS_V4, LOSSLESS_CAPS_V5, LOSSLESS_CAPS_V6, LOSSLESS_CAPS_V7, LOSSLESS_CAPS_CASEOPS_V1}: + raise ValueError(f"unsupported text_transform={text_transform_name!r}") + + title = DEFAULT_V2_TITLE + allcaps = DEFAULT_V2_ALLCAPS + capnext = DEFAULT_V2_CAPNEXT + esc = DEFAULT_V2_ESC + if normalized in {LOSSLESS_CAPS_V2, LOSSLESS_CAPS_CASEOPS_V1}: + _validate_distinct_single_chars(title, allcaps, capnext, esc) + elif normalized in {LOSSLESS_CAPS_V4, LOSSLESS_CAPS_V6, LOSSLESS_CAPS_V7}: + _validate_distinct_single_chars(allcaps, esc) + else: + _validate_distinct_single_chars(title, allcaps, esc) + pending_escape = False + pending_word_mode: str | None = None + active_allcaps = False + pending_capnext = False + in_ascii_word = False + counts: list[int] = [] + for surface in surfaces: + piece_bytes = 0 + for ch in surface: + if pending_escape: + if pending_word_mode is not None and not _is_ascii_alpha(ch): + raise LosslessCapsError("escaped control char cannot satisfy pending word capitalization mode") + piece_bytes += len(ch.encode("utf-8")) + pending_escape = False + if _is_ascii_alpha(ch): + in_ascii_word = True + else: + in_ascii_word = False + active_allcaps = False + continue + if ch == esc: + pending_escape = True + continue + if normalized in {LOSSLESS_CAPS_V2, LOSSLESS_CAPS_V3, LOSSLESS_CAPS_V5, LOSSLESS_CAPS_CASEOPS_V1} and ch == title: + if pending_word_mode is not None or in_ascii_word or pending_capnext: + raise LosslessCapsError("invalid title marker placement") + pending_word_mode = "title" + continue + if ch == allcaps: + if pending_word_mode is not None or in_ascii_word or pending_capnext: + raise LosslessCapsError("invalid allcaps marker placement") + pending_word_mode = "allcaps" + continue + if normalized in {LOSSLESS_CAPS_V2, LOSSLESS_CAPS_CASEOPS_V1} and ch == capnext: + if pending_capnext: + raise LosslessCapsError("duplicate capnext marker") + pending_capnext = True + continue + + if _is_ascii_alpha(ch): + at_word_start = not in_ascii_word + if at_word_start: + piece_bytes += 1 + active_allcaps = pending_word_mode == "allcaps" + pending_word_mode = None + pending_capnext = False + in_ascii_word = True + continue + if pending_word_mode is not None: + raise LosslessCapsError("word capitalization marker leaked into the middle of a word") + piece_bytes += 1 + pending_capnext = False + continue + + if pending_word_mode is not None or pending_capnext: + raise LosslessCapsError("capitalization marker not followed by an ASCII letter") + piece_bytes += len(ch.encode("utf-8")) + in_ascii_word = False + active_allcaps = False + counts.append(piece_bytes) + if pending_escape: + raise LosslessCapsError("dangling escape marker across piece boundary") + if pending_word_mode is not None or pending_capnext: + raise LosslessCapsError("dangling capitalization marker across piece boundary") + return counts diff --git a/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/prepare_caseops_data.py b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/prepare_caseops_data.py new file mode 100644 index 0000000000..5c3f13e69c --- /dev/null +++ b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/prepare_caseops_data.py @@ -0,0 +1,177 @@ +"""Prepare CaseOps-tokenized FineWeb shards + per-token byte sidecar. + +CaseOps (``lossless_caps_caseops_v1``) is a bijective, character-level text +transform that introduces four operator tokens in place of explicit +capitalization: TITLE, ALLCAPS, CAPNEXT, ESC. The transform is fully +reversible — no information is lost relative to the untransformed UTF-8 +text, so BPB stays computable on TRUE byte counts. + +Forward pipeline: + 1. Read the canonical FineWeb-10B doc stream (``docs_selected.jsonl`` + produced by ``data/download_hf_docs_and_tokenize.py`` in the root repo). + 2. Apply ``encode_lossless_caps_v2`` (the caseops_v1 alias) to each doc. + 3. Tokenize with the shipped SP model + ``tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model`` + (reserves TITLE/ALLCAPS/CAPNEXT/ESC + sentinel as user_defined_symbols). + 4. Write uint16 train/val shards (``fineweb_{train,val}_XXXXXX.bin``). + 5. For the VAL stream only, emit per-token byte sidecar shards + (``fineweb_val_bytes_XXXXXX.bin``, uint16 parallel arrays) that record + each token's ORIGINAL pre-transform UTF-8 byte count. BPB is computed + from these canonical bytes so the score is on the untransformed text + (not the transformed representation). + +Output layout — matches what ``train_gpt.py`` expects under +``DATA_DIR=./data`` with ``CASEOPS_ENABLED=1``: + + data/datasets/fineweb10B_sp8192_caseops/datasets/ + tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model + datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/ + fineweb_train_000000.bin + fineweb_train_000001.bin + ... + fineweb_val_000000.bin + fineweb_val_bytes_000000.bin + +Usage: + + python3 prepare_caseops_data.py \\ + --docs ./fineweb10B_raw/docs_selected.jsonl \\ + --out ./data/datasets/fineweb10B_sp8192_caseops/datasets \\ + --sp ./tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model + +Requirements: sentencepiece, numpy. CPU-only. Runs once; reused across seeds. +""" +from __future__ import annotations + +import argparse +import json +import pathlib +import struct +import sys + +import numpy as np +import sentencepiece as spm + +# Local import — lossless_caps.py ships next to this script. +sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent)) +from lossless_caps import ( # noqa: E402 + LOSSLESS_CAPS_CASEOPS_V1, + encode_lossless_caps_v2, + surface_piece_original_byte_counts, +) + + +SHARD_MAGIC = 20240520 +SHARD_VERSION = 1 +SHARD_TOKENS = 10_000_000 # tokens per shard — matches the main pipeline +BOS_ID = 1 # SP model's control token; train_gpt.py:_find_docs requires BOS per doc + + +def _write_shard(out_path: pathlib.Path, arr: np.ndarray) -> None: + """Write a uint16 shard in the standard header-prefixed format.""" + assert arr.dtype == np.uint16 + header = np.zeros(256, dtype=np.int32) + header[0] = SHARD_MAGIC + header[1] = SHARD_VERSION + header[2] = int(arr.size) + with out_path.open("wb") as fh: + fh.write(header.tobytes()) + fh.write(arr.tobytes()) + + +def _iter_docs(docs_path: pathlib.Path): + """Yield doc strings from a jsonl file (one json object per line).""" + with docs_path.open("r", encoding="utf-8") as fh: + for line in fh: + line = line.strip() + if not line: + continue + obj = json.loads(line) + # Support both {"text": ...} and raw strings. + yield obj["text"] if isinstance(obj, dict) else obj + + +def _token_original_byte_counts( + sp: spm.SentencePieceProcessor, + original_text: str, + transformed_text: str, +) -> np.ndarray: + """Per-token canonical (pre-transform) UTF-8 byte counts. + + Delegates to ``surface_piece_original_byte_counts`` in ``lossless_caps.py`` + — the canonical exporter used by the PR #1729 / HF-hosted CaseOps dataset. + Operator pieces (U+E001..U+E004) contribute 0 original bytes; letter pieces + contribute their pre-transform UTF-8 byte count. + """ + proto = sp.encode_as_immutable_proto(transformed_text) + byte_counts = surface_piece_original_byte_counts( + (piece.surface for piece in proto.pieces), + text_transform_name=LOSSLESS_CAPS_CASEOPS_V1, + ) + return np.asarray(list(byte_counts), dtype=np.uint16) + + +def main() -> None: + ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) + ap.add_argument("--docs", required=True, type=pathlib.Path, help="Path to docs_selected.jsonl") + ap.add_argument("--out", required=True, type=pathlib.Path, help="Output datasets dir") + ap.add_argument("--sp", required=True, type=pathlib.Path, help="Path to CaseOps SP model") + ap.add_argument("--val-docs", type=int, default=10_000, help="Validation docs count") + args = ap.parse_args() + + sp = spm.SentencePieceProcessor(model_file=str(args.sp)) + print(f"loaded sp: vocab={sp.vocab_size()}", flush=True) + + train_out = args.out / "datasets" / "fineweb10B_sp8192_lossless_caps_caseops_v1_reserved" + train_out.mkdir(parents=True, exist_ok=True) + + val_buf_tokens: list[int] = [] + val_buf_bytes: list[int] = [] + train_buf: list[int] = [] + val_written = 0 + train_written = 0 + n_docs = 0 + + for text in _iter_docs(args.docs): + transformed = encode_lossless_caps_v2(text) + token_ids = [BOS_ID] + sp.encode(transformed, out_type=int) + if n_docs < args.val_docs: + # Validation doc — also compute byte sidecar + byte_counts = _token_original_byte_counts(sp, text, transformed) + val_buf_tokens.extend(token_ids) + val_buf_bytes.append(0) # BOS contributes 0 original bytes + val_buf_bytes.extend(int(b) for b in byte_counts) + if len(val_buf_tokens) >= SHARD_TOKENS: + _write_shard(train_out / f"fineweb_val_{val_written:06d}.bin", + np.array(val_buf_tokens[:SHARD_TOKENS], dtype=np.uint16)) + _write_shard(train_out / f"fineweb_val_bytes_{val_written:06d}.bin", + np.array(val_buf_bytes[:SHARD_TOKENS], dtype=np.uint16)) + val_buf_tokens = val_buf_tokens[SHARD_TOKENS:] + val_buf_bytes = val_buf_bytes[SHARD_TOKENS:] + val_written += 1 + else: + train_buf.extend(token_ids) + if len(train_buf) >= SHARD_TOKENS: + _write_shard(train_out / f"fineweb_train_{train_written:06d}.bin", + np.array(train_buf[:SHARD_TOKENS], dtype=np.uint16)) + train_buf = train_buf[SHARD_TOKENS:] + train_written += 1 + n_docs += 1 + if n_docs % 10_000 == 0: + print(f" processed {n_docs} docs train_shards={train_written} val_shards={val_written}", flush=True) + + # Flush tail buffers into final (possibly short) shards. + if val_buf_tokens: + _write_shard(train_out / f"fineweb_val_{val_written:06d}.bin", + np.array(val_buf_tokens, dtype=np.uint16)) + _write_shard(train_out / f"fineweb_val_bytes_{val_written:06d}.bin", + np.array(val_buf_bytes, dtype=np.uint16)) + if train_buf: + _write_shard(train_out / f"fineweb_train_{train_written:06d}.bin", + np.array(train_buf, dtype=np.uint16)) + + print(f"done. docs={n_docs} train_shards={train_written + (1 if train_buf else 0)} val_shards={val_written + (1 if val_buf_tokens else 0)}") + + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/requirements.txt b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/requirements.txt new file mode 100644 index 0000000000..b6c55e13aa --- /dev/null +++ b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/requirements.txt @@ -0,0 +1,13 @@ +# Python deps. Install with: pip install -r requirements.txt +torch==2.9.1+cu128 +sentencepiece +brotli +huggingface_hub +numpy +python-minifier + +# FlashAttention 3 must be installed separately (not on PyPI): +# pip install --no-deps flash_attn_3 --find-links https://windreamer.github.io/flash-attention3-wheels/cu128_torch291/ + +# System dep (apt): lrzip (used by per-group compressor) +# apt-get install -y lrzip diff --git a/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/submission.json b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/submission.json new file mode 100644 index 0000000000..4a7c9e6970 --- /dev/null +++ b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/submission.json @@ -0,0 +1,25 @@ +{ + "author": "aquarious dante workman", + "github_id": "aquariouseworkman", + "name": "PR #1855 base + activation-aware GPTQ mixed precision (AWQ-lite)", + "date": "2026-04-29", + "track": "10min_16mb", + "val_bpb": 1.06086, + "val_bpb_std": 0.00069, + "seeds": [42, 314, 1234], + "seed_results": { + "42": {"val_loss": 2.32225638, "val_bpb": 1.06118055, "artifact_bytes": 15978503, "steps": 4989, "eval_time_s": 392.8}, + "314": {"val_loss": 2.31977393, "val_bpb": 1.06004616, "artifact_bytes": 15976469, "steps": 4986, "eval_time_s": 395.8}, + "1234": {"val_loss": 2.32262534, "val_bpb": 1.06134915, "artifact_bytes": 15976673, "steps": 4977, "eval_time_s": 395.5} + }, + "hardware": "8xH100 80GB SXM", + "pytorch_version": "2.9.1+cu128", + "cuda_version": "12.8", + "flash_attn_version": "FA3 3.0.0 (cu128_torch291 wheel)", + "compressor": "pergroup", + "technique_summary": "PR #1855 9-knob hyperparameter stack + PR #1908 activation-aware GPTQ mixed precision (AWQ-lite). Training identical to PR #1855. Post-training: AWQ_LITE_ENABLED=1 selects 1 salient 64-column group per weight matrix via activation RMS * weight magnitude saliency, quantizes at int8 (vs int6 baseline) inside the same GPTQ solve. LQER asymmetric rank-4 correction on top. Per-group lrzip+brotli compression. Phased TTT eval (3 phases, prefix=2500 docs).", + "compliance": { + "artifact_bytes_max": 15978503, + "eval_time_max_s": 395.8 + } +} diff --git a/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model new file mode 100644 index 0000000000..fffc8bb306 Binary files /dev/null and b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model differ diff --git a/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_gpt.py b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_gpt.py new file mode 100644 index 0000000000..a7f006cf7c --- /dev/null +++ b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_gpt.py @@ -0,0 +1,3998 @@ +import base64, collections, copy, fcntl, glob, io, lzma, math, os +from pathlib import Path +import random, re, subprocess, sys, time, uuid, numpy as np, sentencepiece as spm, torch, torch.distributed as dist, torch.nn.functional as F +from torch import Tensor, nn +from flash_attn_interface import ( + flash_attn_func as flash_attn_3_func, + flash_attn_varlen_func, +) +from concurrent.futures import ThreadPoolExecutor +import triton +import triton.language as tl +from triton.tools.tensor_descriptor import TensorDescriptor + + +# ===== Fused softcapped cross-entropy (Triton) — training-only path ===== +# Replaces the eager +# logits_softcap = softcap * tanh(logits / softcap) +# F.cross_entropy(logits_softcap.float(), targets, reduction="mean") +# sequence with a single fused kernel that reads logits_proj once, applies +# softcap in-register, and computes (LSE, loss) in one streaming pass. The +# backward kernel mirrors the forward so there's no stored softcapped logits. +# Numerically identical to the eager path up to fp32 accumulation differences. +_FUSED_CE_LIBRARY = "pgsubmission1draft7fusedce" +_FUSED_CE_BLOCK_SIZE = 1024 +_FUSED_CE_NUM_WARPS = 4 + + +@triton.jit +def _softcapped_ce_fwd_kernel( + logits_ptr, losses_ptr, lse_ptr, targets_ptr, + stride_logits_n, stride_logits_v, + n_rows, n_cols, softcap, + block_size: tl.constexpr, +): + row_idx = tl.program_id(0).to(tl.int64) + logits_row_ptr = logits_ptr + row_idx * stride_logits_n + max_val = -float("inf") + sum_exp = 0.0 + A = 2.0 * softcap + inv_C = 2.0 / softcap + for off in range(0, n_cols, block_size): + cols = off + tl.arange(0, block_size) + mask = cols < n_cols + val = tl.load( + logits_row_ptr + cols * stride_logits_v, + mask=mask, other=-float("inf"), + ).to(tl.float32) + z = A * tl.sigmoid(val * inv_C) + z = tl.where(mask, z, -float("inf")) + curr_max = tl.max(z, axis=0) + new_max = tl.maximum(max_val, curr_max) + sum_exp = sum_exp * tl.exp(max_val - new_max) + tl.sum(tl.exp(z - new_max), axis=0) + max_val = new_max + lse = max_val + tl.log(sum_exp) + tl.store(lse_ptr + row_idx, lse) + target = tl.load(targets_ptr + row_idx).to(tl.int32) + target_val = tl.load(logits_row_ptr + target * stride_logits_v).to(tl.float32) + target_z = A * tl.sigmoid(target_val * inv_C) + tl.store(losses_ptr + row_idx, lse - target_z) + + +@triton.jit +def _softcapped_ce_bwd_kernel( + grad_logits_ptr, grad_losses_ptr, lse_ptr, logits_ptr, targets_ptr, + stride_logits_n, stride_logits_v, + stride_grad_n, stride_grad_v, + n_rows, n_cols, softcap, + block_size: tl.constexpr, +): + row_idx = tl.program_id(0).to(tl.int64) + logits_row_ptr = logits_ptr + row_idx * stride_logits_n + grad_row_ptr = grad_logits_ptr + row_idx * stride_grad_n + lse = tl.load(lse_ptr + row_idx) + grad_loss = tl.load(grad_losses_ptr + row_idx).to(tl.float32) + target = tl.load(targets_ptr + row_idx).to(tl.int32) + A = 2.0 * softcap + inv_C = 2.0 / softcap + dz_dx_scale = A * inv_C + for off in range(0, n_cols, block_size): + cols = off + tl.arange(0, block_size) + mask = cols < n_cols + val = tl.load( + logits_row_ptr + cols * stride_logits_v, + mask=mask, other=0.0, + ).to(tl.float32) + sigmoid_u = tl.sigmoid(val * inv_C) + z = A * sigmoid_u + probs = tl.exp(z - lse) + grad_z = grad_loss * (probs - tl.where(cols == target, 1.0, 0.0)) + grad_x = grad_z * (dz_dx_scale * sigmoid_u * (1.0 - sigmoid_u)) + tl.store(grad_row_ptr + cols * stride_grad_v, grad_x, mask=mask) + + +def _validate_softcapped_ce_inputs( + logits: Tensor, targets: Tensor, softcap: float, +) -> tuple[Tensor, Tensor]: + if logits.ndim != 2: + raise ValueError(f"Expected logits.ndim=2, got {logits.ndim}") + if targets.ndim != 1: + raise ValueError(f"Expected targets.ndim=1, got {targets.ndim}") + if logits.shape[0] != targets.shape[0]: + raise ValueError( + f"Expected matching rows, got logits={tuple(logits.shape)} targets={tuple(targets.shape)}" + ) + if not logits.is_cuda or not targets.is_cuda: + raise ValueError("softcapped_cross_entropy requires CUDA tensors") + if softcap <= 0.0: + raise ValueError(f"softcap must be positive, got {softcap}") + if logits.dtype not in (torch.float16, torch.bfloat16, torch.float32): + raise ValueError(f"Unsupported logits dtype: {logits.dtype}") + logits = logits.contiguous() + targets = targets.contiguous() + if targets.dtype != torch.int64: + targets = targets.to(dtype=torch.int64) + return logits, targets + + +@torch.library.custom_op(f"{_FUSED_CE_LIBRARY}::softcapped_ce", mutates_args=()) +def softcapped_ce_op(logits: Tensor, targets: Tensor, softcap: float) -> tuple[Tensor, Tensor]: + logits, targets = _validate_softcapped_ce_inputs(logits, targets, float(softcap)) + n_rows, n_cols = logits.shape + losses = torch.empty((n_rows,), device=logits.device, dtype=torch.float32) + lse = torch.empty((n_rows,), device=logits.device, dtype=torch.float32) + _softcapped_ce_fwd_kernel[(n_rows,)]( + logits, losses, lse, targets, + logits.stride(0), logits.stride(1), + n_rows, n_cols, float(softcap), + block_size=_FUSED_CE_BLOCK_SIZE, num_warps=_FUSED_CE_NUM_WARPS, + ) + return losses, lse + + +@softcapped_ce_op.register_fake +def _(logits: Tensor, targets: Tensor, softcap: float): + if logits.ndim != 2 or targets.ndim != 1: + raise ValueError("softcapped_ce fake impl expects 2D logits and 1D targets") + if logits.shape[0] != targets.shape[0]: + raise ValueError( + f"Expected matching rows, got logits={tuple(logits.shape)} targets={tuple(targets.shape)}" + ) + n_rows = logits.shape[0] + return ( + logits.new_empty((n_rows,), dtype=torch.float32), + logits.new_empty((n_rows,), dtype=torch.float32), + ) + + +@torch.library.custom_op(f"{_FUSED_CE_LIBRARY}::softcapped_ce_backward", mutates_args=()) +def softcapped_ce_backward_op( + logits: Tensor, targets: Tensor, lse: Tensor, grad_losses: Tensor, softcap: float, +) -> Tensor: + logits, targets = _validate_softcapped_ce_inputs(logits, targets, float(softcap)) + lse = lse.contiguous() + grad_losses = grad_losses.contiguous().to(dtype=torch.float32) + if lse.ndim != 1 or grad_losses.ndim != 1: + raise ValueError("Expected 1D lse and grad_losses") + if lse.shape[0] != logits.shape[0] or grad_losses.shape[0] != logits.shape[0]: + raise ValueError( + f"Expected row-aligned lse/grad_losses, got logits={tuple(logits.shape)} " + f"lse={tuple(lse.shape)} grad_losses={tuple(grad_losses.shape)}" + ) + grad_logits = torch.empty_like(logits) + n_rows, n_cols = logits.shape + _softcapped_ce_bwd_kernel[(n_rows,)]( + grad_logits, grad_losses, lse, logits, targets, + logits.stride(0), logits.stride(1), + grad_logits.stride(0), grad_logits.stride(1), + n_rows, n_cols, float(softcap), + block_size=_FUSED_CE_BLOCK_SIZE, num_warps=_FUSED_CE_NUM_WARPS, + ) + return grad_logits + + +@softcapped_ce_backward_op.register_fake +def _(logits: Tensor, targets: Tensor, lse: Tensor, grad_losses: Tensor, softcap: float): + if logits.ndim != 2 or targets.ndim != 1 or lse.ndim != 1 or grad_losses.ndim != 1: + raise ValueError("softcapped_ce_backward fake impl expects 2D logits and 1D row tensors") + if ( + logits.shape[0] != targets.shape[0] + or logits.shape[0] != lse.shape[0] + or logits.shape[0] != grad_losses.shape[0] + ): + raise ValueError("softcapped_ce_backward fake impl expects row-aligned tensors") + return logits.new_empty(logits.shape) + + +def _softcapped_ce_setup_context( + ctx: torch.autograd.function.FunctionCtx, inputs, output, +) -> None: + logits, targets, softcap = inputs + _losses, lse = output + ctx.save_for_backward(logits, targets, lse) + ctx.softcap = float(softcap) + + +def _softcapped_ce_backward( + ctx: torch.autograd.function.FunctionCtx, grad_losses: Tensor, grad_lse: "Tensor | None", +): + del grad_lse + logits, targets, lse = ctx.saved_tensors + grad_logits = torch.ops.pgsubmission1draft7fusedce.softcapped_ce_backward( + logits, targets, lse, grad_losses, ctx.softcap + ) + return grad_logits, None, None + + +softcapped_ce_op.register_autograd( + _softcapped_ce_backward, setup_context=_softcapped_ce_setup_context, +) + + +def softcapped_cross_entropy( + logits: Tensor, targets: Tensor, softcap: float, reduction: str = "mean", +) -> Tensor: + losses, _lse = torch.ops.pgsubmission1draft7fusedce.softcapped_ce( + logits, targets, float(softcap) + ) + if reduction == "none": + return losses + if reduction == "sum": + return losses.sum() + if reduction == "mean": + return losses.mean() + raise ValueError(f"Unsupported reduction={reduction!r}") + + +class Hyperparameters: + data_dir = os.environ.get("DATA_DIR", "./data/") + seed = int(os.environ.get("SEED", 1337)) + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_frac = float(os.environ.get("WARMDOWN_FRAC", 0.75)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786432)) + # Fused softcapped CE (Triton). Training-only — forward_logits eval path still uses + # eager softcap+F.cross_entropy. Default ON since validated as at-worst neutral. + fused_ce_enabled = bool(int(os.environ.get("FUSED_CE_ENABLED", "1"))) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 6e2)) + val_batch_tokens = int(os.environ.get("VAL_BATCH_TOKENS", 524288)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 8192)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 4.0)) + skip_gates_enabled = bool(int(os.environ.get("SKIP_GATES_ENABLED", "1"))) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 3e1)) + rope_base = float(os.environ.get("ROPE_BASE", 1e4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + rope_train_seq_len = int(os.environ.get("ROPE_TRAIN_SEQ_LEN", 2048)) + rope_yarn = bool(int(os.environ.get("ROPE_YARN", "0"))) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 5.0)) + num_loops = int(os.environ.get("NUM_LOOPS", 2)) + loop_start = int(os.environ.get("LOOP_START", 3)) + loop_end = int(os.environ.get("LOOP_END", 5)) + enable_looping_at = float(os.environ.get("ENABLE_LOOPING_AT", 0.35)) + parallel_start_layer = int(os.environ.get("PARALLEL_START_LAYER", 8)) + parallel_final_lane = os.environ.get("PARALLEL_FINAL_LANE", "mean") + min_lr = float(os.environ.get("MIN_LR", 0.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.026)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.97)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float( + os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92) + ) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + muon_row_normalize = bool(int(os.environ.get("MUON_ROW_NORMALIZE", "1"))) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-08)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + adam_wd = float(os.environ.get("ADAM_WD", 0.02)) + muon_wd = float(os.environ.get("MUON_WD", 0.095)) + embed_wd = float(os.environ.get("EMBED_WD", 0.085)) + ema_decay = float(os.environ.get("EMA_DECAY", 0.9965)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lora_rank = int(os.environ.get("TTT_LORA_RANK", 96)) + ttt_lora_lr = float(os.environ.get("TTT_LORA_LR", 0.0001)) + ttt_chunk_size = int(os.environ.get("TTT_CHUNK_SIZE", 48)) + ttt_eval_seq_len = int(os.environ.get("TTT_EVAL_SEQ_LEN", 2048)) + ttt_batch_size = int(os.environ.get("TTT_BATCH_SIZE", 64)) + ttt_grad_steps = int(os.environ.get("TTT_GRAD_STEPS", 1)) + ttt_weight_decay = float(os.environ.get("TTT_WEIGHT_DECAY", 1.0)) + ttt_beta1 = float(os.environ.get("TTT_BETA1", 0)) + ttt_beta2 = float(os.environ.get("TTT_BETA2", 0.999)) + ttt_k_lora = bool(int(os.environ.get("TTT_K_LORA", "1"))) + ttt_mlp_lora = bool(int(os.environ.get("TTT_MLP_LORA", "1"))) + ttt_o_lora = bool(int(os.environ.get("TTT_O_LORA", "1"))) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adam") + ttt_eval_batches = os.environ.get("TTT_EVAL_BATCHES", "") + val_doc_fraction = float(os.environ.get("VAL_DOC_FRACTION", 1.0)) + compressor = os.environ.get("COMPRESSOR", "brotli") + gptq_calibration_batches = int(os.environ.get("GPTQ_CALIBRATION_BATCHES", 16)) + gptq_reserve_seconds = float(os.environ.get("GPTQ_RESERVE_SECONDS", 4.0)) + phased_ttt_prefix_docs = int(os.environ.get("PHASED_TTT_PREFIX_DOCS", 2000)) + phased_ttt_num_phases = int(os.environ.get("PHASED_TTT_NUM_PHASES", 1)) + global_ttt_lr = float(os.environ.get("GLOBAL_TTT_LR", 0.001)) + global_ttt_momentum = float(os.environ.get("GLOBAL_TTT_MOMENTUM", 0.9)) + global_ttt_epochs = int(os.environ.get("GLOBAL_TTT_EPOCHS", 1)) + global_ttt_chunk_tokens = int(os.environ.get("GLOBAL_TTT_CHUNK_TOKENS", 32768)) + global_ttt_batch_seqs = int(os.environ.get("GLOBAL_TTT_BATCH_SEQS", 32)) + global_ttt_warmup_start_lr = float(os.environ.get("GLOBAL_TTT_WARMUP_START_LR", 0.0)) + global_ttt_warmup_chunks = int(os.environ.get("GLOBAL_TTT_WARMUP_CHUNKS", 0)) + global_ttt_grad_clip = float(os.environ.get("GLOBAL_TTT_GRAD_CLIP", 1.0)) + global_ttt_respect_doc_boundaries = bool(int(os.environ.get("GLOBAL_TTT_RESPECT_DOC_BOUNDARIES", "1"))) + matrix_bits = int(os.environ.get("MATRIX_BITS", 6)) + embed_bits = int(os.environ.get("EMBED_BITS", 8)) + matrix_clip_sigmas = float(os.environ.get("MATRIX_CLIP_SIGMAS", 12.85)) + embed_clip_sigmas = float(os.environ.get("EMBED_CLIP_SIGMAS", 2e1)) + mlp_clip_sigmas = float(os.environ.get("MLP_CLIP_SIGMAS", 10.0)) + attn_clip_sigmas = float(os.environ.get("ATTN_CLIP_SIGMAS", 13.0)) + # AttnOutGate (per-head multiplicative output gate, PR #1667 MarioPaerle). + # Zero-init weight: 2*sigmoid(0)=1 -> transparent at start. Source defaults to + # block input x ('proj'); 'q' uses raw Q projection output. + attn_out_gate_enabled = bool(int(os.environ.get("ATTN_OUT_GATE_ENABLED", "0"))) + attn_out_gate_src = os.environ.get("ATTN_OUT_GATE_SRC", "proj") + # SmearGate (input-dependent forward-1 token smear, modded-nanogpt @classiclarryd + # via PR #1667). x_t <- x_t + lam * sigmoid(W*x_t[:gate_window]) * x_{t-1}. + # lam=0 + W=0 -> transparent at init. + smear_gate_enabled = bool(int(os.environ.get("SMEAR_GATE_ENABLED", "0"))) + # Window: first GATE_WINDOW dims of the source feed the gate projection. + gate_window = int(os.environ.get("GATE_WINDOW", 12)) + # Gated Attention (Qwen, NeurIPS 2025 Best Paper, arXiv:2505.06708; + # qiuzh20/gated_attention). Per-head sigmoid gate on SDPA output, BEFORE + # out_proj. Gate input = full block input x (paper's headwise G1 variant + # driven from hidden_states). W_g shape (num_heads, dim), plain sigmoid. + # Near-zero init gives g~0.5 at step 0 (half attention output); per-block + # attn_scale (init 1.0) compensates during training. Name contains + # "attn_gate" so CONTROL_TENSOR_NAME_PATTERNS routes it to scalar AdamW. + gated_attn_enabled = bool(int(os.environ.get("GATED_ATTN_ENABLED", "0"))) + gated_attn_init_std = float(os.environ.get("GATED_ATTN_INIT_STD", 0.01)) + # Dedicated int8-per-row quantization for `attn_gate_w` tensors. These are + # small ((num_heads, dim) = (8, 512) = 4096 params) and bypass GPTQ via the + # numel<=65536 passthrough branch -> stored as fp16 (8 KB/layer, ~65 KB total + # compressed). int8-per-row cuts the raw tensor in half with negligible BPB + # impact: scales per head (8 values), symmetric quant over [-127, 127]. + # No Hessian needed (gate weights not in collect_hessians()). + gated_attn_quant_gate = bool(int(os.environ.get("GATED_ATTN_QUANT_GATE", "0"))) + # Sparse Attention Gate (modded-nanogpt-style). Keeps dense SDPA and only + # swaps the output-gate input to the first GATE_WINDOW residual dims. + # W_g: (num_heads, gate_window) = (8, 12) = 96 params/layer (~44K total), + # vs dense GatedAttn's (8, 512) = 4K/layer (~44K diff). Name "attn_gate_w" + # is shared so quant routing and int8 gate passthrough Just Work. Gate + # passthrough int8 still applies via GATED_ATTN_QUANT_GATE=1. + # Mutually exclusive with ATTN_OUT_GATE_ENABLED and GATED_ATTN_ENABLED. + sparse_attn_gate_enabled = bool(int(os.environ.get("SPARSE_ATTN_GATE_ENABLED", "0"))) + sparse_attn_gate_init_std = float(os.environ.get("SPARSE_ATTN_GATE_INIT_STD", 0.0)) + sparse_attn_gate_scale = float(os.environ.get("SPARSE_ATTN_GATE_SCALE", 1.0)) + # LQER asymmetric rank-k correction on top-K quant-error tensors (PR #1530 v2 port). + # Computes SVD of E = W_fp - W_quant, packs top-r A,B as INT2/INT4 (asym) or INTk (sym). + lqer_enabled = bool(int(os.environ.get("LQER_ENABLED", "1"))) + lqer_rank = int(os.environ.get("LQER_RANK", 4)) + lqer_top_k = int(os.environ.get("LQER_TOP_K", 3)) + lqer_factor_bits = int(os.environ.get("LQER_FACTOR_BITS", 4)) + lqer_asym_enabled = bool(int(os.environ.get("LQER_ASYM_ENABLED", "1"))) + lqer_asym_group = int(os.environ.get("LQER_ASYM_GROUP", "64")) + lqer_scope = os.environ.get("LQER_SCOPE", "all") + lqer_gain_select = bool(int(os.environ.get("LQER_GAIN_SELECT", "0"))) + awq_lite_enabled = bool(int(os.environ.get("AWQ_LITE_ENABLED", "0"))) + awq_lite_bits = int(os.environ.get("AWQ_LITE_BITS", "8")) + awq_lite_group_top_k = int(os.environ.get("AWQ_LITE_GROUP_TOP_K", "1")) + awq_lite_group_size = int(os.environ.get("AWQ_LITE_GROUP_SIZE", "64")) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + is_main_process = rank == 0 + grad_accum_steps = 8 // world_size + # CaseOps integration: optional override of dataset root + tokenizer path. + # When CASEOPS_ENABLED=1, the wrapper loads a per-token byte sidecar + # (fineweb_val_bytes_*.bin, identical shard layout to val_*.bin) and uses + # it as the canonical raw-byte budget for BPB accounting. The sidecar + # REPLACES the build_sentencepiece_luts byte-counting path entirely. + caseops_enabled = bool(int(os.environ.get("CASEOPS_ENABLED", "0"))) + _default_caseops_data = os.path.join( + data_dir, + "datasets", + "fineweb10B_sp8192_caseops", + "datasets", + "datasets", + "fineweb10B_sp8192_lossless_caps_caseops_v1_reserved", + ) + _default_caseops_tok = os.path.join( + data_dir, + "datasets", + "fineweb10B_sp8192_caseops", + "datasets", + "tokenizers", + "fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model", + ) + if caseops_enabled: + datasets_dir = os.environ.get("DATA_PATH", _default_caseops_data) + tokenizer_path = os.environ.get("TOKENIZER_PATH", _default_caseops_tok) + else: + datasets_dir = os.environ.get( + "DATA_PATH", + os.path.join(data_dir, "datasets", f"fineweb10B_sp{vocab_size}"), + ) + tokenizer_path = os.environ.get( + "TOKENIZER_PATH", + os.path.join(data_dir, "tokenizers", f"fineweb_{vocab_size}_bpe.model"), + ) + train_files = os.path.join(datasets_dir, "fineweb_train_*.bin") + val_files = os.path.join(datasets_dir, "fineweb_val_*.bin") + val_bytes_files = os.path.join(datasets_dir, "fineweb_val_bytes_*.bin") + artifact_dir = os.environ.get("ARTIFACT_DIR", "") + logfile = ( + os.path.join(artifact_dir, f"{run_id}.txt") + if artifact_dir + else f"logs/{run_id}.txt" + ) + model_path = ( + os.path.join(artifact_dir, "final_model.pt") + if artifact_dir + else "final_model.pt" + ) + quantized_model_path = ( + os.path.join(artifact_dir, "final_model.int6.ptz") + if artifact_dir + else "final_model.int6.ptz" + ) + + +_logger_hparams = None + + +def set_logging_hparams(h): + global _logger_hparams + _logger_hparams = h + + +def log(msg, console=True): + if _logger_hparams is None: + print(msg) + return + if _logger_hparams.is_main_process: + if console: + print(msg) + if _logger_hparams.logfile is not None: + with open(_logger_hparams.logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + +class ValidationData: + def __init__(self, h, device): + self.sp = spm.SentencePieceProcessor(model_file=h.tokenizer_path) + if int(self.sp.vocab_size()) != h.vocab_size: + raise ValueError( + f"VOCAB_SIZE={h.vocab_size} does not match tokenizer vocab_size={int(self.sp.vocab_size())}" + ) + self.val_tokens = load_validation_tokens(h.val_files, h.eval_seq_len) + self.caseops_enabled = bool(getattr(h, "caseops_enabled", False)) + if self.caseops_enabled: + self.base_bytes_lut = None + self.has_leading_space_lut = None + self.is_boundary_token_lut = None + else: + ( + self.base_bytes_lut, + self.has_leading_space_lut, + self.is_boundary_token_lut, + ) = build_sentencepiece_luts(self.sp, h.vocab_size, device) + self.val_bytes = None + if self.caseops_enabled: + self.val_bytes = load_validation_byte_sidecar( + h.val_bytes_files, h.eval_seq_len, self.val_tokens.numel() + ) + + +def build_sentencepiece_luts(sp, vocab_size, device): + sp_vocab_size = int(sp.vocab_size()) + assert ( + sp.piece_to_id("▁") != sp.unk_id() + ), "Tokenizer must have '▁' (space) as its own token for correct BPB byte counting" + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern, seq_len): + # Filter out CaseOps byte sidecar shards which share the val_*.bin glob. + files = [ + Path(p) + for p in sorted(glob.glob(pattern)) + if "_bytes_" not in Path(p).name + ] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = (tokens.numel() - 1) // seq_len * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def load_validation_byte_sidecar(pattern, seq_len, expected_len): + """Load CaseOps per-token byte sidecar(s). Same shard layout as token shards + (256 int32 header + uint16 array). Each entry = canonical raw-text byte + budget for that token in the corresponding val shard. Returns a CPU + int16 tensor sliced to match expected_len (i.e. val_tokens length).""" + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No byte sidecar files for pattern: {pattern}") + shards = [load_data_shard(file) for file in files] + # load_data_shard returns uint16 — that's exactly what the sidecar stores. + bytes_full = torch.cat(shards).contiguous() + if bytes_full.numel() < expected_len: + raise ValueError( + f"Byte sidecar too short: {bytes_full.numel()} < val_tokens {expected_len}" + ) + return bytes_full[:expected_len].to(torch.int32) + + +def load_data_shard(file): + header_bytes = 256 * np.dtype(" 0: + pos = start + while pos < end: + seg_starts.append(pos) + pos += max_doc_len + else: + seg_starts.append(start) + boundaries = seg_starts + [total_len] + padded_len = get_next_multiple_of_n(len(boundaries), bucket_size) + cu = torch.full((padded_len,), total_len, dtype=torch.int32, device=device) + cu[: len(boundaries)] = torch.tensor(boundaries, dtype=torch.int32, device=device) + seg_ends = seg_starts[1:] + [total_len] + max_seqlen = max(end - start for start, end in zip(seg_starts, seg_ends)) + return cu, max_seqlen + +class DocumentPackingLoader: + _shard_pool = ThreadPoolExecutor(1) + + def __init__(self, h, device, cu_bucket_size=64): + self.rank = h.rank + self.world_size = h.world_size + self.device = device + self.cu_bucket_size = cu_bucket_size + self.max_seq_len = h.train_seq_len + all_files = [Path(p) for p in sorted(glob.glob(h.train_files))] + if not all_files: + raise FileNotFoundError(f"No files found for pattern: {h.train_files}") + self.files = all_files + self.file_iter = iter(self.files) + self._init_shard(load_data_shard(next(self.file_iter))) + self._next_shard = self._submit_next_shard() + self._batch_pool = ThreadPoolExecutor(1) + self._prefetch_queue = [] + + def _init_shard(self, tokens): + global BOS_ID + self.tokens = tokens + self.shard_size = tokens.numel() + if BOS_ID is None: + BOS_ID = 1 + self.bos_idx = ( + (tokens == BOS_ID).nonzero(as_tuple=True)[0].to(torch.int64).cpu().numpy() + ) + self.cursor = int(self.bos_idx[0]) + + def _submit_next_shard(self): + try: + path = next(self.file_iter) + return self._shard_pool.submit(load_data_shard, path) + except StopIteration: + return None + + def _advance_shard(self): + if self._next_shard is None: + self.file_iter = iter(self.files) + self._next_shard = self._shard_pool.submit( + load_data_shard, next(self.file_iter) + ) + self._init_shard(self._next_shard.result()) + self._next_shard = self._submit_next_shard() + + def _local_doc_starts(self, local_start, total_len): + lo = np.searchsorted(self.bos_idx, local_start, side="left") + hi = np.searchsorted(self.bos_idx, local_start + total_len, side="left") + return (self.bos_idx[lo:hi] - local_start).tolist() + + def _prepare_batch(self, num_tokens_local, max_seq_len): + per_rank_span = num_tokens_local + 1 + global_span = per_rank_span * self.world_size + while self.cursor + global_span > self.shard_size: + self._advance_shard() + local_start = self.cursor + self.rank * per_rank_span + buf = self.tokens[local_start : local_start + per_rank_span] + inputs = torch.empty(per_rank_span - 1, dtype=torch.int64, pin_memory=True) + targets = torch.empty(per_rank_span - 1, dtype=torch.int64, pin_memory=True) + inputs.copy_(buf[:-1]) + targets.copy_(buf[1:]) + starts = self._local_doc_starts(local_start, inputs.numel()) + cu_seqlens, max_seqlen = _build_cu_seqlens( + starts, inputs.numel(), inputs.device, max_seq_len, self.cu_bucket_size + ) + cu_seqlens = cu_seqlens.pin_memory() + self.cursor += global_span + return inputs, targets, cu_seqlens, max_seqlen + + def next_batch(self, global_tokens, grad_accum_steps): + num_tokens_local = global_tokens // (self.world_size * grad_accum_steps) + while len(self._prefetch_queue) < 2: + self._prefetch_queue.append( + self._batch_pool.submit(self._prepare_batch, num_tokens_local, self.max_seq_len)) + inputs, targets, cu_seqlens, max_seqlen = self._prefetch_queue.pop(0).result() + self._prefetch_queue.append( + self._batch_pool.submit(self._prepare_batch, num_tokens_local, self.max_seq_len)) + return ( + inputs[None].to(self.device, non_blocking=True), + targets[None].to(self.device, non_blocking=True), + cu_seqlens.to(self.device, non_blocking=True), + max_seqlen, + ) + + +class ShuffledSequenceLoader: + def __init__(self, h, device): + self.world_size = h.world_size + self.seq_len = h.train_seq_len + self.device = device + all_files = [Path(p) for p in sorted(glob.glob(h.train_files))] + if not all_files: + raise FileNotFoundError(f"No files found for pattern: {h.train_files}") + self.files = all_files[h.rank :: h.world_size] + self.rng = np.random.Generator(np.random.PCG64(h.rank)) + self.num_tokens = [_read_num_tokens(f) for f in self.files] + self.start_inds = [[] for _ in self.files] + for si in range(len(self.files)): + self._reset_shard(si) + + def _reset_shard(self, si): + max_phase = min( + self.seq_len - 1, max(0, self.num_tokens[si] - self.seq_len - 1) + ) + phase = int(self.rng.integers(max_phase + 1)) if max_phase > 0 else 0 + num_sequences = (self.num_tokens[si] - 1 - phase) // self.seq_len + sequence_order = self.rng.permutation(num_sequences) + self.start_inds[si] = (phase + sequence_order * self.seq_len).tolist() + + def next_batch(self, global_tokens, grad_accum_steps): + device_tokens = global_tokens // (self.world_size * grad_accum_steps) + device_batch_size = device_tokens // self.seq_len + remaining = np.array([len(s) for s in self.start_inds], dtype=np.float64) + x = torch.empty((device_batch_size, self.seq_len), dtype=torch.int64) + y = torch.empty((device_batch_size, self.seq_len), dtype=torch.int64) + for bi in range(device_batch_size): + total = remaining.sum() + if total <= 0: + for si in range(len(self.files)): + self._reset_shard(si) + remaining = np.array( + [len(s) for s in self.start_inds], dtype=np.float64 + ) + total = remaining.sum() + probs = remaining / total + si = int(self.rng.choice(len(self.files), p=probs)) + start_ind = self.start_inds[si].pop() + remaining[si] -= 1 + mm = _get_shard_memmap(self.files[si]) + window = torch.as_tensor( + np.array(mm[start_ind : start_ind + self.seq_len + 1], dtype=np.int64) + ) + x[bi] = window[:-1] + y[bi] = window[1:] + return x.to(self.device, non_blocking=True), y.to( + self.device, non_blocking=True + ) + + +class RMSNorm(nn.Module): + def __init__(self, eps=None): + super().__init__() + self.eps = eps + + def forward(self, x): + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + def forward(self, x): + w = self.weight.to(x.dtype) + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + + +@triton.jit +def linear_leaky_relu_square_kernel( + a_desc, + b_desc, + c_desc, + aux_desc, + M, + N, + K, + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + NUM_SMS: tl.constexpr, + FORWARD: tl.constexpr, +): + dtype = tl.bfloat16 + start_pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(M, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + k_tiles = tl.cdiv(K, BLOCK_SIZE_K) + num_tiles = num_pid_m * num_pid_n + tile_id_c = start_pid - NUM_SMS + for tile_id in tl.range(start_pid, num_tiles, NUM_SMS, flatten=True): + pid_m = tile_id // num_pid_n + pid_n = tile_id % num_pid_n + offs_am = pid_m * BLOCK_SIZE_M + offs_bn = pid_n * BLOCK_SIZE_N + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for ki in range(k_tiles): + offs_k = ki * BLOCK_SIZE_K + a = a_desc.load([offs_am, offs_k]) + b = b_desc.load([offs_bn, offs_k]) + accumulator = tl.dot(a, b.T, accumulator) + tile_id_c += NUM_SMS + offs_am_c = offs_am + offs_bn_c = offs_bn + acc = tl.reshape(accumulator, (BLOCK_SIZE_M, 2, BLOCK_SIZE_N // 2)) + acc = tl.permute(acc, (0, 2, 1)) + acc0, acc1 = tl.split(acc) + c0 = acc0.to(dtype) + c1 = acc1.to(dtype) + if not FORWARD: + pre0 = aux_desc.load([offs_am_c, offs_bn_c]) + pre1 = aux_desc.load([offs_am_c, offs_bn_c + BLOCK_SIZE_N // 2]) + c0 = c0 * tl.where(pre0 > 0, 2.0 * pre0, 0.5 * pre0) + c1 = c1 * tl.where(pre1 > 0, 2.0 * pre1, 0.5 * pre1) + c_desc.store([offs_am_c, offs_bn_c], c0) + c_desc.store([offs_am_c, offs_bn_c + BLOCK_SIZE_N // 2], c1) + if FORWARD: + aux0 = tl.where(c0 > 0, c0, 0.5 * c0) + aux1 = tl.where(c1 > 0, c1, 0.5 * c1) + aux_desc.store([offs_am_c, offs_bn_c], aux0 * aux0) + aux_desc.store([offs_am_c, offs_bn_c + BLOCK_SIZE_N // 2], aux1 * aux1) + + +def linear_leaky_relu_square(a, b, aux=None): + M, K = a.shape + N, K2 = b.shape + assert K == K2 + c = torch.empty((M, N), device=a.device, dtype=a.dtype) + forward = aux is None + if aux is None: + aux = torch.empty((M, N), device=a.device, dtype=a.dtype) + num_sms = torch.cuda.get_device_properties(a.device).multi_processor_count + BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K = 256, 128, 64 + num_stages = 4 if forward else 3 + a_desc = TensorDescriptor.from_tensor(a, [BLOCK_SIZE_M, BLOCK_SIZE_K]) + b_desc = TensorDescriptor.from_tensor(b, [BLOCK_SIZE_N, BLOCK_SIZE_K]) + c_desc = TensorDescriptor.from_tensor(c, [BLOCK_SIZE_M, BLOCK_SIZE_N // 2]) + aux_desc = TensorDescriptor.from_tensor(aux, [BLOCK_SIZE_M, BLOCK_SIZE_N // 2]) + grid = lambda _meta: ( + min(num_sms, triton.cdiv(M, BLOCK_SIZE_M) * triton.cdiv(N, BLOCK_SIZE_N)), + ) + linear_leaky_relu_square_kernel[grid]( + a_desc, + b_desc, + c_desc, + aux_desc, + M, + N, + K, + BLOCK_SIZE_M=BLOCK_SIZE_M, + BLOCK_SIZE_N=BLOCK_SIZE_N, + BLOCK_SIZE_K=BLOCK_SIZE_K, + NUM_SMS=num_sms, + FORWARD=forward, + num_stages=num_stages, + num_warps=8, + ) + if forward: + return c, aux + return c + + +class FusedLinearLeakyReLUSquareFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, x, w1, w2): + x_flat = x.reshape(-1, x.shape[-1]) + pre, post = linear_leaky_relu_square(x_flat, w1) + out = F.linear(post, w2) + ctx.save_for_backward(x, w1, w2, pre, post) + return out.view(*x.shape[:-1], out.shape[-1]) + + @staticmethod + def backward(ctx, grad_output): + x, w1, w2, pre, post = ctx.saved_tensors + x_flat = x.reshape(-1, x.shape[-1]) + grad_output_flat = grad_output.reshape(-1, grad_output.shape[-1]) + dw2 = grad_output_flat.T @ post + dpre = linear_leaky_relu_square(grad_output_flat, w2.T.contiguous(), aux=pre) + dw1 = dpre.T @ x_flat + dx = dpre @ w1 + return dx.view_as(x), dw1, dw2 + + +FusedLeakyReLUSquareMLP = FusedLinearLeakyReLUSquareFunction.apply + + +class Rotary(nn.Module): + def __init__(self, dim, base=1e4, train_seq_len=1024, rope_dims=0, yarn=True): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.yarn = yarn + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / base ** ( + torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims + ) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached = None + self._sin_cached = None + + def forward(self, seq_len, device, dtype): + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached < seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if self.yarn and seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * scale ** (rd / (rd - 2)) + inv_freq = 1.0 / new_base ** ( + torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd + ) + else: + inv_freq = self.inv_freq.float().to(device) + t = torch.arange(seq_len, device=device, dtype=torch.float32) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached[:, :seq_len].to(dtype=dtype), self._sin_cached[:, :seq_len].to(dtype=dtype) + + +def apply_rotary_emb(x, cos, sin, rope_dims=0): + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * -sin + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * -sin + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__( + self, dim, num_heads, num_kv_heads, rope_base, qk_gain_init, train_seq_len, yarn=True, + attn_out_gate=False, attn_out_gate_src="proj", gate_window=12, + gated_attn=False, gated_attn_init_std=0.01, + sparse_attn_gate=False, sparse_attn_gate_init_std=0.0, sparse_attn_gate_scale=1.0, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + if int(attn_out_gate) + int(gated_attn) + int(sparse_attn_gate) > 1: + raise ValueError( + "attn_out_gate, gated_attn, and sparse_attn_gate are mutually exclusive" + ) + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + self.q_gain = nn.Parameter( + torch.full((num_heads,), qk_gain_init, dtype=torch.float32) + ) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=train_seq_len, yarn=yarn) + self.use_xsa = False + # AttnOutGate (PR #1667 MarioPaerle): per-head multiplicative gate on attention + # output. CastedLinear so restore_fp32_params casts back to fp32 for GPTQ. + # _zero_init -> 2*sigmoid(0)=1 -> transparent at init. + self.attn_out_gate = attn_out_gate + self.attn_out_gate_src = attn_out_gate_src + self.gate_window = gate_window + if attn_out_gate: + self.attn_gate_proj = CastedLinear(gate_window, num_heads, bias=False) + self.attn_gate_proj._zero_init = True + # Gated Attention (arXiv:2505.06708, Qwen, NeurIPS 2025). Per-head sigmoid + # gate on SDPA output, BEFORE out_proj. Gate projection W_g: (num_heads, dim). + # Name "attn_gate_w" contains "attn_gate" substring so it matches + # CONTROL_TENSOR_NAME_PATTERNS and routes to the scalar AdamW group. + # fp32 Parameter -> restore_fp32_params path covers it via the ndim<2 OR + # name-pattern check (name matches "attn_gate"). Cast to x.dtype on use. + self.gated_attn = gated_attn + if gated_attn: + W = torch.empty(num_heads, dim, dtype=torch.float32) + nn.init.normal_(W, mean=0.0, std=gated_attn_init_std) + self.attn_gate_w = nn.Parameter(W) + # Sparse attention head-output gate (modded-nanogpt style). Keeps dense SDPA + # and only narrows the gate input to the first gate_window residual dims. + # W_g: (num_heads, gate_window). y_{t,h} <- sigmoid(scale * W_g_h @ x_t[:gate_window]) * y_{t,h}. + # Shares attn_gate_w name with dense GatedAttn so the quant routing + # (CONTROL_TENSOR_NAME_PATTERNS / attn_gate_w int8 passthrough) is unchanged. + self.sparse_attn_gate = sparse_attn_gate + self.sparse_attn_gate_scale = sparse_attn_gate_scale + if sparse_attn_gate: + W = torch.empty(num_heads, gate_window, dtype=torch.float32) + if sparse_attn_gate_init_std > 0: + nn.init.normal_(W, mean=0.0, std=sparse_attn_gate_init_std) + else: + nn.init.zeros_(W) + self.attn_gate_w = nn.Parameter(W) + + def _xsa_efficient(self, y, v): + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + + def forward(self, x, q_w, k_w, v_w, out_w, cu_seqlens=None, max_seqlen=0): + bsz, seqlen, dim = x.shape + # q_raw kept around as a tap point for attn_out_gate_src='q' (post-projection, + # pre-reshape, pre-RoPE). + q_raw = F.linear(x, q_w.to(x.dtype)) + q = q_raw.reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if cu_seqlens is not None: + y = flash_attn_varlen_func( + q[0], + k[0], + v[0], + cu_seqlens_q=cu_seqlens, + cu_seqlens_k=cu_seqlens, + max_seqlen_q=max_seqlen, + max_seqlen_k=max_seqlen, + causal=True, + window_size=(-1, -1), + )[None] + else: + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + # AttnOutGate inlined (PR #1667). Inline + .contiguous() barrier so torch.compile + # fullgraph=True is happy (this avoids the @torch.compiler.disable trap that + # crashed gates v3). Per-head gate on (B,T,H,D) tensor: g shape [B,T,H], broadcast + # over D via [..., None]. zero-init weight -> 2*sigmoid(0)=1 -> transparent. + if self.attn_out_gate: + gate_src = q_raw if self.attn_out_gate_src == "q" else x + gate_in = gate_src[..., : self.gate_window].contiguous() + g = 2.0 * torch.sigmoid(self.attn_gate_proj(gate_in)) + y = y * g[..., None] + # Gated Attention (arXiv:2505.06708 G1). Inline + .contiguous() barrier so + # torch.compile fullgraph=True is happy. Per-head gate on (B,T,H,D): g shape + # [B,T,H], broadcast over D via [..., None]. Paper: g = sigmoid(x @ W_g.T) + # where W_g: (H, dim). .to(x.dtype) on fp32 param before broadcast with bf16. + if self.gated_attn: + x_c = x.contiguous() + g = torch.sigmoid(F.linear(x_c, self.attn_gate_w.to(x.dtype))) + y = y * g[..., None] + # Sparse head-output gate: narrower (gate_window) input, same shape g as GatedAttn. + if self.sparse_attn_gate: + gate_in = x[..., : self.gate_window].contiguous() + g = torch.sigmoid( + self.sparse_attn_gate_scale + * F.linear(gate_in, self.attn_gate_w.to(x.dtype)) + ) + y = y * g[..., None] + y = y.reshape(bsz, seqlen, dim) + self._last_proj_input = y.detach() if getattr(self, "_calib", False) else None + return F.linear(y, out_w.to(x.dtype)) + + +class MLP(nn.Module): + def __init__(self, dim, mlp_mult): + super().__init__() + self.use_fused = True + + def forward(self, x, up_w, down_w): + if self.training and self.use_fused: + return FusedLeakyReLUSquareMLP(x, up_w.to(x.dtype), down_w.to(x.dtype)) + hidden = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5).square() + self._last_down_input = hidden.detach() if getattr(self, "_calib", False) else None + return F.linear(hidden, down_w.to(x.dtype)) + + +class Block(nn.Module): + def __init__( + self, + dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + train_seq_len, + layer_idx=0, + ln_scale=False, + yarn=True, + attn_out_gate=False, + attn_out_gate_src="proj", + gate_window=12, + gated_attn=False, + gated_attn_init_std=0.01, + sparse_attn_gate=False, + sparse_attn_gate_init_std=0.0, + sparse_attn_gate_scale=1.0, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention( + dim, num_heads, num_kv_heads, rope_base, qk_gain_init, train_seq_len, yarn=yarn, + attn_out_gate=attn_out_gate, attn_out_gate_src=attn_out_gate_src, gate_window=gate_window, + gated_attn=gated_attn, gated_attn_init_std=gated_attn_init_std, + sparse_attn_gate=sparse_attn_gate, + sparse_attn_gate_init_std=sparse_attn_gate_init_std, + sparse_attn_gate_scale=sparse_attn_gate_scale, + ) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter( + torch.stack((torch.ones(dim), torch.zeros(dim))).float() + ) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + + def forward(self, x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=None, max_seqlen=0): + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn( + self.attn_norm(x_in) * self.ln_scale_factor, + q_w, k_w, v_w, out_w, + cu_seqlens=cu_seqlens, + max_seqlen=max_seqlen, + ) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[ + None, None, : + ] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + return x_out + +class GPT(nn.Module): + def __init__(self, h): + super().__init__() + if h.logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {h.logit_softcap}") + self.tie_embeddings = h.tie_embeddings + self.tied_embed_init_std = h.tied_embed_init_std + self.logit_softcap = h.logit_softcap + self.fused_ce_enabled = bool(h.fused_ce_enabled) + self.tok_emb = nn.Embedding(h.vocab_size, h.model_dim) + self.num_layers = h.num_layers + head_dim = h.model_dim // h.num_heads + kv_dim = h.num_kv_heads * head_dim + hidden_dim = int(h.mlp_mult * h.model_dim) + self.qo_bank = nn.Parameter(torch.empty(2 * h.num_layers, h.model_dim, h.model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * h.num_layers, kv_dim, h.model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(h.num_layers, hidden_dim, h.model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(h.num_layers, h.model_dim, hidden_dim)) + self.num_encoder_layers = h.num_layers // 2 + self.num_decoder_layers = h.num_layers - self.num_encoder_layers + self.blocks = nn.ModuleList( + [ + Block( + h.model_dim, + h.num_heads, + h.num_kv_heads, + h.mlp_mult, + h.rope_base, + h.qk_gain_init, + h.train_seq_len, + layer_idx=i, + ln_scale=h.ln_scale, + yarn=h.rope_yarn, + attn_out_gate=h.attn_out_gate_enabled, + attn_out_gate_src=h.attn_out_gate_src, + gate_window=h.gate_window, + gated_attn=h.gated_attn_enabled, + gated_attn_init_std=h.gated_attn_init_std, + sparse_attn_gate=h.sparse_attn_gate_enabled, + sparse_attn_gate_init_std=h.sparse_attn_gate_init_std, + sparse_attn_gate_scale=h.sparse_attn_gate_scale, + ) + for i in range(h.num_layers) + ] + ) + if h.rope_dims > 0: + head_dim = h.model_dim // h.num_heads + for block in self.blocks: + block.attn.rope_dims = h.rope_dims + block.attn.rotary = Rotary( + head_dim, + base=h.rope_base, + train_seq_len=h.train_seq_len, + rope_dims=h.rope_dims, + yarn=h.rope_yarn, + ) + self.final_norm = RMSNorm() + self.lm_head = ( + None + if h.tie_embeddings + else CastedLinear(h.model_dim, h.vocab_size, bias=False) + ) + if self.lm_head is not None: + self.lm_head._zero_init = True + if h.xsa_last_n > 0: + for i in range(max(0, h.num_layers - h.xsa_last_n), h.num_layers): + self.blocks[i].attn.use_xsa = True + self.looping_active = False + if h.num_loops > 0: + loop_seg = list(range(h.loop_start, h.loop_end + 1)) + all_indices = list(range(h.loop_start)) + for _ in range(h.num_loops + 1): + all_indices.extend(loop_seg) + all_indices.extend(range(h.loop_end + 1, h.num_layers)) + num_enc = len(all_indices) // 2 + self.encoder_indices = all_indices[:num_enc] + self.decoder_indices = all_indices[num_enc:] + else: + self.encoder_indices = list(range(self.num_encoder_layers)) + self.decoder_indices = list(range(self.num_encoder_layers, h.num_layers)) + self.num_skip_weights = min( + len(self.encoder_indices), len(self.decoder_indices) + ) + self.skip_weights = nn.Parameter( + torch.ones(self.num_skip_weights, h.model_dim, dtype=torch.float32) + ) + self.skip_gates = ( + nn.Parameter( + torch.zeros(self.num_skip_weights, h.model_dim, dtype=torch.float32) + ) + if h.skip_gates_enabled + else None + ) + self.parallel_start_layer = h.parallel_start_layer + self.parallel_final_lane = h.parallel_final_lane.lower() + self.parallel_post_lambdas = nn.Parameter( + torch.ones(h.num_layers, 2, 2, dtype=torch.float32) + ) + self.parallel_resid_lambdas = nn.Parameter( + torch.full((h.num_layers, 2), 1.1, dtype=torch.float32) + ) + # SmearGate (PR #1667 / modded-nanogpt @classiclarryd): + # x_t <- x_t + lam * sigmoid(W * x_t[:gate_window]) * x_{t-1}. + # Per-token forward-1 smear of the embedding lane. W zero-init + lam=0 -> + # transparent at init. Uses CastedLinear so restore_fp32_params handles dtype. + self.smear_gate_enabled = h.smear_gate_enabled + if self.smear_gate_enabled: + self.smear_window = h.gate_window + self.smear_gate = CastedLinear(self.smear_window, 1, bias=False) + self.smear_gate._zero_init = True + self.smear_lambda = nn.Parameter(torch.zeros(1, dtype=torch.float32)) + self._init_weights() + + def _init_weights(self): + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) + nn.init.zeros_(self.qo_bank.data[n + i]) + self.qo_bank.data[n + i].mul_(proj_scale) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) + for i in range(n): + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) + nn.init.zeros_(self.mlp_down_bank.data[i]) + self.mlp_down_bank.data[i].mul_(proj_scale) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif ( + module.weight.ndim == 2 + and module.weight.shape[0] >= 64 + and module.weight.shape[1] >= 64 + ): + nn.init.orthogonal_(module.weight, gain=1.0) + + def _bank_weights(self, i): + n = self.num_layers + return ( + self.qo_bank[i], + self.kv_bank[i], + self.kv_bank[n + i], + self.qo_bank[n + i], + self.mlp_up_bank[i], + self.mlp_down_bank[i], + ) + + def _parallel_block( + self, block_idx, lane0, lane1, x0, + q_w, k_w, v_w, out_w, up_w, down_w, + cu_seqlens=None, max_seqlen=0, + ): + block = self.blocks[block_idx] + mix = block.resid_mix.to(dtype=lane0.dtype) + attn_read = mix[0][None, None, :] * lane0 + mix[1][None, None, :] * x0 + attn_out = block.attn( + block.attn_norm(attn_read) * block.ln_scale_factor, + q_w, k_w, v_w, out_w, + cu_seqlens=cu_seqlens, max_seqlen=max_seqlen, + ) + attn_out = block.attn_scale.to(dtype=attn_out.dtype)[None, None, :] * attn_out + mlp_read = lane1 + mlp_out = block.mlp_scale.to(dtype=lane1.dtype)[None, None, :] * block.mlp( + block.mlp_norm(mlp_read) * block.ln_scale_factor, up_w, down_w + ) + attn_resid = self.parallel_resid_lambdas[block_idx, 0].to(dtype=lane0.dtype) + attn_post = self.parallel_post_lambdas[block_idx, 0].to(dtype=lane0.dtype) + mlp_resid = self.parallel_resid_lambdas[block_idx, 1].to(dtype=lane0.dtype) + mlp_post = self.parallel_post_lambdas[block_idx, 1].to(dtype=lane0.dtype) + lane0 = attn_resid * lane0 + attn_post[0] * attn_out + mlp_post[0] * mlp_out + lane1 = mlp_resid * lane1 + attn_post[1] * attn_out + mlp_post[1] * mlp_out + return lane0, lane1 + + def _final_parallel_hidden(self, lane0, lane1): + if self.parallel_final_lane == "mlp": + return lane1 + if self.parallel_final_lane == "attn": + return lane0 + return 0.5 * (lane0 + lane1) + + def _forward_hidden(self, input_ids, cu_seqlens=None, max_seqlen=0): + """Run the encoder/decoder stack to the final RMSNorm; returns pre-projection hidden. + Shared by eval (softcap+projection via forward_logits) and train (fused CE path).""" + x = self.tok_emb(input_ids) + # SmearGate (PR #1667). lam=0 + W=0 -> identity at init. + # Cross-doc leak fix: zero the prev-token smear at any position whose current token + # is BOS, so the BOS embedding starting doc N+1 in a packed stream is not + # contaminated by doc N's last token (audited issue on PR#1797 base). + if self.smear_gate_enabled: + sl = self.smear_lambda.to(dtype=x.dtype) + gate_in = x[:, 1:, : self.smear_window].contiguous() + g = sl * torch.sigmoid(self.smear_gate(gate_in)) + not_bos = (input_ids[:, 1:] != BOS_ID).to(x.dtype).unsqueeze(-1) + x = torch.cat([x[:, :1], x[:, 1:] + g * x[:, :-1] * not_bos], dim=1) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips = [] + enc_iter = ( + self.encoder_indices + if self.looping_active + else range(self.num_encoder_layers) + ) + dec_iter = ( + self.decoder_indices + if self.looping_active + else range( + self.num_encoder_layers, + self.num_encoder_layers + self.num_decoder_layers, + ) + ) + for i in enc_iter: + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + x = self.blocks[i](x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + skips.append(x) + psl = self.parallel_start_layer + lane0 = None + lane1 = None + for skip_idx, i in enumerate(dec_iter): + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + if i >= psl and psl > 0: + if lane0 is None: + lane0 = x + lane1 = x + if skip_idx < self.num_skip_weights and skips: + skip = skips.pop() + w = self.skip_weights[skip_idx].to(dtype=lane0.dtype)[None, None, :] + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=lane0.dtype))[None, None, :] + lane0 = torch.lerp(w * skip, lane0, g) + else: + lane0 = lane0 + w * skip + lane0, lane1 = self._parallel_block( + i, lane0, lane1, x0, q_w, k_w, v_w, out_w, up_w, down_w, + cu_seqlens=cu_seqlens, max_seqlen=max_seqlen, + ) + else: + if skip_idx < self.num_skip_weights and skips: + scaled_skip = ( + self.skip_weights[skip_idx].to(dtype=x.dtype)[None, None, :] + * skips.pop() + ) + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=x.dtype))[None, None, :] + x = torch.lerp(scaled_skip, x, g) + else: + x = x + scaled_skip + x = self.blocks[i](x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + if lane0 is not None: + x = self._final_parallel_hidden(lane0, lane1) + x = self.final_norm(x) + return x + + def _project_logits(self, hidden): + if self.tie_embeddings: + return F.linear(hidden, self.tok_emb.weight) + return self.lm_head(hidden) + + def forward_logits(self, input_ids, cu_seqlens=None, max_seqlen=0): + hidden = self._forward_hidden(input_ids, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + logits_proj = self._project_logits(hidden) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + def forward(self, input_ids, target_ids, cu_seqlens=None, max_seqlen=0): + hidden = self._forward_hidden(input_ids, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + logits_proj = self._project_logits(hidden) + flat_targets = target_ids.reshape(-1) + # Fused softcapped-CE kernel (training path only). Applies softcap inside the + # Triton kernel; takes pre-softcap logits_proj. Non-fused path matches stock + # PR-1736 numerics exactly (softcap in fp32, then F.cross_entropy on fp32). + if self.fused_ce_enabled: + return softcapped_cross_entropy( + logits_proj.reshape(-1, logits_proj.size(-1)), + flat_targets, + self.logit_softcap, + reduction="mean", + ) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + flat_targets, + reduction="mean", + ) + + def forward_ttt(self, input_ids, target_ids, lora): + x = self.tok_emb(input_ids) + # SmearGate on the TTT path — same inline compute as forward_logits. + # Cross-doc leak fix: see _forward_hidden comment. + if self.smear_gate_enabled: + sl = self.smear_lambda.to(dtype=x.dtype) + gate_in = x[:, 1:, : self.smear_window].contiguous() + g = sl * torch.sigmoid(self.smear_gate(gate_in)) + not_bos = (input_ids[:, 1:] != BOS_ID).to(x.dtype).unsqueeze(-1) + x = torch.cat([x[:, :1], x[:, 1:] + g * x[:, :-1] * not_bos], dim=1) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips = [] + enc_iter = ( + self.encoder_indices + if self.looping_active + else list(range(self.num_encoder_layers)) + ) + dec_iter = ( + self.decoder_indices + if self.looping_active + else list( + range( + self.num_encoder_layers, + self.num_encoder_layers + self.num_decoder_layers, + ) + ) + ) + slot = 0 + for i in enc_iter: + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + x = self._block_with_lora(self.blocks[i], x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w) + slot += 1 + skips.append(x) + psl = self.parallel_start_layer + lane0 = None + lane1 = None + for skip_idx, i in enumerate(dec_iter): + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + if i >= psl and psl > 0: + if lane0 is None: + lane0 = x + lane1 = x + if skip_idx < self.num_skip_weights and skips: + skip = skips.pop() + w = self.skip_weights[skip_idx].to(dtype=lane0.dtype)[None, None, :] + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=lane0.dtype))[None, None, :] + lane0 = torch.lerp(w * skip, lane0, g) + else: + lane0 = lane0 + w * skip + lane0, lane1 = self._parallel_block_with_lora( + i, lane0, lane1, x0, lora, slot, + q_w, k_w, v_w, out_w, up_w, down_w, + ) + else: + if skip_idx < self.num_skip_weights and skips: + scaled_skip = ( + self.skip_weights[skip_idx].to(dtype=x.dtype)[None, None, :] + * skips.pop() + ) + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=x.dtype))[None, None, :] + x = torch.lerp(scaled_skip, x, g) + else: + x = x + scaled_skip + x = self._block_with_lora(self.blocks[i], x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w) + slot += 1 + if lane0 is not None: + x = self._final_parallel_hidden(lane0, lane1) + x = self.final_norm(x) + if self.tie_embeddings: + logits = F.linear(x, self.tok_emb.weight) + else: + logits = self.lm_head(x) + logits = logits + lora.lm_head_lora(x) + logits = self.logit_softcap * torch.tanh(logits / self.logit_softcap) + bsz, sl, V = logits.shape + return F.cross_entropy( + logits.float().reshape(-1, V), target_ids.reshape(-1), reduction="none" + ).reshape(bsz, sl) + + def _block_with_lora(self, block, x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w): + mix = block.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = block.attn_norm(x_in) * block.ln_scale_factor + attn = block.attn + bsz, seqlen, dim = n.shape + # Keep raw Q for AttnOutGate src='q' (matches forward path semantics). + q_raw = F.linear(n, q_w.to(n.dtype)) + lora.q_loras[slot](n) + q = q_raw.reshape(bsz, seqlen, attn.num_heads, attn.head_dim) + k = F.linear(n, k_w.to(n.dtype)) + if lora.k_loras is not None: + k = k + lora.k_loras[slot](n) + k = k.reshape(bsz, seqlen, attn.num_kv_heads, attn.head_dim) + v = (F.linear(n, v_w.to(n.dtype)) + lora.v_loras[slot](n)).reshape( + bsz, seqlen, attn.num_kv_heads, attn.head_dim + ) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = attn.rotary(seqlen, n.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, attn.rope_dims) + k = apply_rotary_emb(k, cos, sin, attn.rope_dims) + q = q * attn.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if attn.use_xsa: + y = attn._xsa_efficient(y, v) + # AttnOutGate (TTT path) — inline + .contiguous() barrier, same as the eval path. + if attn.attn_out_gate: + gate_src = q_raw if attn.attn_out_gate_src == "q" else n + gate_in = gate_src[..., : attn.gate_window].contiguous() + g = 2.0 * torch.sigmoid(attn.attn_gate_proj(gate_in)) + y = y * g[..., None] + # Gated Attention (TTT path). Gate input is n (post-norm block input), same + # as eval path. .to(n.dtype) on fp32 param before bf16 broadcast. + if attn.gated_attn: + n_c = n.contiguous() + g = torch.sigmoid(F.linear(n_c, attn.attn_gate_w.to(n.dtype))) + y = y * g[..., None] + # Sparse attention head-output gate (TTT path) — must match the eval path in + # forward() exactly, else training (which applied the gate) and TTT eval (which + # skipped it) produce mismatched representations and catastrophic BPB regression. + if attn.sparse_attn_gate: + gate_in = n[..., : attn.gate_window].contiguous() + g = torch.sigmoid( + attn.sparse_attn_gate_scale + * F.linear(gate_in, attn.attn_gate_w.to(n.dtype)) + ) + y = y * g[..., None] + y = y.reshape(bsz, seqlen, dim) + attn_out = F.linear(y, out_w.to(n.dtype)) + if lora.o_loras is not None: + attn_out = attn_out + lora.o_loras[slot](n) + x_out = x_in + block.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + mlp_n = block.mlp_norm(x_out) * block.ln_scale_factor + mlp_out = block.mlp(mlp_n, up_w, down_w) + if lora.mlp_loras is not None: + mlp_out = mlp_out + lora.mlp_loras[slot](mlp_n) + x_out = x_out + block.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * mlp_out + return x_out + + def _parallel_block_with_lora( + self, block_idx, lane0, lane1, x0, lora, slot, + q_w, k_w, v_w, out_w, up_w, down_w, + ): + block = self.blocks[block_idx] + mix = block.resid_mix.to(dtype=lane0.dtype) + attn_read = mix[0][None, None, :] * lane0 + mix[1][None, None, :] * x0 + n = block.attn_norm(attn_read) * block.ln_scale_factor + attn = block.attn + bsz, seqlen, dim = n.shape + q_raw = F.linear(n, q_w.to(n.dtype)) + lora.q_loras[slot](n) + q = q_raw.reshape(bsz, seqlen, attn.num_heads, attn.head_dim) + k = F.linear(n, k_w.to(n.dtype)) + if lora.k_loras is not None: + k = k + lora.k_loras[slot](n) + k = k.reshape(bsz, seqlen, attn.num_kv_heads, attn.head_dim) + v = (F.linear(n, v_w.to(n.dtype)) + lora.v_loras[slot](n)).reshape( + bsz, seqlen, attn.num_kv_heads, attn.head_dim + ) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = attn.rotary(seqlen, n.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, attn.rope_dims) + k = apply_rotary_emb(k, cos, sin, attn.rope_dims) + q = q * attn.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if attn.use_xsa: + y = attn._xsa_efficient(y, v) + # AttnOutGate (TTT parallel path) — inline + .contiguous() barrier. + if attn.attn_out_gate: + gate_src = q_raw if attn.attn_out_gate_src == "q" else n + gate_in = gate_src[..., : attn.gate_window].contiguous() + g = 2.0 * torch.sigmoid(attn.attn_gate_proj(gate_in)) + y = y * g[..., None] + # Gated Attention (TTT parallel path). Gate input is n (post-norm block input). + if attn.gated_attn: + n_c = n.contiguous() + g = torch.sigmoid(F.linear(n_c, attn.attn_gate_w.to(n.dtype))) + y = y * g[..., None] + # Sparse attention head-output gate (TTT parallel path) — must match the + # eval path in forward() to keep train/eval semantics in sync. + if attn.sparse_attn_gate: + gate_in = n[..., : attn.gate_window].contiguous() + g = torch.sigmoid( + attn.sparse_attn_gate_scale + * F.linear(gate_in, attn.attn_gate_w.to(n.dtype)) + ) + y = y * g[..., None] + y = y.reshape(bsz, seqlen, dim) + attn_out = F.linear(y, out_w.to(n.dtype)) + if lora.o_loras is not None: + attn_out = attn_out + lora.o_loras[slot](n) + attn_out = block.attn_scale.to(dtype=attn_out.dtype)[None, None, :] * attn_out + mlp_read = lane1 + mlp_n = block.mlp_norm(mlp_read) * block.ln_scale_factor + mlp_out = block.mlp(mlp_n, up_w, down_w) + if lora.mlp_loras is not None: + mlp_out = mlp_out + lora.mlp_loras[slot](mlp_n) + mlp_out = block.mlp_scale.to(dtype=lane1.dtype)[None, None, :] * mlp_out + attn_resid = self.parallel_resid_lambdas[block_idx, 0].to(dtype=lane0.dtype) + attn_post = self.parallel_post_lambdas[block_idx, 0].to(dtype=lane0.dtype) + mlp_resid = self.parallel_resid_lambdas[block_idx, 1].to(dtype=lane0.dtype) + mlp_post = self.parallel_post_lambdas[block_idx, 1].to(dtype=lane0.dtype) + lane0 = attn_resid * lane0 + attn_post[0] * attn_out + mlp_post[0] * mlp_out + lane1 = mlp_resid * lane1 + attn_post[1] * attn_out + mlp_post[1] * mlp_out + return lane0, lane1 + + +class BatchedLinearLoRA(nn.Module): + # PR-1767: rank-scaled output (alpha/rank), like standard LoRA. Decouples + # effective magnitude from rank so changing rank does not change LR scale. + _ALPHA = float(os.environ.get("TTT_LORA_ALPHA", "144")) + # PR-1767: optionally keep A warm across per-doc resets (only B is zeroed). + # Accumulates useful feature directions across documents within a TTT phase. + _WARM_START_A = bool(int(os.environ.get("TTT_WARM_START_A", "1"))) + + def __init__(self, bsz, in_features, out_features, rank): + super().__init__() + self._bound = 1.0 / math.sqrt(in_features) + self._scale = self._ALPHA / rank + self.A = nn.Parameter( + torch.empty(bsz, rank, in_features).uniform_(-self._bound, self._bound) + ) + self.B = nn.Parameter(torch.zeros(bsz, out_features, rank)) + + def reset(self): + with torch.no_grad(): + if not self._WARM_START_A: + self.A.uniform_(-self._bound, self._bound) + self.B.zero_() + + def forward(self, x): + return ((x @ self.A.transpose(1, 2)) @ self.B.transpose(1, 2)) * self._scale + + +class BatchedTTTLoRA(nn.Module): + def __init__(self, bsz, model, rank, k_lora=True, mlp_lora=True, o_lora=True): + super().__init__() + self.bsz = bsz + dim = model.qo_bank.shape[-1] + vocab = model.tok_emb.num_embeddings + if getattr(model, "looping_active", False): + num_slots = len(model.encoder_indices) + len(model.decoder_indices) + else: + num_slots = len(model.blocks) + kv_dim = model.blocks[0].attn.num_kv_heads * ( + dim // model.blocks[0].attn.num_heads + ) + embed_dim = model.tok_emb.embedding_dim + self.lm_head_lora = BatchedLinearLoRA(bsz, embed_dim, vocab, rank) + self.q_loras = nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + self.v_loras = nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, kv_dim, rank) for _ in range(num_slots)] + ) + self.k_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, kv_dim, rank) for _ in range(num_slots)] + ) + if k_lora + else None + ) + self.mlp_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + if mlp_lora + else None + ) + self.o_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + if o_lora + else None + ) + + def reset(self): + with torch.no_grad(): + self.lm_head_lora.reset() + for loras in [self.q_loras, self.v_loras, self.k_loras, + self.mlp_loras, self.o_loras]: + if loras is not None: + for lora in loras: + lora.reset() + + +# Polar Express per-iteration minimax Newton-Schulz coefficients (PR #1344). +# Replaces the fixed (3.4445, -4.775, 2.0315) coefficients of stock Muon. +# Applied at backend_steps=5 — taking more than 5 iterations from this list +# falls back to the final (converged) tuple via the slice guard below. +_PE_COEFFS = ( + (8.156554524902461, -22.48329292557795, 15.878769915207462), + (4.042929935166739, -2.808917465908714, 0.5000178451051316), + (3.8916678022926607, -2.772484153217685, 0.5060648178503393), + (3.285753657755655, -2.3681294933425376, 0.46449024233003106), + (2.3465413258596377, -1.7097828382687081, 0.42323551169305323), +) + + +@torch.compile +def zeropower_via_newtonschulz5(G, steps=10, eps=1e-07): + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + coeffs = _PE_COEFFS[:steps] if steps <= len(_PE_COEFFS) else _PE_COEFFS + for a, b, c in coeffs: + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + + +class Muon(torch.optim.Optimizer): + def __init__( + self, + params, + lr, + momentum, + backend_steps, + nesterov=True, + weight_decay=0.0, + row_normalize=False, + ): + super().__init__( + params, + dict( + lr=lr, + momentum=momentum, + backend_steps=backend_steps, + nesterov=nesterov, + weight_decay=weight_decay, + row_normalize=row_normalize, + ), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + "p": p, + "B": B, + "padded_grad": torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + "shard": torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + "shard_mom": torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + "full_update": torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + "scale": max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + self._bank_meta.sort(key=lambda m: -m["p"].numel()) + self._built = True + + def launch_reduce_scatters(self): + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m["p"] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m["padded_grad"] + pg[: m["B"]].copy_(p.grad) + fut = dist.reduce_scatter_tensor( + m["shard"], pg, op=dist.ReduceOp.AVG, async_op=True + ) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + if not self._built: + self._build() + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + row_normalize = group.get("row_normalize", False) + prev_ag_handle = None + prev_m = None + sharded = self._distributed and hasattr(self, "_rs_futures") + for idx, m in enumerate(self._bank_meta): + p = m["p"] + if p.grad is None: + continue + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m["p"] + upd = prev_m["full_update"][: prev_m["B"]] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd, alpha=-lr * prev_m["scale"]) + if sharded and self._rs_futures[idx] is not None: + self._rs_futures[idx].wait() + g = m["shard"] + buf = m["shard_mom"] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + if row_normalize: + rn = update.float().norm(dim=-1, keepdim=True).clamp_min(1e-07) + update = update / rn.to(update.dtype) + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m["full_update"], update, async_op=True + ) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update, alpha=-lr * m["scale"]) + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m["p"] + upd = prev_m["full_update"][: prev_m["B"]] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd, alpha=-lr * prev_m["scale"]) + if hasattr(self, "_rs_futures"): + del self._rs_futures + return loss + + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,skip_gates,parallel_post_lambdas,parallel_resid_lambdas,attn_gate_proj,attn_gate_w,smear_gate,smear_lambda", + ).split(",") + if pattern +) + + +PACKED_REPLICATED_GRAD_MAX_NUMEL = 1 << 15 + + +class Optimizers: + def __init__(self, h, base_model): + matrix_params = [ + base_model.qo_bank, + base_model.kv_bank, + base_model.mlp_up_bank, + base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for (name, p) in block_named_params + if p.ndim < 2 + or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + if base_model.skip_gates is not None and base_model.skip_gates.numel() > 0: + scalar_params.append(base_model.skip_gates) + if base_model.parallel_post_lambdas is not None: + scalar_params.append(base_model.parallel_post_lambdas) + if base_model.parallel_resid_lambdas is not None: + scalar_params.append(base_model.parallel_resid_lambdas) + # SmearGate params live on GPT root (not in .blocks), so add them by hand. + # Both are tiny (gate_window scalars + 1 lambda). Optimized via scalar Adam. + if getattr(base_model, "smear_gate_enabled", False): + scalar_params.append(base_model.smear_gate.weight) + scalar_params.append(base_model.smear_lambda) + token_lr = h.tied_embed_lr if h.tie_embeddings else h.embed_lr + tok_params = [ + {"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr} + ] + self.optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(h.beta1, h.beta2), + eps=h.adam_eps, + weight_decay=h.embed_wd, + fused=True, + ) + self.optimizer_muon = Muon( + matrix_params, + lr=h.matrix_lr, + momentum=h.muon_momentum, + backend_steps=h.muon_backend_steps, + weight_decay=h.muon_wd, + row_normalize=h.muon_row_normalize, + ) + for group in self.optimizer_muon.param_groups: + group["base_lr"] = h.matrix_lr + self.optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": h.scalar_lr, "base_lr": h.scalar_lr}], + betas=(h.beta1, h.beta2), + eps=h.adam_eps, + weight_decay=h.adam_wd, + fused=True, + ) + self.optimizers = [ + self.optimizer_tok, + self.optimizer_muon, + self.optimizer_scalar, + ] + self.replicated_params = list(tok_params[0]["params"]) + self.replicated_params.extend(scalar_params) + self.replicated_large_params = [] + self.replicated_packed_params = [] + for p in self.replicated_params: + if p.numel() <= PACKED_REPLICATED_GRAD_MAX_NUMEL: + self.replicated_packed_params.append(p) + else: + self.replicated_large_params.append(p) + self._aux_stream = torch.cuda.Stream() + + def __iter__(self): + return iter(self.optimizers) + + def zero_grad_all(self): + for opt in self.optimizers: + opt.zero_grad(set_to_none=True) + + def _all_reduce_packed_grads(self): + grads_by_key = collections.defaultdict(list) + for p in self.replicated_packed_params: + if p.grad is not None: + grads_by_key[(p.grad.device, p.grad.dtype)].append(p.grad) + for grads in grads_by_key.values(): + flat = torch.empty( + sum(g.numel() for g in grads), + device=grads[0].device, + dtype=grads[0].dtype, + ) + offset = 0 + for g in grads: + n = g.numel() + flat[offset : offset + n].copy_(g.contiguous().view(-1)) + offset += n + dist.all_reduce(flat, op=dist.ReduceOp.AVG) + offset = 0 + for g in grads: + n = g.numel() + g.copy_(flat[offset : offset + n].view_as(g)) + offset += n + + def step(self, distributed=False): + self.optimizer_muon.launch_reduce_scatters() + if distributed: + reduce_handles = [ + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG, async_op=True) + for p in self.replicated_large_params + if p.grad is not None + ] + self._all_reduce_packed_grads() + for handle in reduce_handles: + handle.wait() + self._aux_stream.wait_stream(torch.cuda.current_stream()) + with torch.cuda.stream(self._aux_stream): + self.optimizer_tok.step() + self.optimizer_scalar.step() + self.optimizer_muon.step() + torch.cuda.current_stream().wait_stream(self._aux_stream) + self.zero_grad_all() + + +def restore_fp32_params(model): + for module in model.modules(): + if isinstance(module, CastedLinear): + module.float() + for name, param in model.named_parameters(): + if ( + param.ndim < 2 + or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ) and param.dtype != torch.float32: + param.data = param.data.float() + if hasattr(model, "qo_bank") and model.qo_bank is not None: + model.qo_bank.data = model.qo_bank.data.float() + model.kv_bank.data = model.kv_bank.data.float() + model.mlp_up_bank.data = model.mlp_up_bank.data.float() + model.mlp_down_bank.data = model.mlp_down_bank.data.float() + + +def collect_hessians(model, train_loader, h, device, n_calibration_batches=64): + hessians = {} + act_sumsq = {} + act_counts = {} + hooks = [] + for i, block in enumerate(model.blocks): + block.attn._calib = True + block.mlp._calib = True + block.mlp.use_fused = False + + def make_attn_hook(layer_idx): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + x_sq = x.square().sum(dim=0) + x_count = x.shape[0] + for suffix in ["c_q", "c_k", "c_v"]: + name = f"blocks.{layer_idx}.attn.{suffix}.weight" + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + if name not in act_sumsq: + act_sumsq[name] = torch.zeros( + x.shape[1], dtype=torch.float32, device=device + ) + act_counts[name] = 0 + act_sumsq[name] += x_sq + act_counts[name] += x_count + y = module._last_proj_input + if y is not None: + y = y.float() + if y.ndim == 3: + y = y.reshape(-1, y.shape[-1]) + name = f"blocks.{layer_idx}.attn.proj.weight" + if name not in hessians: + hessians[name] = torch.zeros( + y.shape[1], y.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(y.T, y) + if name not in act_sumsq: + act_sumsq[name] = torch.zeros( + y.shape[1], dtype=torch.float32, device=device + ) + act_counts[name] = 0 + act_sumsq[name] += y.square().sum(dim=0) + act_counts[name] += y.shape[0] + return hook_fn + + def make_mlp_hook(layer_idx): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + name = f"blocks.{layer_idx}.mlp.fc.weight" + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + if name not in act_sumsq: + act_sumsq[name] = torch.zeros( + x.shape[1], dtype=torch.float32, device=device + ) + act_counts[name] = 0 + act_sumsq[name] += x.square().sum(dim=0) + act_counts[name] += x.shape[0] + h_act = module._last_down_input + if h_act is not None: + h_act = h_act.float() + if h_act.ndim == 3: + h_act = h_act.reshape(-1, h_act.shape[-1]) + name = f"blocks.{layer_idx}.mlp.proj.weight" + if name not in hessians: + hessians[name] = torch.zeros( + h_act.shape[1], h_act.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(h_act.T, h_act) + if name not in act_sumsq: + act_sumsq[name] = torch.zeros( + h_act.shape[1], dtype=torch.float32, device=device + ) + act_counts[name] = 0 + act_sumsq[name] += h_act.square().sum(dim=0) + act_counts[name] += h_act.shape[0] + return hook_fn + + for i, block in enumerate(model.blocks): + hooks.append(block.attn.register_forward_hook(make_attn_hook(i))) + hooks.append(block.mlp.register_forward_hook(make_mlp_hook(i))) + + # Hessian hooks for embedding factorization projection layers + def make_linear_input_hook(weight_name): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if weight_name not in hessians: + hessians[weight_name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[weight_name].addmm_(x.T, x) + return hook_fn + + if model.tie_embeddings: + hook_module = model.final_norm + + def make_output_hook(name): + def hook_fn(module, inp, out): + x = out.detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + if name not in act_sumsq: + act_sumsq[name] = torch.zeros( + x.shape[1], dtype=torch.float32, device=device + ) + act_counts[name] = 0 + act_sumsq[name] += x.square().sum(dim=0) + act_counts[name] += x.shape[0] + return hook_fn + + hooks.append( + hook_module.register_forward_hook(make_output_hook("tok_emb.weight")) + ) + model.eval() + with torch.no_grad(): + for _ in range(n_calibration_batches): + x, _ = train_loader.next_batch(h.train_batch_tokens, h.grad_accum_steps) + model.forward_logits(x) + for hook in hooks: + hook.remove() + for i, block in enumerate(model.blocks): + block.attn._calib = False + block.mlp._calib = False + block.mlp.use_fused = True + for name in hessians: + hessians[name] = hessians[name].cpu() / n_calibration_batches + act_stats = {} + for name, sumsq in act_sumsq.items(): + count = max(act_counts.get(name, 0), 1) + act_stats[name] = (sumsq / count).sqrt().cpu() + return hessians, act_stats + + +def gptq_quantize_weight( + w, + H, + clip_sigmas=3.0, + clip_range=63, + block_size=128, + protect_groups=None, + group_size=None, + protect_clip_range=None, +): + W_orig = w.float().clone() + rows, cols = W_orig.shape + H = H.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = 0.01 * H.diag().mean() + H.diagonal().add_(damp) + perm = torch.argsort(H.diag(), descending=True) + invperm = torch.argsort(perm) + W_perm = W_orig[:, perm].clone() + W_perm[:, dead[perm]] = 0 + H = H[perm][:, perm] + Hinv = torch.cholesky_inverse(torch.linalg.cholesky(H)) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + row_std = W_orig.std(dim=1) + s = (clip_sigmas * row_std / clip_range).clamp_min(1e-10).to(torch.float16) + sf = s.float() + protect_meta = None + protect_mask_perm = None + s_hi = None + sf_hi = None + if ( + protect_groups + and group_size is not None + and protect_clip_range is not None + and protect_clip_range > clip_range + ): + protect_mask = torch.zeros(cols, dtype=torch.bool) + starts = [] + for (start, end) in protect_groups: + if start < 0 or end > cols or end <= start: + continue + protect_mask[start:end] = True + starts.append(start) + if starts: + protect_mask_perm = protect_mask[perm] + s_hi = (clip_sigmas * row_std / protect_clip_range).clamp_min(1e-10).to( + torch.float16 + ) + sf_hi = s_hi.float() + protect_meta = { + "starts": torch.tensor(starts, dtype=torch.int16), + "size": int(group_size), + "s_hi": s_hi, + } + Q = torch.zeros(rows, cols, dtype=torch.int8) + W_work = W_perm.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W_work[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros(rows, i2 - i1) + for j in range(i2 - i1): + w_col = W_block[:, j] + d = Hinv_block[j, j] + if protect_mask_perm is not None and bool(protect_mask_perm[i1 + j]): + q_col = torch.clamp( + torch.round(w_col / sf_hi), + -protect_clip_range, + protect_clip_range, + ) + w_recon = q_col.float() * sf_hi + else: + q_col = torch.clamp(torch.round(w_col / sf), -clip_range, clip_range) + w_recon = q_col.float() * sf + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - w_recon) / d + Err[:, j] = err + W_block[:, j:] -= err.unsqueeze(1) * Hinv_block[j, j:].unsqueeze(0) + if i2 < cols: + W_work[:, i2:] -= Err @ Hinv[i1:i2, i2:] + return Q[:, invperm], s, protect_meta + + +def _quantize_gate_int8_row(w): + # Symmetric int8-per-row quantization for small gate tensors. w shape + # (R, C) -> (R,) scales in fp16, int8 values in [-127, 127]. Single scale + # per row keeps accuracy high while halving storage vs fp16. + W = w.float().contiguous() + row_max = W.abs().amax(dim=1).clamp_min(1e-10) + s = (row_max / 127.0).to(torch.float16) + sf = s.float().view(-1, 1) + q = torch.clamp(torch.round(W / sf), -127, 127).to(torch.int8) + return q, s + + +def _lqer_pack(A, B, bits): + rng = 2 ** (bits - 1) - 1 + sA = (A.abs().amax(dim=1).clamp_min(1e-10) / rng).to(torch.float16) + sB = (B.abs().amax(dim=1).clamp_min(1e-10) / rng).to(torch.float16) + qA = torch.clamp(torch.round(A / sA.float().view(-1, 1)), -rng, rng).to(torch.int8) + qB = torch.clamp(torch.round(B / sB.float().view(-1, 1)), -rng, rng).to(torch.int8) + return qA, sA, qB, sB + + +def _lqer_pack_asym(A, B, g=64): + # A: INT2 per-matrix scalar (signed [-2,1], scale = |A|max/1.5). + sA = (A.abs().amax().clamp_min(1e-10) / 1.5).to(torch.float16) + qA = torch.clamp(torch.round(A / sA.float()), -2, 1).to(torch.int8) + # B: INT4 groupwise g over flattened B (signed [-8,7], per-group scale). + Bf = B.reshape(-1, g) + Bmax = Bf.abs().amax(dim=-1, keepdim=True).clamp_min(1e-10) + sB = (Bmax / 7.5).to(torch.float16).reshape(-1) + qB = torch.clamp(torch.round(Bf / sB.float().reshape(-1, 1)), -8, 7).to( + torch.int8 + ).reshape(B.shape) + return qA, sA, qB, sB + + +def _lqer_fit_quantized(E, h): + U, S, Vh = torch.linalg.svd(E, full_matrices=False) + r = min(h.lqer_rank, S.numel()) + if r <= 0: + return None + A = (U[:, :r] * S[:r]).contiguous() + B = Vh[:r, :].contiguous() + asym_on = bool(getattr(h, "lqer_asym_enabled", False)) + asym_g = int(getattr(h, "lqer_asym_group", 64)) + if asym_on and B.numel() % asym_g == 0: + qA, sA, qB, sB = _lqer_pack_asym(A, B, asym_g) + A_hat = qA.float() * float(sA) + g_sz = qB.numel() // sB.numel() + B_hat = (qB.reshape(-1, g_sz).float() * sB.float().view(-1, 1)).reshape( + qB.shape + ) + return { + "kind": "asym", + "qA": qA, + "sA": sA, + "qB": qB, + "sB": sB, + "delta": A_hat @ B_hat, + } + qA, sA, qB, sB = _lqer_pack(A, B, h.lqer_factor_bits) + A_hat = qA.float() * sA.float().view(-1, 1) + B_hat = qB.float() * sB.float().view(-1, 1) + return { + "kind": "sym", + "qA": qA, + "sA": sA, + "qB": qB, + "sB": sB, + "delta": A_hat @ B_hat, + } + + +def _awq_lite_group_candidates(w, act_rms, group_size): + cols = w.shape[1] + n_groups = cols // group_size + if n_groups <= 0: + return [] + weight_score = w.float().abs().mean(dim=0) + saliency = act_rms.float() * weight_score + cands = [] + for gi in range(n_groups): + start = gi * group_size + end = start + group_size + score = float(saliency[start:end].sum()) + cands.append((score, start, end)) + return cands + + +def gptq_mixed_quantize(state_dict, hessians, act_stats, h): + result = {} + meta = {} + quant_gate = bool(getattr(h, "gated_attn_quant_gate", False)) + lqer_on = bool(getattr(h, "lqer_enabled", False)) + awq_on = bool(getattr(h, "awq_lite_enabled", False)) + lqer_cands = {} + awq_selected = collections.defaultdict(list) + if awq_on: + awq_cands = [] + for (name, tensor) in state_dict.items(): + t = tensor.detach().cpu().contiguous() + if t.is_floating_point() and t.numel() > 65536 and name in act_stats: + bits = h.embed_bits if "tok_emb" in name else h.matrix_bits + if bits < h.awq_lite_bits: + for score, start, end in _awq_lite_group_candidates( + t, act_stats[name], h.awq_lite_group_size + ): + awq_cands.append((score, name, start, end)) + awq_cands.sort(key=lambda x: -x[0]) + for (_score, name, start, end) in awq_cands[: h.awq_lite_group_top_k]: + awq_selected[name].append((start, end)) + for (name, tensor) in state_dict.items(): + t = tensor.detach().cpu().contiguous() + # Dedicated int8-per-row path for attn_gate_w (bypasses both GPTQ and + # fp16 passthrough). Applied BEFORE the numel<=65536 passthrough check + # so the gate tensor is routed here instead of to fp16. + if ( + quant_gate + and t.is_floating_point() + and t.ndim == 2 + and name.endswith(".attn_gate_w") + # Dense GatedAttn: (num_heads, dim) = (8, 512) = 4096. + # Sparse gate: (num_heads, gate_window) = (8, 12) = 96. + # Both need int8-per-row routing; the 1024 lower bound in stock + # PR-1736 presumed dense-only. Widen to catch both. + and 32 <= t.numel() <= 8192 + ): + gq, gs = _quantize_gate_int8_row(t) + result[name + ".gq"] = gq + result[name + ".gs"] = gs + meta[name] = "gate_int8_row" + continue + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough (float16)" + continue + if "tok_emb" in name: + cs = h.embed_clip_sigmas + elif ".mlp." in name: + cs = h.mlp_clip_sigmas + elif ".attn." in name: + cs = h.attn_clip_sigmas + else: + cs = h.matrix_clip_sigmas + bits = h.embed_bits if "tok_emb" in name else h.matrix_bits + clip_range = 2 ** (bits - 1) - 1 + q, s, protect_meta = gptq_quantize_weight( + t, + hessians[name], + clip_sigmas=cs, + clip_range=clip_range, + protect_groups=awq_selected.get(name), + group_size=h.awq_lite_group_size if name in awq_selected else None, + protect_clip_range=(2 ** (h.awq_lite_bits - 1) - 1) + if name in awq_selected + else None, + ) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = f"gptq (int{bits})" + W_q = q.float() * s.float().view(-1, 1) + if protect_meta is not None: + result[name + ".awqg_start"] = protect_meta["starts"] + result[name + ".awqg_s_hi"] = protect_meta["s_hi"] + result[name + ".awqg_size"] = torch.tensor( + protect_meta["size"], dtype=torch.int16 + ) + meta[name] = meta[name] + f"+awqgrpint{h.awq_lite_bits}" + gsz = protect_meta["size"] + for start in protect_meta["starts"].tolist(): + W_q[:, start : start + gsz] = ( + q[:, start : start + gsz].float() + * protect_meta["s_hi"].float().view(-1, 1) + ) + if lqer_on: + # LQER is fit on top of the fully realized GPTQ base, which already + # includes any higher-precision AWQ-protected groups. + scope = str(getattr(h, "lqer_scope", "all")).lower() + scope_ok = ( + scope == "all" + or (scope == "mlp" and ".mlp." in name) + or (scope == "attn" and ".attn." in name) + or (scope == "embed" and "tok_emb" in name) + ) + if scope_ok: + E = t.float() - W_q + err_norm = float(E.norm()) + if err_norm > 0: + lqer_cands[name] = (E, err_norm) + if lqer_on and lqer_cands: + if bool(getattr(h, "lqer_gain_select", False)): + scored = [] + for (name, (E, base_err)) in lqer_cands.items(): + fit = _lqer_fit_quantized(E, h) + if fit is None: + continue + new_err = float((E - fit["delta"]).norm()) + gain = base_err - new_err + if gain > 0: + scored.append((gain, name, fit)) + scored.sort(key=lambda x: -x[0]) + for (_gain, name, fit) in scored[: h.lqer_top_k]: + if fit["kind"] == "asym": + result[name + ".lqA_a"] = fit["qA"] + result[name + ".lqAs_a"] = fit["sA"] + result[name + ".lqB_a"] = fit["qB"] + result[name + ".lqBs_a"] = fit["sB"] + meta[name] = meta[name] + "+lqer_asym" + else: + result[name + ".lqA"] = fit["qA"] + result[name + ".lqAs"] = fit["sA"] + result[name + ".lqB"] = fit["qB"] + result[name + ".lqBs"] = fit["sB"] + meta[name] = meta[name] + "+lqer" + else: + top = sorted(lqer_cands.items(), key=lambda kv: -kv[1][1])[: h.lqer_top_k] + asym_on = bool(getattr(h, "lqer_asym_enabled", False)) + asym_g = int(getattr(h, "lqer_asym_group", 64)) + for (name, (E, _)) in top: + U, S, Vh = torch.linalg.svd(E, full_matrices=False) + r = min(h.lqer_rank, S.numel()) + A = (U[:, :r] * S[:r]).contiguous() + B = Vh[:r, :].contiguous() + if asym_on and B.numel() % asym_g == 0: + qA, sA, qB, sB = _lqer_pack_asym(A, B, asym_g) + result[name + ".lqA_a"] = qA + result[name + ".lqAs_a"] = sA + result[name + ".lqB_a"] = qB + result[name + ".lqBs_a"] = sB + meta[name] = meta[name] + "+lqer_asym" + else: + qA, sA, qB, sB = _lqer_pack(A, B, h.lqer_factor_bits) + result[name + ".lqA"] = qA + result[name + ".lqAs"] = sA + result[name + ".lqB"] = qB + result[name + ".lqBs"] = sB + meta[name] = meta[name] + "+lqer" + categories = collections.defaultdict(set) + for (name, cat) in meta.items(): + short = re.sub("\\.\\d+$", "", re.sub("blocks\\.\\d+", "blocks", name)) + categories[cat].add(short) + log("Quantized weights:") + for cat in sorted(categories): + log(f" {cat}: {', '.join(sorted(categories[cat]))}") + return result, meta + +def dequantize_mixed(result, meta, template_sd): + out = {} + for (name, orig) in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if "passthrough" in info: + t = result[name] + if t.dtype == torch.float16 and orig_dtype in ( + torch.float32, + torch.bfloat16, + ): + t = t.to(orig_dtype) + out[name] = t + continue + if info == "gate_int8_row": + gq = result[name + ".gq"] + gs = result[name + ".gs"] + out[name] = (gq.float() * gs.float().view(-1, 1)).to(orig_dtype) + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + W = q.float() * s.float().view(q.shape[0], *[1] * (q.ndim - 1)) + else: + W = q.float() * float(s.item()) + if "awqgrpint" in info: + starts = result[name + ".awqg_start"].tolist() + s_hi = result[name + ".awqg_s_hi"].float() + gsz = int(result[name + ".awqg_size"].item()) + for start in starts: + W[:, start : start + gsz] = ( + q[:, start : start + gsz].float() * s_hi.view(-1, 1) + ) + if "lqer_asym" in info: + qA_t = result[name + ".lqA_a"] + sA_t = result[name + ".lqAs_a"] + qB_t = result[name + ".lqB_a"] + sB_t = result[name + ".lqBs_a"] + qA = qA_t.float() * float(sA_t) + g_sz = qB_t.numel() // sB_t.numel() + qB = (qB_t.reshape(-1, g_sz).float() * sB_t.float().view(-1, 1)).reshape( + qB_t.shape + ) + W = W + qA @ qB + elif "lqer" in info: + qA = result[name + ".lqA"].float() * result[name + ".lqAs"].float().view(-1, 1) + qB = result[name + ".lqB"].float() * result[name + ".lqBs"].float().view(-1, 1) + W = W + qA @ qB + out[name] = W.to(orig_dtype) + return out + + +_BSHF_MAGIC = b"BSHF" + + +# ── Per-group lrzip compression (ported from PR#1586 via PR#1667/1729) ──────── + +_GROUP_ORDER = [ + "_tok_emb.weight.q", + "attn.c_k.weight.q", "attn.c_q.weight.q", + "attn.c_v.weight.q", "attn.proj.weight.q", + "mlp.fc.weight.q", "mlp.proj.weight.q", +] +_SIMSORT_KEYS = {"_tok_emb.weight.q", "attn.c_q.weight.q", "mlp.fc.weight.q"} +_PACK_MAGIC = b"PGRP" + + +def _similarity_sort_l1(matrix): + import numpy as _np + n = matrix.shape[0] + used = _np.zeros(n, dtype=bool) + order = [0] + used[0] = True + cur = matrix[0].astype(_np.float32) + for _ in range(n - 1): + dists = _np.sum(_np.abs(matrix[~used].astype(_np.float32) - cur), axis=1) + unused = _np.where(~used)[0] + best = unused[_np.argmin(dists)] + order.append(best) + used[best] = True + cur = matrix[best].astype(_np.float32) + return _np.array(order, dtype=_np.uint16) + + +def _lrzip_compress(data, tmpdir, label): + inp = os.path.join(tmpdir, f"{label}.bin") + out = f"{inp}.lrz" + with open(inp, "wb") as f: + f.write(data) + subprocess.run(["lrzip", "-z", "-L", "9", "-o", out, inp], capture_output=True, check=True) + with open(out, "rb") as f: + result = f.read() + os.remove(inp); os.remove(out) + return result + + +def _lrzip_decompress(data, tmpdir, label): + inp = os.path.join(tmpdir, f"{label}.lrz") + out = os.path.join(tmpdir, f"{label}.bin") + with open(inp, "wb") as f: + f.write(data) + subprocess.run(["lrzip", "-d", "-f", "-o", out, inp], capture_output=True, check=True) + with open(out, "rb") as f: + result = f.read() + os.remove(inp); os.remove(out) + return result + + +def _pack_streams(streams): + import struct + n = len(streams) + hdr = _PACK_MAGIC + struct.pack("= 2 + docs.append((start, end - start)) + return docs + + +def _build_ttt_global_batches(doc_entries, h, ascending=False): + batch_size = h.ttt_batch_size + global_doc_entries = sorted(doc_entries, key=lambda x: x[1][1]) + global_batches = [ + global_doc_entries[i : i + batch_size] + for i in range(0, len(global_doc_entries), batch_size) + ] + indexed = list(enumerate(global_batches)) + if not ascending: + indexed.sort(key=lambda ib: -max(dl for _, (_, dl) in ib[1])) + return indexed + + +def _init_batch_counter(path): + with open(path, "wb") as f: + f.write((0).to_bytes(4, "little")) + + +def _claim_next_batch(counter_path, queue_len): + try: + with open(counter_path, "r+b") as f: + fcntl.flock(f, fcntl.LOCK_EX) + idx = int.from_bytes(f.read(4), "little") + f.seek(0) + f.write((idx + 1).to_bytes(4, "little")) + f.flush() + except FileNotFoundError: + return queue_len + return idx + + +def _compute_chunk_window(ci, pred_len, num_chunks, chunk_size, eval_seq_len): + chunk_end = pred_len if ci == num_chunks - 1 else (ci + 1) * chunk_size + win_start = max(0, chunk_end - eval_seq_len) + win_len = chunk_end - win_start + chunk_start = ci * chunk_size + chunk_offset = chunk_start - win_start + chunk_len = chunk_end - chunk_start + return win_start, win_len, chunk_offset, chunk_len + + +def _accumulate_bpb( + ptl, + x, + y, + chunk_offsets, + chunk_lens, + pos_idx, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + loss_sum, + byte_sum, + token_count, + y_bytes=None, +): + pos = pos_idx[: x.size(1)].unsqueeze(0) + mask = ( + (chunk_lens.unsqueeze(1) > 0) + & (pos >= chunk_offsets.unsqueeze(1)) + & (pos < (chunk_offsets + chunk_lens).unsqueeze(1)) + ) + mask_f64 = mask.to(torch.float64) + if y_bytes is not None: + tok_bytes = y_bytes.to(torch.float64) + else: + tok_bytes = base_bytes_lut[y].to(torch.float64) + tok_bytes += (has_leading_space_lut[y] & ~is_boundary_token_lut[x]).to( + torch.float64 + ) + loss_sum += (ptl.to(torch.float64) * mask_f64).sum() + byte_sum += (tok_bytes * mask_f64).sum() + token_count += chunk_lens.to(torch.float64).sum() + + +def _loss_bpb_from_sums(loss_sum, token_count, byte_sum): + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_sum.item()) + return val_loss, val_bpb + + +def _add_to_counter(path, delta): + try: + with open(path, "r+b") as f: + fcntl.flock(f, fcntl.LOCK_EX) + cur = int.from_bytes(f.read(8), "little", signed=True) + cur += int(delta) + f.seek(0) + f.write(int(cur).to_bytes(8, "little", signed=True)) + f.flush() + return cur + except FileNotFoundError: + return int(delta) + + +def _init_int64_counter(path): + with open(path, "wb") as f: + f.write((0).to_bytes(8, "little", signed=True)) + + +def _select_ttt_doc_entries(docs, h): + doc_entries = list(enumerate(docs)) + if h.val_doc_fraction < 1.0: + sample_n = max(1, int(round(len(docs) * h.val_doc_fraction))) + sampled_indices = sorted( + random.Random(h.seed).sample(range(len(docs)), sample_n) + ) + return [(i, docs[i]) for i in sampled_indices] + return doc_entries + + +def train_val_ttt_global_sgd_distributed(h, device, val_data, base_model, val_tokens, batch_seqs=None): + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + base_model.eval() + seq_len = h.eval_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = h.global_ttt_chunk_tokens + batch_seqs = h.global_ttt_batch_seqs if batch_seqs is None else batch_seqs + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + ttt_params = [p for p in base_model.parameters()] + for p in ttt_params: + p.requires_grad_(True) + optimizer = torch.optim.SGD( + ttt_params, lr=h.global_ttt_lr, momentum=h.global_ttt_momentum + ) + t_start = time.perf_counter() + for ci in range(num_chunks): + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + is_last_chunk = ci == num_chunks - 1 + if is_last_chunk or h.global_ttt_epochs <= 0: + continue + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs <= 0: + continue + warmup_chunks = max(0, min(h.global_ttt_warmup_chunks, num_chunks - 1)) + if warmup_chunks > 0 and ci < warmup_chunks: + warmup_denom = max(warmup_chunks - 1, 1) + warmup_t = ci / warmup_denom + lr_now = ( + h.global_ttt_warmup_start_lr + + (h.global_ttt_lr - h.global_ttt_warmup_start_lr) * warmup_t + ) + else: + decay_steps = max(num_chunks - 1 - warmup_chunks, 1) + decay_ci = max(ci - warmup_chunks, 0) + lr_now = h.global_ttt_lr * 0.5 * ( + 1.0 + math.cos(math.pi * decay_ci / decay_steps) + ) + for pg in optimizer.param_groups: + pg["lr"] = lr_now + my_seq_s = chunk_seqs * h.rank // h.world_size + my_seq_e = chunk_seqs * (h.rank + 1) // h.world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ in range(h.global_ttt_epochs): + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x_flat = local[:-1] + y_flat = local[1:] + optimizer.zero_grad(set_to_none=True) + with torch.enable_grad(): + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + if h.global_ttt_respect_doc_boundaries: + bos_pos = (x_flat == BOS_ID).nonzero(as_tuple=True)[0].tolist() + cu_seqlens, max_seqlen = _build_cu_seqlens( + bos_pos, x_flat.numel(), x_flat.device, h.eval_seq_len, 64 + ) + loss = base_model( + x_flat[None], + y_flat[None], + cu_seqlens=cu_seqlens, + max_seqlen=max_seqlen, + ) + else: + x = x_flat.reshape(-1, seq_len) + y = y_flat.reshape(-1, seq_len) + loss = base_model(x, y) + loss.backward() + if dist.is_available() and dist.is_initialized(): + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.SUM) + p.grad.mul_(1.0 / h.world_size) + if h.global_ttt_grad_clip > 0: + torch.nn.utils.clip_grad_norm_(ttt_params, h.global_ttt_grad_clip) + optimizer.step() + base_model.eval() + if h.rank == 0: + elapsed = time.perf_counter() - t_start + log( + f"tttg: c{ci+1}/{num_chunks} lr:{lr_now:.6f} t:{elapsed:.1f}s" + ) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + +def eval_val_ttt_phased(h, base_model, device, val_data, forward_ttt_train): + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(False) + all_tokens = val_data.val_tokens + all_tokens_idx = all_tokens.to(torch.int32) + docs = _find_docs(all_tokens) + doc_entries = _select_ttt_doc_entries(docs, h) + prefix_doc_limit = max(0, min(len(doc_entries), int(h.phased_ttt_prefix_docs))) + num_phases = max(1, int(h.phased_ttt_num_phases)) + phase_boundaries = [] + for pi in range(num_phases): + boundary = prefix_doc_limit * (pi + 1) // num_phases + phase_boundaries.append(boundary) + current_phase = 0 + current_phase_boundary = phase_boundaries[0] + log( + "ttt_phased:" + f" total_docs:{len(doc_entries)} prefix_docs:{prefix_doc_limit} " + f"suffix_docs:{len(doc_entries) - prefix_doc_limit}" + f" num_phases:{num_phases} boundaries:{phase_boundaries}" + ) + chunk_size, eval_seq_len = h.ttt_chunk_size, h.ttt_eval_seq_len + eval_batch_set = None + if h.ttt_eval_batches: + eval_batch_set = set(int(x) for x in h.ttt_eval_batches.split(",") if x.strip()) + use_ascending = eval_batch_set is not None + global_batches_sorted = _build_ttt_global_batches( + doc_entries, h, ascending=use_ascending + ) + queue_len = len(global_batches_sorted) + counter_path = f"/tmp/ttt_counter_{h.run_id}" + prefix_counter_path = f"/tmp/ttt_prefix_counter_{h.run_id}" + pause_flag_path = f"/tmp/ttt_pause_flag_{h.run_id}" + if h.rank == 0: + _init_batch_counter(counter_path) + _init_int64_counter(prefix_counter_path) + try: + os.remove(pause_flag_path) + except FileNotFoundError: + pass + if dist.is_available() and dist.is_initialized(): + path_list = [counter_path, prefix_counter_path, pause_flag_path] + dist.broadcast_object_list(path_list, src=0) + counter_path, prefix_counter_path, pause_flag_path = path_list + dist.barrier() + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + byte_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + t_start = time.perf_counter() + reusable_lora = BatchedTTTLoRA( + h.ttt_batch_size, base_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + + def _build_opt(lora): + if h.ttt_optimizer == "sgd": + return torch.optim.SGD( + lora.parameters(), lr=h.ttt_lora_lr, + momentum=h.ttt_beta1, weight_decay=h.ttt_weight_decay, + ) + return torch.optim.AdamW( + lora.parameters(), lr=h.ttt_lora_lr, + betas=(h.ttt_beta1, h.ttt_beta2), + eps=1e-10, weight_decay=h.ttt_weight_decay, fused=True, + ) + + reusable_opt = _build_opt(reusable_lora) + local_scored_docs = [] + global_ttt_done = prefix_doc_limit == 0 + try: + while True: + queue_idx = _claim_next_batch(counter_path, queue_len) + if queue_idx >= queue_len: + break + orig_batch_idx, batch_entries = global_batches_sorted[queue_idx] + batch = [doc for _, doc in batch_entries] + bsz = len(batch) + prev_loss = loss_sum.item() + prev_bytes = byte_sum.item() + prev_tokens = token_count.item() + if bsz == reusable_lora.bsz: + reusable_lora.reset() + for s in reusable_opt.state.values(): + for k, v in s.items(): + if isinstance(v, torch.Tensor): + v.zero_() + elif k == "step": + s[k] = 0 + cur_lora = reusable_lora + cur_opt = reusable_opt + else: + cur_lora = BatchedTTTLoRA( + bsz, base_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + cur_opt = _build_opt(cur_lora) + pred_lens = [doc_len - 1 for _, doc_len in batch] + num_chunks = [(pl + chunk_size - 1) // chunk_size for pl in pred_lens] + max_nc = max(num_chunks) + num_chunks_t = torch.tensor(num_chunks, dtype=torch.int64, device=device) + for ci in range(max_nc): + active = [ci < nc for nc in num_chunks] + needs_train = any(ci < nc - 1 for nc in num_chunks) + tok_starts = torch.zeros(bsz, dtype=torch.int64) + tok_wls = torch.zeros(bsz, dtype=torch.int64) + chunk_offsets_cpu = torch.zeros(bsz, dtype=torch.int64) + chunk_lens_cpu = torch.zeros(bsz, dtype=torch.int64) + for b in range(bsz): + if not active[b]: + continue + doc_start, doc_len = batch[b] + win_start, win_len, chunk_offset, chunk_len = _compute_chunk_window( + ci, pred_lens[b], num_chunks[b], chunk_size, eval_seq_len + ) + tok_starts[b] = doc_start + win_start + tok_wls[b] = win_len + chunk_offsets_cpu[b] = chunk_offset + chunk_lens_cpu[b] = chunk_len + _, context_size, chunk_offset, _ = _compute_chunk_window( + ci, (ci + 1) * chunk_size, ci + 1, chunk_size, eval_seq_len + ) + col_idx = torch.arange(context_size + 1) + idx = tok_starts.unsqueeze(1) + col_idx.unsqueeze(0) + idx.clamp_(max=all_tokens.numel() - 1) + gathered_gpu = all_tokens_idx[idx].to( + device=device, dtype=torch.int64, non_blocking=True + ) + valid = (col_idx[:context_size].unsqueeze(0) < tok_wls.unsqueeze(1)).to( + device, non_blocking=True + ) + chunk_offsets = chunk_offsets_cpu.to(device, non_blocking=True) + chunk_lens = chunk_lens_cpu.to(device, non_blocking=True) + x = torch.where(valid, gathered_gpu[:, :context_size], 0) + y = torch.where(valid, gathered_gpu[:, 1 : context_size + 1], 0) + ctx_pos = torch.arange(context_size, device=device, dtype=torch.int64) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + per_tok_loss = forward_ttt_train(x, y, lora=cur_lora) + # CaseOps sidecar-driven byte budget. Mirror the index pattern + # used to build y from all_tokens: y[b, j] corresponds to the + # token at global position tok_starts[b] + 1 + j (when valid). + y_bytes_arg = None + if val_data.caseops_enabled and val_data.val_bytes is not None: + y_idx = ( + tok_starts.unsqueeze(1) + + 1 + + col_idx[:context_size].unsqueeze(0) + ) + y_idx = y_idx.clamp_(max=val_data.val_bytes.numel() - 1) + y_bytes_arg = val_data.val_bytes[y_idx].to( + device=device, dtype=torch.int32, non_blocking=True + ) + # Mirror the `valid` masking used for y so out-of-range tokens + # contribute zero bytes (matches y=0 substitution above). + y_bytes_arg = torch.where( + valid, y_bytes_arg, torch.zeros_like(y_bytes_arg) + ) + with torch.no_grad(): + _accumulate_bpb( + per_tok_loss, + x, + y, + chunk_offsets, + chunk_lens, + ctx_pos, + val_data.base_bytes_lut, + val_data.has_leading_space_lut, + val_data.is_boundary_token_lut, + loss_sum, + byte_sum, + token_count, + y_bytes=y_bytes_arg, + ) + if needs_train: + activate_chunk_mask = (num_chunks_t - 1 > ci).float() + for gi in range(h.ttt_grad_steps): + if gi > 0: + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + per_tok_loss = forward_ttt_train(x, y, lora=cur_lora) + per_doc = per_tok_loss[ + :, chunk_offset : chunk_offset + chunk_size + ].mean(dim=-1) + cur_opt.zero_grad(set_to_none=True) + (per_doc * activate_chunk_mask).sum().backward() + cur_opt.step() + else: + del per_tok_loss + batch_num = orig_batch_idx + 1 + doc_lens = [dl for _, dl in batch] + should_report = batch_num in eval_batch_set if eval_batch_set is not None else True + if should_report: + cur_tokens = token_count.item() + cur_loss_val = loss_sum.item() + cur_bytes_val = byte_sum.item() + dt = cur_tokens - prev_tokens + db = cur_bytes_val - prev_bytes + if dt > 0 and db > 0: + b_loss = (cur_loss_val - prev_loss) / dt + b_bpb = b_loss / math.log(2.0) * (dt / db) + else: + b_loss = b_bpb = 0.0 + r_loss = cur_loss_val / max(cur_tokens, 1) + r_bpb = r_loss / math.log(2.0) * (cur_tokens / max(cur_bytes_val, 1)) + elapsed = time.perf_counter() - t_start + log( + f"ttp: b{batch_num}/{queue_len} bl:{b_loss:.4f} bb:{b_bpb:.4f} " + f"rl:{r_loss:.4f} rb:{r_bpb:.4f} dl:{min(doc_lens)}-{max(doc_lens)} " + f"gd:{int(global_ttt_done)}" + ) + if not global_ttt_done: + local_scored_docs.extend( + (orig_batch_idx, pos, doc_start, doc_len) + for pos, (doc_start, doc_len) in enumerate(batch) + ) + prefix_done = _add_to_counter(prefix_counter_path, len(batch_entries)) + if prefix_done >= current_phase_boundary: + try: + with open(pause_flag_path, "x"): + pass + except FileExistsError: + pass + should_pause = os.path.exists(pause_flag_path) + if should_pause: + if dist.is_available() and dist.is_initialized(): + dist.barrier() + gathered_scored_docs = [None] * h.world_size + if dist.is_available() and dist.is_initialized(): + dist.all_gather_object(gathered_scored_docs, local_scored_docs) + else: + gathered_scored_docs = [local_scored_docs] + scored_docs_for_global = [] + for rank_docs in gathered_scored_docs: + if rank_docs: + scored_docs_for_global.extend(rank_docs) + scored_docs_for_global.sort(key=lambda x: (x[0], x[1])) + scored_docs_for_global = scored_docs_for_global[:current_phase_boundary] + scored_token_chunks = [ + val_data.val_tokens[doc_start : doc_start + doc_len] + for _, _, doc_start, doc_len in scored_docs_for_global + ] + if scored_token_chunks: + global_ttt_tokens = torch.cat(scored_token_chunks) + else: + global_ttt_tokens = val_data.val_tokens[:0] + if h.rank == 0: + prefix_done = 0 + try: + with open(prefix_counter_path, "rb") as f: + prefix_done = int.from_bytes( + f.read(8), "little", signed=True + ) + except FileNotFoundError: + pass + log( + f"ttpp: phase:{current_phase + 1}/{num_phases} pd:{prefix_done} " + f"gd:{len(scored_docs_for_global)} " + f"t:{time.perf_counter() - t_start:.1f}s" + ) + train_val_ttt_global_sgd_distributed( + h, device, val_data, base_model, global_ttt_tokens + ) + for p in base_model.parameters(): + p.requires_grad_(False) + reusable_lora = BatchedTTTLoRA( + h.ttt_batch_size, base_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + reusable_opt = _build_opt(reusable_lora) + current_phase += 1 + if current_phase >= num_phases: + global_ttt_done = True + else: + current_phase_boundary = phase_boundaries[current_phase] + if h.rank == 0: + try: + os.remove(pause_flag_path) + except FileNotFoundError: + pass + if dist.is_available() and dist.is_initialized(): + dist.barrier() + if h.rank == 0: + log(f"ttpr: phase:{current_phase}/{num_phases} t:{time.perf_counter() - t_start:.1f}s") + del cur_lora, cur_opt + finally: + pass + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.train() + return _loss_bpb_from_sums(loss_sum, token_count, byte_sum) + + +def timed_eval(label, fn, *args, **kwargs): + torch.cuda.synchronize() + t0 = time.perf_counter() + val_loss, val_bpb = fn(*args, **kwargs) + torch.cuda.synchronize() + elapsed_ms = 1e3 * (time.perf_counter() - t0) + log( + f"{label} val_loss:{val_loss:.8f} val_bpb:{val_bpb:.8f} eval_time:{elapsed_ms:.0f}ms" + ) + return val_loss, val_bpb + + +def train_model(h, device, val_data): + base_model = GPT(h).to(device).bfloat16() + restore_fp32_params(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + compiled_forward_logits = torch.compile( + base_model.forward_logits, dynamic=False, fullgraph=True + ) + model = compiled_model + log(f"model_params:{sum(p.numel()for p in base_model.parameters())}") + optimizers = Optimizers(h, base_model) + train_loader = DocumentPackingLoader(h, device) + max_wallclock_ms = ( + 1e3 * h.max_wallclock_seconds if h.max_wallclock_seconds > 0 else None + ) + if max_wallclock_ms is not None: + max_wallclock_ms -= h.gptq_reserve_seconds * 1e3 + log( + f"gptq:reserving {h.gptq_reserve_seconds:.0f}s, effective={max_wallclock_ms:.0f}ms" + ) + + def training_frac(step, elapsed_ms): + if max_wallclock_ms is None: + return step / max(h.iterations, 1) + return elapsed_ms / max(max_wallclock_ms, 1e-09) + + def lr_mul(frac): + if h.warmdown_frac <= 0: + return 1.0 + if frac >= 1.0 - h.warmdown_frac: + return max((1.0 - frac) / h.warmdown_frac, h.min_lr) + return 1.0 + + _clip_params = [p for p in base_model.parameters() if p.requires_grad] + def step_fn(step, lr_scale): + train_loss = torch.zeros((), device=device) + for micro_step in range(h.grad_accum_steps): + x, y, cu_seqlens, _max_seqlen = train_loader.next_batch( + h.train_batch_tokens, h.grad_accum_steps + ) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y, cu_seqlens=cu_seqlens, max_seqlen=h.train_seq_len) + train_loss += loss.detach() + (loss / h.grad_accum_steps).backward() + train_loss /= h.grad_accum_steps + if step <= h.muon_momentum_warmup_steps: + + frac = ( + + min(step / h.muon_momentum_warmup_steps, 1.0) + + if h.muon_momentum_warmup_steps > 0 + + else 1.0 + + ) + + muon_momentum = ( + + 1 - frac + + ) * h.muon_momentum_warmup_start + frac * h.muon_momentum + + for group in optimizers.optimizer_muon.param_groups: + + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * lr_scale + if h.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(_clip_params, h.grad_clip_norm) + optimizers.step(distributed=h.distributed) + return train_loss + + if h.warmup_steps > 0: + initial_model_state = { + name: tensor.detach().cpu().clone() + for (name, tensor) in base_model.state_dict().items() + } + initial_optimizer_states = [ + copy.deepcopy(opt.state_dict()) for opt in optimizers + ] + model.train() + num_tokens_local = h.train_batch_tokens // h.world_size + for blk in base_model.blocks: + blk.attn.rotary(num_tokens_local, device, torch.bfloat16) + cu_bucket_size = train_loader.cu_bucket_size + warmup_cu_buckets = tuple(cu_bucket_size * i for i in range(1, 5)) + warmup_cu_iters = 3 + x, y, cu_seqlens, _ = train_loader.next_batch( + h.train_batch_tokens, h.grad_accum_steps + ) + log(f"warmup_cu_buckets:{','.join(str(b) for b in warmup_cu_buckets)} iters_each:{warmup_cu_iters}") + def _run_cu_bucket_warmup(): + for bucket_len in warmup_cu_buckets: + boundaries = list(range(0, x.size(1), max(h.train_seq_len, 1))) + if boundaries[-1] != x.size(1): + boundaries.append(x.size(1)) + cu = torch.full((bucket_len,), x.size(1), dtype=torch.int32, device=device) + cu[: len(boundaries)] = torch.tensor(boundaries, dtype=torch.int32, device=device) + for _ in range(warmup_cu_iters): + optimizers.zero_grad_all() + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + wloss = model(x, y, cu_seqlens=cu, max_seqlen=h.train_seq_len) + (wloss / h.grad_accum_steps).backward() + optimizers.zero_grad_all() + _run_cu_bucket_warmup() + if h.num_loops > 0: + base_model.looping_active = True + _run_cu_bucket_warmup() + base_model.looping_active = False + for warmup_step in range(h.warmup_steps): + step_fn(warmup_step, 1.0) + if ( + warmup_step <= 5 + or (warmup_step + 1) % 10 == 0 + or warmup_step + 1 == h.warmup_steps + ): + log(f"warmup_step: {warmup_step+1}/{h.warmup_steps}") + if h.num_loops > 0: + base_model.looping_active = True + log( + f"loop_warmup:enabled encoder:{base_model.encoder_indices} decoder:{base_model.decoder_indices}" + ) + for warmup_step in range(h.warmup_steps): + step_fn(warmup_step, 1.0) + if ( + warmup_step <= 5 + or (warmup_step + 1) % 10 == 0 + or warmup_step + 1 == h.warmup_steps + ): + log(f"loop_warmup_step: {warmup_step+1}/{h.warmup_steps}") + base_model.looping_active = False + base_model.load_state_dict(initial_model_state, strict=True) + for (opt, state) in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + optimizers.zero_grad_all() + train_loader = DocumentPackingLoader(h, device) + _live_state = base_model.state_dict(keep_vars=True) + ema_state = { + name: t.detach().float().clone() + for (name, t) in _live_state.items() + } + _ema_pairs = [(ema_state[name], t) for (name, t) in _live_state.items()] + ema_decay = h.ema_decay + training_time_ms = 0.0 + forced_stop_step = int(os.environ.get("FORCE_STOP_STEP", "0")) + stop_after_step = forced_stop_step if forced_stop_step > 0 else None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = ( + step == h.iterations + or stop_after_step is not None + and step >= stop_after_step + ) + should_validate = ( + last_step or h.val_loss_every > 0 and step % h.val_loss_every == 0 + ) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1e3 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + h, device, val_data, model, compiled_forward_logits + ) + log( + f"{step}/{h.iterations} val_loss: {val_loss:.4f} val_bpb: {val_bpb:.4f}" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < h.iterations: + log( + f"stopping_early: wallclock_cap train_time: {training_time_ms:.0f}ms step: {step}/{h.iterations}" + ) + break + elapsed_ms = training_time_ms + 1e3 * (time.perf_counter() - t0) + frac = training_frac(step, elapsed_ms) + scale = lr_mul(frac) + if ( + h.num_loops > 0 + and not base_model.looping_active + and frac >= h.enable_looping_at + ): + base_model.looping_active = True + log( + f"layer_loop:enabled step:{step} frac:{frac:.3f} encoder:{base_model.encoder_indices} decoder:{base_model.decoder_indices}" + ) + train_loss = step_fn(step, scale) + with torch.no_grad(): + for ema_t, t in _ema_pairs: + ema_t.mul_(ema_decay).add_(t.detach(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1e3 * (time.perf_counter() - t0) + should_log_train = h.train_log_every > 0 and ( + step <= 5 or step % h.train_log_every == 0 or stop_after_step is not None + ) + if should_log_train: + tok_per_sec = step * h.train_batch_tokens / (approx_training_time_ms / 1e3) + log( + f"{step}/{h.iterations} train_loss: {train_loss.item():.4f} train_time: {approx_training_time_ms/60000:.1f}m tok/s: {tok_per_sec:.0f}" + ) + reached_cap = ( + forced_stop_step <= 0 + and max_wallclock_ms is not None + and approx_training_time_ms >= max_wallclock_ms + ) + if h.distributed and forced_stop_step <= 0 and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log( + f"peak memory allocated: {torch.cuda.max_memory_allocated()//1024//1024} MiB reserved: {torch.cuda.max_memory_reserved()//1024//1024} MiB" + ) + log("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = { + name: t.to(dtype=current_state[name].dtype) for (name, t) in ema_state.items() + } + base_model.load_state_dict(avg_state, strict=True) + return base_model, compiled_model, compiled_forward_logits + + +def train_and_eval(h, device): + global BOS_ID + random.seed(h.seed) + np.random.seed(h.seed) + torch.manual_seed(h.seed) + torch.cuda.manual_seed_all(h.seed) + if h.artifact_dir and h.is_main_process: + os.makedirs(h.artifact_dir, exist_ok=True) + val_data = ValidationData(h, device) + log( + f"train_shards: {len(list(Path(h.datasets_dir).resolve().glob('fineweb_train_*.bin')))}" + ) + log(f"val_tokens: {val_data.val_tokens.numel()-1}") + # TTT_EVAL_ONLY: skip training + GPTQ, jump straight to TTT eval on a + # pre-existing quantized artifact. Used to test TTT-only improvements + # (e.g., PR-1767's alpha/warm-start/WD) without retraining. + ttt_eval_only = os.environ.get("TTT_EVAL_ONLY", "0") == "1" + quantize_only = os.environ.get("QUANTIZE_ONLY", "0") == "1" + if ttt_eval_only: + log("TTT_EVAL_ONLY=1 — skipping training + GPTQ, loading saved artifact for TTT eval") + log(f"ttt_lora_alpha: {BatchedLinearLoRA._ALPHA}") + log(f"ttt_warm_start_a: {BatchedLinearLoRA._WARM_START_A}") + log(f"ttt_weight_decay: {h.ttt_weight_decay}") + elif quantize_only: + log("QUANTIZE_ONLY=1 — skipping training, loading saved full-precision checkpoint") + log(f"quantize_only checkpoint: {h.model_path}") + if BOS_ID is None: + BOS_ID = 1 + base_model = GPT(h).to(device).bfloat16() + state = torch.load(h.model_path, map_location="cpu") + base_model.load_state_dict(state, strict=True) + del state + serialize(h, base_model, Path(__file__).read_text(encoding="utf-8")) + if h.distributed: + dist.barrier() + else: + base_model, compiled_model, compiled_forward_logits = train_model( + h, device, val_data + ) + torch._dynamo.reset() + timed_eval( + "diagnostic pre-quantization post-ema", + eval_val, + h, + device, + val_data, + compiled_model, + compiled_forward_logits, + ) + if os.environ.get("PREQUANT_ONLY", "0") == "1": + log("PREQUANT_ONLY=1 — skipping serialize/GPTQ/post-quant eval/TTT") + return + serialize(h, base_model, Path(__file__).read_text(encoding="utf-8")) + if h.distributed: + dist.barrier() + eval_model = deserialize(h, device) + if h.num_loops > 0: + eval_model.looping_active = True + if not ttt_eval_only: + compiled_model = torch.compile(eval_model, dynamic=False, fullgraph=True) + compiled_forward_logits = torch.compile( + eval_model.forward_logits, dynamic=False, fullgraph=True + ) + timed_eval( + "diagnostic quantized", + eval_val, + h, + device, + val_data, + compiled_model, + compiled_forward_logits, + ) + del eval_model + if h.ttt_enabled: + if not ttt_eval_only: + del compiled_model + if ttt_eval_only: + del eval_model + torch._dynamo.reset() + torch.cuda.empty_cache() + ttt_model = deserialize(h, device) + if h.num_loops > 0: + ttt_model.looping_active = True + for p in ttt_model.parameters(): + p.requires_grad_(False) + + if h.rope_yarn: + _yarn_seqlen = h.train_batch_tokens // h.grad_accum_steps + for block in ttt_model.blocks: + block.attn.rotary(_yarn_seqlen, device, torch.bfloat16) + else: + for block in ttt_model.blocks: + block.attn.rotary._cos_cached = None + block.attn.rotary._sin_cached = None + block.attn.rotary._seq_len_cached = 0 + block.attn.rotary(h.ttt_eval_seq_len, device, torch.bfloat16) + + def _fwd_ttt_inner(input_ids, target_ids, lora): + return ttt_model.forward_ttt(input_ids, target_ids, lora=lora) + + _fwd_ttt_compiled_inner = None + + def _fwd_ttt(input_ids, target_ids, lora): + nonlocal _fwd_ttt_compiled_inner + if _fwd_ttt_compiled_inner is None: + _fwd_ttt_compiled_inner = torch.compile(_fwd_ttt_inner, dynamic=True) + return _fwd_ttt_compiled_inner(input_ids, target_ids, lora=lora) + + fwd_ttt_compiled = _fwd_ttt + log(f"ttt_lora:warming up compile (random tokens, no val data)") + if BOS_ID is None: + BOS_ID = 1 + t_warmup = time.perf_counter() + warmup_bszes = [h.ttt_batch_size] + for bsz in warmup_bszes: + wl = BatchedTTTLoRA( + bsz, ttt_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + wo = torch.optim.AdamW( + wl.parameters(), + lr=h.ttt_lora_lr, + betas=(h.ttt_beta1, h.ttt_beta2), + eps=1e-10, + weight_decay=h.ttt_weight_decay, + fused=True, + ) + for ctx_len in (h.ttt_chunk_size, h.ttt_eval_seq_len): + xw = torch.randint(0, h.vocab_size, (bsz, ctx_len), device=device, dtype=torch.int64) + yw = torch.randint(0, h.vocab_size, (bsz, ctx_len), device=device, dtype=torch.int64) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + ptl = fwd_ttt_compiled(xw, yw, lora=wl) + ptl[:, : min(h.ttt_chunk_size, ctx_len)].mean(dim=-1).sum().backward() + wo.step() + wo.zero_grad(set_to_none=True) + del wl, wo + torch.cuda.empty_cache() + compile_elapsed = time.perf_counter() - t_warmup + log(f"ttt_lora:compile warmup done ({compile_elapsed:.1f}s)") + log("\nbeginning TTT eval timer") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_ttt_phased( + h, ttt_model, device, val_data, forward_ttt_train=fwd_ttt_compiled + ) + torch.cuda.synchronize() + ttt_eval_elapsed = time.perf_counter() - t_ttt + log( + "quantized_ttt_phased " + f"val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f} " + f"eval_time:{1e3*ttt_eval_elapsed:.0f}ms" + ) + log(f"total_eval_time:{ttt_eval_elapsed:.1f}s") + del ttt_model + + +def main(): + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError( + f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral" + ) + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + torch.set_float32_matmul_precision("high") + from torch.backends.cuda import ( + enable_cudnn_sdp, + enable_flash_sdp, + enable_math_sdp, + enable_mem_efficient_sdp, + ) + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + torch._dynamo.config.optimize_ddp = False + torch._dynamo.config.cache_size_limit = 64 + h = Hyperparameters() + set_logging_hparams(h) + if h.is_main_process: + os.makedirs(h.artifact_dir if h.artifact_dir else "logs", exist_ok=True) + log(100 * "=", console=False) + log("Hyperparameters:", console=True) + for (k, v) in sorted(vars(type(h)).items()): + if not k.startswith("_"): + log(f" {k}: {v}", console=True) + log("=" * 100, console=False) + log("Source code:", console=False) + log("=" * 100, console=False) + with open(__file__, "r", encoding="utf-8") as _src: + log(_src.read(), console=False) + log("=" * 100, console=False) + log(f"Running Python {sys.version}", console=False) + log(f"Running PyTorch {torch.__version__}", console=False) + log("=" * 100, console=False) + train_and_eval(h, device) + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_seed1234.log b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_seed1234.log new file mode 100644 index 0000000000..73879a8861 --- /dev/null +++ b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_seed1234.log @@ -0,0 +1,947 @@ +=== Starting training: SEED=1234 RUN_ID=awq_seed1234_20260429_055105 === +=== Artifact dir: /workspace/pargolf-work/runs/awq_seed1234_20260429_055105 === +=== NPROC=8 === +W0429 05:51:07.057000 272044 torch/distributed/run.py:803] +W0429 05:51:07.057000 272044 torch/distributed/run.py:803] ***************************************** +W0429 05:51:07.057000 272044 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0429 05:51:07.057000 272044 torch/distributed/run.py:803] ***************************************** +Hyperparameters: + adam_eps: 1e-08 + adam_wd: 0.02 + artifact_dir: /workspace/pargolf-work/runs/awq_seed1234_20260429_055105 + attn_clip_sigmas: 13.0 + attn_out_gate_enabled: False + attn_out_gate_src: proj + awq_lite_bits: 8 + awq_lite_enabled: True + awq_lite_group_size: 64 + awq_lite_group_top_k: 1 + beta1: 0.9 + beta2: 0.99 + caseops_enabled: True + compressor: pergroup + data_dir: ./data/ + datasets_dir: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved + distributed: True + ema_decay: 0.9965 + embed_bits: 7 + embed_clip_sigmas: 14.0 + embed_lr: 0.6 + embed_wd: 0.085 + enable_looping_at: 0.35 + eval_seq_len: 2048 + eval_stride: 64 + fused_ce_enabled: True + gate_window: 12 + gated_attn_enabled: False + gated_attn_init_std: 0.01 + gated_attn_quant_gate: True + global_ttt_batch_seqs: 32 + global_ttt_chunk_tokens: 32768 + global_ttt_epochs: 1 + global_ttt_grad_clip: 1.0 + global_ttt_lr: 0.001 + global_ttt_momentum: 0.9 + global_ttt_respect_doc_boundaries: True + global_ttt_warmup_chunks: 0 + global_ttt_warmup_start_lr: 0.0 + gptq_calibration_batches: 16 + gptq_reserve_seconds: 0.5 + grad_accum_steps: 1 + grad_clip_norm: 0.3 + is_main_process: True + iterations: 20000 + ln_scale: True + local_rank: 0 + logfile: /workspace/pargolf-work/runs/awq_seed1234_20260429_055105/awq_seed1234_20260429_055105.txt + logit_softcap: 30.0 + loop_end: 5 + loop_start: 3 + lqer_asym_enabled: True + lqer_asym_group: 64 + lqer_enabled: True + lqer_factor_bits: 4 + lqer_gain_select: False + lqer_rank: 4 + lqer_scope: all + lqer_top_k: 3 + matrix_bits: 6 + matrix_clip_sigmas: 12.85 + matrix_lr: 0.026 + max_wallclock_seconds: 600.0 + min_lr: 0.1 + mlp_clip_sigmas: 11.5 + mlp_mult: 4.0 + model_dim: 512 + model_path: /workspace/pargolf-work/runs/awq_seed1234_20260429_055105/final_model.pt + muon_backend_steps: 5 + muon_momentum: 0.97 + muon_momentum_warmup_start: 0.92 + muon_momentum_warmup_steps: 1500 + muon_row_normalize: True + muon_wd: 0.095 + num_heads: 8 + num_kv_heads: 4 + num_layers: 11 + num_loops: 2 + parallel_final_lane: mean + parallel_start_layer: 8 + phased_ttt_num_phases: 3 + phased_ttt_prefix_docs: 2500 + qk_gain_init: 5.0 + quantized_model_path: /workspace/pargolf-work/runs/awq_seed1234_20260429_055105/final_model.int6.ptz + rank: 0 + rope_base: 10000.0 + rope_dims: 16 + rope_train_seq_len: 2048 + rope_yarn: False + run_id: awq_seed1234_20260429_055105 + scalar_lr: 0.02 + seed: 1234 + skip_gates_enabled: True + smear_gate_enabled: True + sparse_attn_gate_enabled: True + sparse_attn_gate_init_std: 0.0 + sparse_attn_gate_scale: 0.5 + tie_embeddings: True + tied_embed_init_std: 0.005 + tied_embed_lr: 0.03 + tokenizer_path: /workspace/pargolf/records/track_10min_16mb/2026-04-27_SP8192_LQER_SparseGate_BOSSmearFix_9HpStack_1.0611/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model + train_batch_tokens: 786432 + train_files: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/fineweb_train_*.bin + train_log_every: 500 + train_seq_len: 2048 + ttt_batch_size: 64 + ttt_beta1: 0.0 + ttt_beta2: 0.99 + ttt_chunk_size: 48 + ttt_enabled: True + ttt_eval_batches: + ttt_eval_seq_len: 2048 + ttt_grad_steps: 1 + ttt_k_lora: True + ttt_lora_lr: 0.0001 + ttt_lora_rank: 80 + ttt_mlp_lora: True + ttt_o_lora: True + ttt_optimizer: adam + ttt_weight_decay: 0.5 + val_batch_tokens: 524288 + val_bytes_files: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/fineweb_val_bytes_*.bin + val_doc_fraction: 1.0 + val_files: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/fineweb_val_*.bin + val_loss_every: 4000 + vocab_size: 8192 + warmdown_frac: 0.85 + warmup_steps: 20 + world_size: 8 + xsa_last_n: 11 +train_shards: 80 +val_tokens: 47851520 +model_params:35945671 +gptq:reserving 0s, effective=599500ms +warmup_cu_buckets:64,128,192,256 iters_each:3 +warmup_step: 1/20 +warmup_step: 2/20 +warmup_step: 3/20 +warmup_step: 4/20 +warmup_step: 5/20 +warmup_step: 6/20 +warmup_step: 10/20 +warmup_step: 20/20 +loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +loop_warmup_step: 1/20 +loop_warmup_step: 2/20 +loop_warmup_step: 3/20 +loop_warmup_step: 4/20 +loop_warmup_step: 5/20 +loop_warmup_step: 6/20 +loop_warmup_step: 10/20 +loop_warmup_step: 20/20 +0/20000 val_loss: 9.0008 val_bpb: 4.1127 +1/20000 train_loss: 9.0048 train_time: 0.0m tok/s: 17443421 +2/20000 train_loss: 12.7447 train_time: 0.0m tok/s: 11135875 +3/20000 train_loss: 10.3870 train_time: 0.0m tok/s: 9840663 +4/20000 train_loss: 9.1924 train_time: 0.0m tok/s: 9247123 +5/20000 train_loss: 7.9185 train_time: 0.0m tok/s: 8954019 +500/20000 train_loss: 2.6288 train_time: 0.8m tok/s: 8301535 +1000/20000 train_loss: 2.7393 train_time: 1.6m tok/s: 8296244 +1500/20000 train_loss: 2.7178 train_time: 2.4m tok/s: 8296853 +2000/20000 train_loss: 2.6255 train_time: 3.2m tok/s: 8291262 +layer_loop:enabled step:2212 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +2500/20000 train_loss: 2.4698 train_time: 4.2m tok/s: 7864474 +3000/20000 train_loss: 2.4962 train_time: 5.4m tok/s: 7320722 +3500/20000 train_loss: 2.6220 train_time: 6.5m tok/s: 7021693 +4000/20000 train_loss: 2.3670 train_time: 7.7m tok/s: 6795873 +4000/20000 val_loss: 2.4308 val_bpb: 1.1107 +4500/20000 train_loss: 2.5373 train_time: 8.9m tok/s: 6645807 +4982/20000 val_loss: 2.3576 val_bpb: 1.0772 +stopping_early: wallclock_cap train_time: 599654ms step: 4982/20000 +peak memory allocated: 41707 MiB reserved: 47048 MiB +ema:applying EMA weights +diagnostic pre-quantization post-ema val_loss:2.33178511 val_bpb:1.06546561 eval_time:8209ms +Serialized model: 135417533 bytes +Code size (uncompressed): 168543 bytes +Code size (compressed): 33787 bytes +GPTQ:collecting Hessians from calibration data... +GPTQ:collected 67 Hessians in 4.0s +Quantized weights: + gate_int8_row: blocks.attn.attn_gate_w + gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight + gptq (int6)+lqer_asym: blocks.mlp.fc.weight + gptq (int7)+awqgrpint8+lqer_asym: tok_emb.weight + passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, parallel_post_lambdas, parallel_resid_lambdas, skip_gates, skip_weights, smear_gate.weight, smear_lambda +Serialize: per-group lrzip compression... +Serialize: per-group compression done in 108.1s +Serialized model quantized+pergroup: 15942886 bytes +Total submission size quantized+pergroup: 15976673 bytes +Deserialize: per-group lrzip decompression... +Deserialize: decompression done in 19.4s +diagnostic quantized val_loss:2.35052869 val_bpb:1.07403014 eval_time:13429ms +Deserialize: per-group lrzip decompression... +Deserialize: decompression done in 19.7s +ttt_lora:warming up compile (random tokens, no val data) +ttt_lora:compile warmup done (97.9s) + +beginning TTT eval timer +ttt_phased: total_docs:50000 prefix_docs:2500 suffix_docs:47500 num_phases:3 boundaries:[833, 1666, 2500] +ttp: b780/782 bl:2.2373 bb:1.0778 rl:2.2373 rb:1.0778 dl:13091-17244 gd:0 +ttp: b765/782 bl:2.3181 bb:1.0841 rl:2.2561 rb:1.0793 dl:4393-4510 gd:0 +ttpp: phase:1/3 pd:1296 gd:833 t:169.6s +tttg: c1/131 lr:0.001000 t:2.4s +tttg: c2/131 lr:0.001000 t:2.5s +tttg: c3/131 lr:0.000999 t:2.5s +tttg: c4/131 lr:0.000999 t:2.6s +tttg: c5/131 lr:0.000998 t:2.7s +tttg: c6/131 lr:0.000996 t:2.7s +tttg: c7/131 lr:0.000995 t:2.8s +tttg: c8/131 lr:0.000993 t:2.9s +tttg: c9/131 lr:0.000991 t:2.9s +tttg: c10/131 lr:0.000988 t:3.0s +tttg: c11/131 lr:0.000985 t:3.1s +tttg: c12/131 lr:0.000982 t:3.1s +tttg: c13/131 lr:0.000979 t:3.2s +tttg: c14/131 lr:0.000976 t:3.3s +tttg: c15/131 lr:0.000972 t:3.3s +tttg: c16/131 lr:0.000968 t:3.4s +tttg: c17/131 lr:0.000963 t:3.5s +tttg: c18/131 lr:0.000958 t:3.5s +tttg: c19/131 lr:0.000953 t:3.6s +tttg: c20/131 lr:0.000948 t:3.7s +tttg: c21/131 lr:0.000943 t:3.8s +tttg: c22/131 lr:0.000937 t:3.8s +tttg: c23/131 lr:0.000931 t:3.9s +tttg: c24/131 lr:0.000925 t:4.0s +tttg: c25/131 lr:0.000918 t:4.0s +tttg: c26/131 lr:0.000911 t:4.1s +tttg: c27/131 lr:0.000905 t:4.2s +tttg: c28/131 lr:0.000897 t:4.2s +tttg: c29/131 lr:0.000890 t:5.8s +tttg: c30/131 lr:0.000882 t:5.9s +tttg: c31/131 lr:0.000874 t:5.9s +tttg: c32/131 lr:0.000866 t:6.0s +tttg: c33/131 lr:0.000858 t:6.1s +tttg: c34/131 lr:0.000849 t:6.1s +tttg: c35/131 lr:0.000841 t:6.2s +tttg: c36/131 lr:0.000832 t:6.3s +tttg: c37/131 lr:0.000822 t:6.3s +tttg: c38/131 lr:0.000813 t:6.4s +tttg: c39/131 lr:0.000804 t:6.5s +tttg: c40/131 lr:0.000794 t:6.5s +tttg: c41/131 lr:0.000784 t:6.6s +tttg: c42/131 lr:0.000774 t:6.7s +tttg: c43/131 lr:0.000764 t:6.7s +tttg: c44/131 lr:0.000753 t:6.8s +tttg: c45/131 lr:0.000743 t:6.9s +tttg: c46/131 lr:0.000732 t:7.0s +tttg: c47/131 lr:0.000722 t:7.0s +tttg: c48/131 lr:0.000711 t:7.1s +tttg: c49/131 lr:0.000700 t:7.2s +tttg: c50/131 lr:0.000689 t:7.2s +tttg: c51/131 lr:0.000677 t:7.3s +tttg: c52/131 lr:0.000666 t:7.4s +tttg: c53/131 lr:0.000655 t:7.4s +tttg: c54/131 lr:0.000643 t:7.5s +tttg: c55/131 lr:0.000631 t:7.6s +tttg: c56/131 lr:0.000620 t:7.6s +tttg: c57/131 lr:0.000608 t:7.7s +tttg: c58/131 lr:0.000596 t:7.8s +tttg: c59/131 lr:0.000584 t:7.8s +tttg: c60/131 lr:0.000572 t:7.9s +tttg: c61/131 lr:0.000560 t:8.0s +tttg: c62/131 lr:0.000548 t:8.1s +tttg: c63/131 lr:0.000536 t:8.1s +tttg: c64/131 lr:0.000524 t:8.2s +tttg: c65/131 lr:0.000512 t:8.3s +tttg: c66/131 lr:0.000500 t:8.3s +tttg: c67/131 lr:0.000488 t:8.4s +tttg: c68/131 lr:0.000476 t:8.5s +tttg: c69/131 lr:0.000464 t:8.6s +tttg: c70/131 lr:0.000452 t:8.6s +tttg: c71/131 lr:0.000440 t:8.7s +tttg: c72/131 lr:0.000428 t:8.8s +tttg: c73/131 lr:0.000416 t:8.8s +tttg: c74/131 lr:0.000404 t:8.9s +tttg: c75/131 lr:0.000392 t:9.0s +tttg: c76/131 lr:0.000380 t:9.0s +tttg: c77/131 lr:0.000369 t:9.1s +tttg: c78/131 lr:0.000357 t:9.2s +tttg: c79/131 lr:0.000345 t:9.2s +tttg: c80/131 lr:0.000334 t:9.3s +tttg: c81/131 lr:0.000323 t:9.4s +tttg: c82/131 lr:0.000311 t:9.4s +tttg: c83/131 lr:0.000300 t:9.5s +tttg: c84/131 lr:0.000289 t:9.6s +tttg: c85/131 lr:0.000278 t:9.6s +tttg: c86/131 lr:0.000268 t:9.7s +tttg: c87/131 lr:0.000257 t:9.8s +tttg: c88/131 lr:0.000247 t:9.8s +tttg: c89/131 lr:0.000236 t:9.9s +tttg: c90/131 lr:0.000226 t:10.0s +tttg: c91/131 lr:0.000216 t:10.0s +tttg: c92/131 lr:0.000206 t:10.1s +tttg: c93/131 lr:0.000196 t:10.2s +tttg: c94/131 lr:0.000187 t:10.3s +tttg: c95/131 lr:0.000178 t:10.3s +tttg: c96/131 lr:0.000168 t:10.4s +tttg: c97/131 lr:0.000159 t:10.5s +tttg: c98/131 lr:0.000151 t:10.5s +tttg: c99/131 lr:0.000142 t:10.6s +tttg: c100/131 lr:0.000134 t:10.7s +tttg: c101/131 lr:0.000126 t:10.7s +tttg: c102/131 lr:0.000118 t:10.8s +tttg: c103/131 lr:0.000110 t:12.3s +tttg: c104/131 lr:0.000103 t:12.4s +tttg: c105/131 lr:0.000095 t:12.5s +tttg: c106/131 lr:0.000089 t:12.6s +tttg: c107/131 lr:0.000082 t:12.6s +tttg: c108/131 lr:0.000075 t:12.7s +tttg: c109/131 lr:0.000069 t:12.8s +tttg: c110/131 lr:0.000063 t:12.8s +tttg: c111/131 lr:0.000057 t:12.9s +tttg: c112/131 lr:0.000052 t:13.0s +tttg: c113/131 lr:0.000047 t:13.0s +tttg: c114/131 lr:0.000042 t:13.1s +tttg: c115/131 lr:0.000037 t:13.2s +tttg: c116/131 lr:0.000032 t:13.2s +tttg: c117/131 lr:0.000028 t:13.3s +tttg: c118/131 lr:0.000024 t:13.4s +tttg: c119/131 lr:0.000021 t:13.4s +tttg: c120/131 lr:0.000018 t:13.5s +tttg: c121/131 lr:0.000015 t:13.6s +tttg: c122/131 lr:0.000012 t:13.6s +tttg: c123/131 lr:0.000009 t:13.7s +tttg: c124/131 lr:0.000007 t:13.8s +tttg: c125/131 lr:0.000005 t:13.8s +tttg: c126/131 lr:0.000004 t:13.9s +tttg: c127/131 lr:0.000002 t:14.0s +tttg: c128/131 lr:0.000001 t:14.0s +tttg: c129/131 lr:0.000001 t:14.1s +tttg: c130/131 lr:0.000000 t:14.2s +ttpr: phase:1/3 t:185.3s +ttp: b754/782 bl:2.2870 bb:1.0579 rl:2.2607 rb:1.0760 dl:3345-3397 gd:0 +ttp: b751/782 bl:2.3137 bb:1.0358 rl:2.2673 rb:1.0707 dl:3150-3221 gd:0 +ttpp: phase:2/3 pd:2128 gd:1666 t:251.4s +tttg: c1/219 lr:0.001000 t:0.1s +tttg: c2/219 lr:0.001000 t:0.1s +tttg: c3/219 lr:0.001000 t:0.2s +tttg: c4/219 lr:0.001000 t:0.3s +tttg: c5/219 lr:0.000999 t:0.3s +tttg: c6/219 lr:0.000999 t:0.4s +tttg: c7/219 lr:0.000998 t:0.5s +tttg: c8/219 lr:0.000997 t:0.5s +tttg: c9/219 lr:0.000997 t:0.6s +tttg: c10/219 lr:0.000996 t:0.7s +tttg: c11/219 lr:0.000995 t:0.7s +tttg: c12/219 lr:0.000994 t:0.8s +tttg: c13/219 lr:0.000993 t:0.9s +tttg: c14/219 lr:0.000991 t:1.0s +tttg: c15/219 lr:0.000990 t:1.0s +tttg: c16/219 lr:0.000988 t:1.1s +tttg: c17/219 lr:0.000987 t:1.2s +tttg: c18/219 lr:0.000985 t:1.2s +tttg: c19/219 lr:0.000983 t:1.3s +tttg: c20/219 lr:0.000981 t:1.4s +tttg: c21/219 lr:0.000979 t:1.4s +tttg: c22/219 lr:0.000977 t:1.5s +tttg: c23/219 lr:0.000975 t:1.6s +tttg: c24/219 lr:0.000973 t:1.6s +tttg: c25/219 lr:0.000970 t:1.7s +tttg: c26/219 lr:0.000968 t:1.8s +tttg: c27/219 lr:0.000965 t:1.8s +tttg: c28/219 lr:0.000963 t:1.9s +tttg: c29/219 lr:0.000960 t:2.0s +tttg: c30/219 lr:0.000957 t:2.0s +tttg: c31/219 lr:0.000954 t:2.1s +tttg: c32/219 lr:0.000951 t:2.2s +tttg: c33/219 lr:0.000948 t:2.3s +tttg: c34/219 lr:0.000945 t:2.3s +tttg: c35/219 lr:0.000941 t:2.4s +tttg: c36/219 lr:0.000938 t:2.5s +tttg: c37/219 lr:0.000934 t:2.5s +tttg: c38/219 lr:0.000931 t:2.6s +tttg: c39/219 lr:0.000927 t:2.7s +tttg: c40/219 lr:0.000923 t:2.7s +tttg: c41/219 lr:0.000919 t:2.8s +tttg: c42/219 lr:0.000915 t:2.9s +tttg: c43/219 lr:0.000911 t:2.9s +tttg: c44/219 lr:0.000907 t:3.0s +tttg: c45/219 lr:0.000903 t:3.1s +tttg: c46/219 lr:0.000898 t:3.1s +tttg: c47/219 lr:0.000894 t:3.2s +tttg: c48/219 lr:0.000890 t:3.3s +tttg: c49/219 lr:0.000885 t:3.3s +tttg: c50/219 lr:0.000880 t:3.4s +tttg: c51/219 lr:0.000876 t:3.5s +tttg: c52/219 lr:0.000871 t:3.5s +tttg: c53/219 lr:0.000866 t:3.6s +tttg: c54/219 lr:0.000861 t:3.7s +tttg: c55/219 lr:0.000856 t:3.8s +tttg: c56/219 lr:0.000851 t:3.8s +tttg: c57/219 lr:0.000846 t:3.9s +tttg: c58/219 lr:0.000841 t:4.0s +tttg: c59/219 lr:0.000835 t:4.0s +tttg: c60/219 lr:0.000830 t:4.1s +tttg: c61/219 lr:0.000824 t:4.2s +tttg: c62/219 lr:0.000819 t:4.2s +tttg: c63/219 lr:0.000813 t:4.3s +tttg: c64/219 lr:0.000808 t:4.4s +tttg: c65/219 lr:0.000802 t:4.4s +tttg: c66/219 lr:0.000796 t:4.5s +tttg: c67/219 lr:0.000790 t:4.6s +tttg: c68/219 lr:0.000784 t:4.6s +tttg: c69/219 lr:0.000779 t:4.7s +tttg: c70/219 lr:0.000773 t:4.8s +tttg: c71/219 lr:0.000766 t:4.8s +tttg: c72/219 lr:0.000760 t:4.9s +tttg: c73/219 lr:0.000754 t:5.0s +tttg: c74/219 lr:0.000748 t:5.1s +tttg: c75/219 lr:0.000742 t:5.1s +tttg: c76/219 lr:0.000735 t:5.2s +tttg: c77/219 lr:0.000729 t:5.3s +tttg: c78/219 lr:0.000722 t:5.3s +tttg: c79/219 lr:0.000716 t:5.4s +tttg: c80/219 lr:0.000709 t:5.5s +tttg: c81/219 lr:0.000703 t:5.5s +tttg: c82/219 lr:0.000696 t:5.6s +tttg: c83/219 lr:0.000690 t:5.7s +tttg: c84/219 lr:0.000683 t:5.7s +tttg: c85/219 lr:0.000676 t:5.8s +tttg: c86/219 lr:0.000670 t:5.9s +tttg: c87/219 lr:0.000663 t:5.9s +tttg: c88/219 lr:0.000656 t:6.0s +tttg: c89/219 lr:0.000649 t:6.1s +tttg: c90/219 lr:0.000642 t:6.1s +tttg: c91/219 lr:0.000635 t:6.2s +tttg: c92/219 lr:0.000628 t:6.3s +tttg: c93/219 lr:0.000621 t:6.3s +tttg: c94/219 lr:0.000614 t:6.4s +tttg: c95/219 lr:0.000607 t:6.5s +tttg: c96/219 lr:0.000600 t:6.6s +tttg: c97/219 lr:0.000593 t:6.6s +tttg: c98/219 lr:0.000586 t:6.7s +tttg: c99/219 lr:0.000579 t:6.8s +tttg: c100/219 lr:0.000572 t:6.8s +tttg: c101/219 lr:0.000565 t:6.9s +tttg: c102/219 lr:0.000558 t:7.0s +tttg: c103/219 lr:0.000550 t:7.0s +tttg: c104/219 lr:0.000543 t:7.1s +tttg: c105/219 lr:0.000536 t:7.2s +tttg: c106/219 lr:0.000529 t:7.2s +tttg: c107/219 lr:0.000522 t:7.3s +tttg: c108/219 lr:0.000514 t:7.4s +tttg: c109/219 lr:0.000507 t:7.4s +tttg: c110/219 lr:0.000500 t:7.5s +tttg: c111/219 lr:0.000493 t:7.6s +tttg: c112/219 lr:0.000486 t:7.6s +tttg: c113/219 lr:0.000478 t:7.7s +tttg: c114/219 lr:0.000471 t:7.8s +tttg: c115/219 lr:0.000464 t:7.8s +tttg: c116/219 lr:0.000457 t:7.9s +tttg: c117/219 lr:0.000450 t:8.0s +tttg: c118/219 lr:0.000442 t:8.1s +tttg: c119/219 lr:0.000435 t:8.1s +tttg: c120/219 lr:0.000428 t:8.2s +tttg: c121/219 lr:0.000421 t:8.3s +tttg: c122/219 lr:0.000414 t:8.3s +tttg: c123/219 lr:0.000407 t:8.4s +tttg: c124/219 lr:0.000400 t:8.5s +tttg: c125/219 lr:0.000393 t:8.5s +tttg: c126/219 lr:0.000386 t:8.6s +tttg: c127/219 lr:0.000379 t:8.7s +tttg: c128/219 lr:0.000372 t:8.7s +tttg: c129/219 lr:0.000365 t:8.8s +tttg: c130/219 lr:0.000358 t:8.9s +tttg: c131/219 lr:0.000351 t:9.0s +tttg: c132/219 lr:0.000344 t:9.0s +tttg: c133/219 lr:0.000337 t:9.1s +tttg: c134/219 lr:0.000330 t:9.2s +tttg: c135/219 lr:0.000324 t:9.2s +tttg: c136/219 lr:0.000317 t:9.3s +tttg: c137/219 lr:0.000310 t:9.4s +tttg: c138/219 lr:0.000304 t:9.4s +tttg: c139/219 lr:0.000297 t:9.5s +tttg: c140/219 lr:0.000291 t:9.6s +tttg: c141/219 lr:0.000284 t:9.6s +tttg: c142/219 lr:0.000278 t:9.7s +tttg: c143/219 lr:0.000271 t:9.8s +tttg: c144/219 lr:0.000265 t:9.8s +tttg: c145/219 lr:0.000258 t:9.9s +tttg: c146/219 lr:0.000252 t:10.0s +tttg: c147/219 lr:0.000246 t:10.1s +tttg: c148/219 lr:0.000240 t:10.1s +tttg: c149/219 lr:0.000234 t:10.2s +tttg: c150/219 lr:0.000227 t:10.3s +tttg: c151/219 lr:0.000221 t:10.3s +tttg: c152/219 lr:0.000216 t:10.4s +tttg: c153/219 lr:0.000210 t:10.5s +tttg: c154/219 lr:0.000204 t:10.5s +tttg: c155/219 lr:0.000198 t:10.6s +tttg: c156/219 lr:0.000192 t:10.7s +tttg: c157/219 lr:0.000187 t:10.7s +tttg: c158/219 lr:0.000181 t:10.8s +tttg: c159/219 lr:0.000176 t:10.9s +tttg: c160/219 lr:0.000170 t:11.0s +tttg: c161/219 lr:0.000165 t:11.0s +tttg: c162/219 lr:0.000159 t:11.1s +tttg: c163/219 lr:0.000154 t:11.2s +tttg: c164/219 lr:0.000149 t:11.2s +tttg: c165/219 lr:0.000144 t:11.3s +tttg: c166/219 lr:0.000139 t:11.4s +tttg: c167/219 lr:0.000134 t:11.4s +tttg: c168/219 lr:0.000129 t:11.5s +tttg: c169/219 lr:0.000124 t:11.6s +tttg: c170/219 lr:0.000120 t:11.7s +tttg: c171/219 lr:0.000115 t:11.7s +tttg: c172/219 lr:0.000110 t:11.8s +tttg: c173/219 lr:0.000106 t:11.9s +tttg: c174/219 lr:0.000102 t:11.9s +tttg: c175/219 lr:0.000097 t:12.0s +tttg: c176/219 lr:0.000093 t:12.1s +tttg: c177/219 lr:0.000089 t:12.1s +tttg: c178/219 lr:0.000085 t:12.2s +tttg: c179/219 lr:0.000081 t:12.3s +tttg: c180/219 lr:0.000077 t:12.3s +tttg: c181/219 lr:0.000073 t:12.4s +tttg: c182/219 lr:0.000069 t:12.5s +tttg: c183/219 lr:0.000066 t:12.6s +tttg: c184/219 lr:0.000062 t:12.6s +tttg: c185/219 lr:0.000059 t:12.7s +tttg: c186/219 lr:0.000055 t:12.8s +tttg: c187/219 lr:0.000052 t:12.8s +tttg: c188/219 lr:0.000049 t:12.9s +tttg: c189/219 lr:0.000046 t:13.0s +tttg: c190/219 lr:0.000043 t:13.0s +tttg: c191/219 lr:0.000040 t:13.1s +tttg: c192/219 lr:0.000037 t:13.2s +tttg: c193/219 lr:0.000035 t:13.2s +tttg: c194/219 lr:0.000032 t:13.3s +tttg: c195/219 lr:0.000030 t:13.4s +tttg: c196/219 lr:0.000027 t:13.4s +tttg: c197/219 lr:0.000025 t:13.5s +tttg: c198/219 lr:0.000023 t:13.6s +tttg: c199/219 lr:0.000021 t:13.6s +tttg: c200/219 lr:0.000019 t:13.7s +tttg: c201/219 lr:0.000017 t:13.8s +tttg: c202/219 lr:0.000015 t:13.8s +tttg: c203/219 lr:0.000013 t:13.9s +tttg: c204/219 lr:0.000012 t:14.0s +tttg: c205/219 lr:0.000010 t:14.1s +tttg: c206/219 lr:0.000009 t:14.1s +tttg: c207/219 lr:0.000007 t:14.2s +tttg: c208/219 lr:0.000006 t:14.3s +tttg: c209/219 lr:0.000005 t:14.3s +tttg: c210/219 lr:0.000004 t:14.4s +tttg: c211/219 lr:0.000003 t:14.5s +tttg: c212/219 lr:0.000003 t:14.5s +tttg: c213/219 lr:0.000002 t:14.6s +tttg: c214/219 lr:0.000001 t:14.7s +tttg: c215/219 lr:0.000001 t:14.7s +tttg: c216/219 lr:0.000000 t:14.8s +tttg: c217/219 lr:0.000000 t:14.9s +tttg: c218/219 lr:0.000000 t:15.0s +ttpr: phase:2/3 t:267.9s +ttp: b741/782 bl:2.3227 bb:1.0416 rl:2.2726 rb:1.0678 dl:2686-2730 gd:0 +ttp: b740/782 bl:2.2599 bb:1.0374 rl:2.2715 rb:1.0652 dl:2653-2686 gd:0 +ttpp: phase:3/3 pd:2960 gd:2500 t:283.1s +tttg: c1/289 lr:0.001000 t:0.1s +tttg: c2/289 lr:0.001000 t:0.1s +tttg: c3/289 lr:0.001000 t:0.2s +tttg: c4/289 lr:0.001000 t:0.3s +tttg: c5/289 lr:0.001000 t:0.3s +tttg: c6/289 lr:0.000999 t:0.4s +tttg: c7/289 lr:0.000999 t:0.5s +tttg: c8/289 lr:0.000999 t:0.5s +tttg: c9/289 lr:0.000998 t:0.6s +tttg: c10/289 lr:0.000998 t:0.7s +tttg: c11/289 lr:0.000997 t:0.8s +tttg: c12/289 lr:0.000996 t:0.8s +tttg: c13/289 lr:0.000996 t:0.9s +tttg: c14/289 lr:0.000995 t:1.0s +tttg: c15/289 lr:0.000994 t:1.0s +tttg: c16/289 lr:0.000993 t:1.1s +tttg: c17/289 lr:0.000992 t:1.2s +tttg: c18/289 lr:0.000991 t:1.2s +tttg: c19/289 lr:0.000990 t:1.3s +tttg: c20/289 lr:0.000989 t:1.4s +tttg: c21/289 lr:0.000988 t:1.4s +tttg: c22/289 lr:0.000987 t:1.5s +tttg: c23/289 lr:0.000986 t:1.6s +tttg: c24/289 lr:0.000984 t:1.6s +tttg: c25/289 lr:0.000983 t:1.7s +tttg: c26/289 lr:0.000982 t:1.8s +tttg: c27/289 lr:0.000980 t:1.8s +tttg: c28/289 lr:0.000978 t:1.9s +tttg: c29/289 lr:0.000977 t:2.0s +tttg: c30/289 lr:0.000975 t:2.0s +tttg: c31/289 lr:0.000973 t:2.1s +tttg: c32/289 lr:0.000972 t:2.2s +tttg: c33/289 lr:0.000970 t:2.2s +tttg: c34/289 lr:0.000968 t:2.3s +tttg: c35/289 lr:0.000966 t:2.4s +tttg: c36/289 lr:0.000964 t:2.4s +tttg: c37/289 lr:0.000962 t:2.5s +tttg: c38/289 lr:0.000960 t:2.6s +tttg: c39/289 lr:0.000958 t:2.7s +tttg: c40/289 lr:0.000955 t:2.7s +tttg: c41/289 lr:0.000953 t:2.8s +tttg: c42/289 lr:0.000951 t:2.9s +tttg: c43/289 lr:0.000948 t:2.9s +tttg: c44/289 lr:0.000946 t:3.0s +tttg: c45/289 lr:0.000944 t:3.1s +tttg: c46/289 lr:0.000941 t:3.1s +tttg: c47/289 lr:0.000938 t:3.2s +tttg: c48/289 lr:0.000936 t:3.3s +tttg: c49/289 lr:0.000933 t:3.3s +tttg: c50/289 lr:0.000930 t:3.4s +tttg: c51/289 lr:0.000927 t:3.5s +tttg: c52/289 lr:0.000925 t:3.5s +tttg: c53/289 lr:0.000922 t:3.6s +tttg: c54/289 lr:0.000919 t:3.7s +tttg: c55/289 lr:0.000916 t:3.7s +tttg: c56/289 lr:0.000913 t:3.8s +tttg: c57/289 lr:0.000910 t:3.9s +tttg: c58/289 lr:0.000906 t:3.9s +tttg: c59/289 lr:0.000903 t:4.0s +tttg: c60/289 lr:0.000900 t:4.1s +tttg: c61/289 lr:0.000897 t:4.1s +tttg: c62/289 lr:0.000893 t:4.2s +tttg: c63/289 lr:0.000890 t:4.3s +tttg: c64/289 lr:0.000887 t:4.4s +tttg: c65/289 lr:0.000883 t:4.4s +tttg: c66/289 lr:0.000879 t:4.5s +tttg: c67/289 lr:0.000876 t:4.6s +tttg: c68/289 lr:0.000872 t:4.6s +tttg: c69/289 lr:0.000869 t:4.7s +tttg: c70/289 lr:0.000865 t:4.8s +tttg: c71/289 lr:0.000861 t:4.8s +tttg: c72/289 lr:0.000857 t:4.9s +tttg: c73/289 lr:0.000854 t:5.0s +tttg: c74/289 lr:0.000850 t:5.0s +tttg: c75/289 lr:0.000846 t:5.1s +tttg: c76/289 lr:0.000842 t:5.2s +tttg: c77/289 lr:0.000838 t:5.2s +tttg: c78/289 lr:0.000834 t:5.3s +tttg: c79/289 lr:0.000830 t:5.4s +tttg: c80/289 lr:0.000826 t:5.4s +tttg: c81/289 lr:0.000821 t:5.5s +tttg: c82/289 lr:0.000817 t:5.6s +tttg: c83/289 lr:0.000813 t:5.6s +tttg: c84/289 lr:0.000809 t:5.7s +tttg: c85/289 lr:0.000804 t:5.8s +tttg: c86/289 lr:0.000800 t:5.8s +tttg: c87/289 lr:0.000796 t:5.9s +tttg: c88/289 lr:0.000791 t:6.0s +tttg: c89/289 lr:0.000787 t:6.1s +tttg: c90/289 lr:0.000782 t:6.1s +tttg: c91/289 lr:0.000778 t:6.2s +tttg: c92/289 lr:0.000773 t:6.3s +tttg: c93/289 lr:0.000769 t:6.3s +tttg: c94/289 lr:0.000764 t:6.4s +tttg: c95/289 lr:0.000759 t:6.5s +tttg: c96/289 lr:0.000755 t:6.5s +tttg: c97/289 lr:0.000750 t:6.6s +tttg: c98/289 lr:0.000745 t:6.7s +tttg: c99/289 lr:0.000740 t:6.7s +tttg: c100/289 lr:0.000736 t:6.8s +tttg: c101/289 lr:0.000731 t:6.9s +tttg: c102/289 lr:0.000726 t:6.9s +tttg: c103/289 lr:0.000721 t:7.0s +tttg: c104/289 lr:0.000716 t:7.1s +tttg: c105/289 lr:0.000711 t:7.1s +tttg: c106/289 lr:0.000706 t:7.2s +tttg: c107/289 lr:0.000701 t:7.3s +tttg: c108/289 lr:0.000696 t:7.4s +tttg: c109/289 lr:0.000691 t:7.4s +tttg: c110/289 lr:0.000686 t:7.5s +tttg: c111/289 lr:0.000681 t:7.6s +tttg: c112/289 lr:0.000676 t:7.6s +tttg: c113/289 lr:0.000671 t:7.7s +tttg: c114/289 lr:0.000666 t:7.8s +tttg: c115/289 lr:0.000661 t:7.8s +tttg: c116/289 lr:0.000656 t:7.9s +tttg: c117/289 lr:0.000650 t:8.0s +tttg: c118/289 lr:0.000645 t:8.0s +tttg: c119/289 lr:0.000640 t:8.1s +tttg: c120/289 lr:0.000635 t:8.2s +tttg: c121/289 lr:0.000629 t:8.2s +tttg: c122/289 lr:0.000624 t:8.3s +tttg: c123/289 lr:0.000619 t:8.4s +tttg: c124/289 lr:0.000614 t:8.4s +tttg: c125/289 lr:0.000608 t:8.5s +tttg: c126/289 lr:0.000603 t:8.6s +tttg: c127/289 lr:0.000598 t:8.6s +tttg: c128/289 lr:0.000592 t:8.7s +tttg: c129/289 lr:0.000587 t:8.8s +tttg: c130/289 lr:0.000581 t:8.8s +tttg: c131/289 lr:0.000576 t:8.9s +tttg: c132/289 lr:0.000571 t:9.0s +tttg: c133/289 lr:0.000565 t:9.0s +tttg: c134/289 lr:0.000560 t:9.1s +tttg: c135/289 lr:0.000554 t:9.2s +tttg: c136/289 lr:0.000549 t:9.3s +tttg: c137/289 lr:0.000544 t:9.3s +tttg: c138/289 lr:0.000538 t:9.4s +tttg: c139/289 lr:0.000533 t:9.5s +tttg: c140/289 lr:0.000527 t:9.5s +tttg: c141/289 lr:0.000522 t:9.6s +tttg: c142/289 lr:0.000516 t:9.7s +tttg: c143/289 lr:0.000511 t:9.7s +tttg: c144/289 lr:0.000505 t:9.8s +tttg: c145/289 lr:0.000500 t:9.9s +tttg: c146/289 lr:0.000495 t:9.9s +tttg: c147/289 lr:0.000489 t:10.0s +tttg: c148/289 lr:0.000484 t:10.1s +tttg: c149/289 lr:0.000478 t:10.1s +tttg: c150/289 lr:0.000473 t:10.2s +tttg: c151/289 lr:0.000467 t:10.3s +tttg: c152/289 lr:0.000462 t:10.3s +tttg: c153/289 lr:0.000456 t:10.4s +tttg: c154/289 lr:0.000451 t:10.5s +tttg: c155/289 lr:0.000446 t:10.5s +tttg: c156/289 lr:0.000440 t:10.6s +tttg: c157/289 lr:0.000435 t:10.7s +tttg: c158/289 lr:0.000429 t:10.7s +tttg: c159/289 lr:0.000424 t:10.8s +tttg: c160/289 lr:0.000419 t:10.9s +tttg: c161/289 lr:0.000413 t:11.0s +tttg: c162/289 lr:0.000408 t:11.0s +tttg: c163/289 lr:0.000402 t:11.1s +tttg: c164/289 lr:0.000397 t:11.2s +tttg: c165/289 lr:0.000392 t:11.2s +tttg: c166/289 lr:0.000386 t:11.3s +tttg: c167/289 lr:0.000381 t:11.4s +tttg: c168/289 lr:0.000376 t:11.4s +tttg: c169/289 lr:0.000371 t:11.5s +tttg: c170/289 lr:0.000365 t:11.6s +tttg: c171/289 lr:0.000360 t:11.6s +tttg: c172/289 lr:0.000355 t:11.7s +tttg: c173/289 lr:0.000350 t:11.8s +tttg: c174/289 lr:0.000344 t:11.8s +tttg: c175/289 lr:0.000339 t:11.9s +tttg: c176/289 lr:0.000334 t:12.0s +tttg: c177/289 lr:0.000329 t:12.0s +tttg: c178/289 lr:0.000324 t:12.1s +tttg: c179/289 lr:0.000319 t:12.2s +tttg: c180/289 lr:0.000314 t:12.2s +tttg: c181/289 lr:0.000309 t:12.3s +tttg: c182/289 lr:0.000304 t:12.4s +tttg: c183/289 lr:0.000299 t:12.4s +tttg: c184/289 lr:0.000294 t:12.5s +tttg: c185/289 lr:0.000289 t:12.6s +tttg: c186/289 lr:0.000284 t:12.7s +tttg: c187/289 lr:0.000279 t:12.7s +tttg: c188/289 lr:0.000274 t:12.8s +tttg: c189/289 lr:0.000269 t:12.9s +tttg: c190/289 lr:0.000264 t:12.9s +tttg: c191/289 lr:0.000260 t:13.0s +tttg: c192/289 lr:0.000255 t:13.1s +tttg: c193/289 lr:0.000250 t:13.1s +tttg: c194/289 lr:0.000245 t:13.2s +tttg: c195/289 lr:0.000241 t:13.3s +tttg: c196/289 lr:0.000236 t:13.3s +tttg: c197/289 lr:0.000231 t:13.4s +tttg: c198/289 lr:0.000227 t:13.5s +tttg: c199/289 lr:0.000222 t:13.5s +tttg: c200/289 lr:0.000218 t:13.6s +tttg: c201/289 lr:0.000213 t:13.7s +tttg: c202/289 lr:0.000209 t:13.8s +tttg: c203/289 lr:0.000204 t:13.8s +tttg: c204/289 lr:0.000200 t:13.9s +tttg: c205/289 lr:0.000196 t:14.0s +tttg: c206/289 lr:0.000191 t:14.0s +tttg: c207/289 lr:0.000187 t:14.1s +tttg: c208/289 lr:0.000183 t:14.2s +tttg: c209/289 lr:0.000179 t:14.2s +tttg: c210/289 lr:0.000174 t:14.3s +tttg: c211/289 lr:0.000170 t:14.4s +tttg: c212/289 lr:0.000166 t:14.4s +tttg: c213/289 lr:0.000162 t:14.5s +tttg: c214/289 lr:0.000158 t:14.6s +tttg: c215/289 lr:0.000154 t:14.6s +tttg: c216/289 lr:0.000150 t:14.7s +tttg: c217/289 lr:0.000146 t:14.8s +tttg: c218/289 lr:0.000143 t:14.8s +tttg: c219/289 lr:0.000139 t:14.9s +tttg: c220/289 lr:0.000135 t:15.0s +tttg: c221/289 lr:0.000131 t:15.1s +tttg: c222/289 lr:0.000128 t:15.1s +tttg: c223/289 lr:0.000124 t:15.2s +tttg: c224/289 lr:0.000121 t:15.3s +tttg: c225/289 lr:0.000117 t:15.3s +tttg: c226/289 lr:0.000113 t:15.4s +tttg: c227/289 lr:0.000110 t:15.5s +tttg: c228/289 lr:0.000107 t:15.5s +tttg: c229/289 lr:0.000103 t:15.6s +tttg: c230/289 lr:0.000100 t:15.7s +tttg: c231/289 lr:0.000097 t:15.7s +tttg: c232/289 lr:0.000094 t:15.8s +tttg: c233/289 lr:0.000090 t:15.9s +tttg: c234/289 lr:0.000087 t:15.9s +tttg: c235/289 lr:0.000084 t:16.0s +tttg: c236/289 lr:0.000081 t:16.1s +tttg: c237/289 lr:0.000078 t:16.1s +tttg: c238/289 lr:0.000075 t:16.2s +tttg: c239/289 lr:0.000073 t:16.3s +tttg: c240/289 lr:0.000070 t:16.3s +tttg: c241/289 lr:0.000067 t:16.4s +tttg: c242/289 lr:0.000064 t:16.5s +tttg: c243/289 lr:0.000062 t:16.5s +tttg: c244/289 lr:0.000059 t:16.6s +tttg: c245/289 lr:0.000056 t:16.7s +tttg: c246/289 lr:0.000054 t:16.8s +tttg: c247/289 lr:0.000052 t:16.8s +tttg: c248/289 lr:0.000049 t:16.9s +tttg: c249/289 lr:0.000047 t:17.0s +tttg: c250/289 lr:0.000045 t:17.0s +tttg: c251/289 lr:0.000042 t:17.1s +tttg: c252/289 lr:0.000040 t:17.2s +tttg: c253/289 lr:0.000038 t:17.2s +tttg: c254/289 lr:0.000036 t:17.3s +tttg: c255/289 lr:0.000034 t:17.4s +tttg: c256/289 lr:0.000032 t:17.4s +tttg: c257/289 lr:0.000030 t:17.5s +tttg: c258/289 lr:0.000028 t:17.6s +tttg: c259/289 lr:0.000027 t:17.6s +tttg: c260/289 lr:0.000025 t:17.7s +tttg: c261/289 lr:0.000023 t:17.8s +tttg: c262/289 lr:0.000022 t:17.8s +tttg: c263/289 lr:0.000020 t:17.9s +tttg: c264/289 lr:0.000018 t:18.0s +tttg: c265/289 lr:0.000017 t:18.0s +tttg: c266/289 lr:0.000016 t:18.1s +tttg: c267/289 lr:0.000014 t:18.2s +tttg: c268/289 lr:0.000013 t:18.3s +tttg: c269/289 lr:0.000012 t:18.3s +tttg: c270/289 lr:0.000011 t:18.4s +tttg: c271/289 lr:0.000010 t:18.5s +tttg: c272/289 lr:0.000009 t:18.5s +tttg: c273/289 lr:0.000008 t:18.6s +tttg: c274/289 lr:0.000007 t:18.7s +tttg: c275/289 lr:0.000006 t:18.7s +tttg: c276/289 lr:0.000005 t:18.8s +tttg: c277/289 lr:0.000004 t:18.9s +tttg: c278/289 lr:0.000004 t:18.9s +tttg: c279/289 lr:0.000003 t:19.0s +tttg: c280/289 lr:0.000002 t:19.1s +tttg: c281/289 lr:0.000002 t:19.1s +tttg: c282/289 lr:0.000001 t:19.2s +tttg: c283/289 lr:0.000001 t:19.3s +tttg: c284/289 lr:0.000001 t:19.3s +tttg: c285/289 lr:0.000000 t:19.4s +tttg: c286/289 lr:0.000000 t:19.5s +tttg: c287/289 lr:0.000000 t:19.6s +tttg: c288/289 lr:0.000000 t:19.6s +ttpr: phase:3/3 t:304.3s +ttp: b728/782 bl:2.3652 bb:1.0828 rl:2.2780 rb:1.0664 dl:2306-2324 gd:1 +ttp: b726/782 bl:2.3362 bb:1.0391 rl:2.2817 rb:1.0646 dl:2254-2276 gd:1 +ttp: b716/782 bl:2.2543 bb:1.0417 rl:2.2802 rb:1.0633 dl:2054-2069 gd:1 +ttp: b707/782 bl:2.3574 bb:1.0476 rl:2.2839 rb:1.0625 dl:1910-1923 gd:1 +ttp: b697/782 bl:2.3273 bb:1.0326 rl:2.2858 rb:1.0612 dl:1790-1803 gd:1 +ttp: b692/782 bl:2.2947 bb:1.0301 rl:2.2862 rb:1.0599 dl:1737-1746 gd:1 +ttp: b682/782 bl:2.3448 bb:1.0581 rl:2.2883 rb:1.0598 dl:1638-1646 gd:1 +ttp: b674/782 bl:2.4074 bb:1.0903 rl:2.2923 rb:1.0609 dl:1571-1578 gd:1 +ttp: b670/782 bl:2.3505 bb:1.0696 rl:2.2942 rb:1.0612 dl:1537-1544 gd:1 +ttp: b656/782 bl:2.3300 bb:1.1115 rl:2.2953 rb:1.0626 dl:1439-1445 gd:1 +ttp: b648/782 bl:2.2839 bb:1.0078 rl:2.2949 rb:1.0610 dl:1387-1392 gd:1 +ttp: b647/782 bl:2.2781 bb:1.0339 rl:2.2945 rb:1.0603 dl:1382-1387 gd:1 +ttp: b638/782 bl:2.3468 bb:1.0692 rl:2.2958 rb:1.0605 dl:1325-1331 gd:1 +ttp: b627/782 bl:2.3785 bb:1.0709 rl:2.2977 rb:1.0607 dl:1266-1271 gd:1 +ttp: b619/782 bl:2.3254 bb:1.0605 rl:2.2983 rb:1.0607 dl:1221-1226 gd:1 +ttp: b609/782 bl:2.2724 bb:1.0181 rl:2.2978 rb:1.0598 dl:1172-1177 gd:1 +ttp: b602/782 bl:2.3796 bb:1.0496 rl:2.2994 rb:1.0596 dl:1141-1146 gd:1 +ttp: b598/782 bl:2.3564 bb:1.0658 rl:2.3005 rb:1.0597 dl:1124-1129 gd:1 +ttp: b587/782 bl:2.4053 bb:1.0674 rl:2.3023 rb:1.0599 dl:1077-1081 gd:1 +ttp: b579/782 bl:2.3443 bb:1.0361 rl:2.3031 rb:1.0595 dl:1044-1048 gd:1 +ttp: b573/782 bl:2.3675 bb:1.0672 rl:2.3041 rb:1.0596 dl:1021-1025 gd:1 +ttp: b564/782 bl:2.2867 bb:1.0175 rl:2.3038 rb:1.0589 dl:990-993 gd:1 +ttp: b556/782 bl:2.3773 bb:1.0688 rl:2.3049 rb:1.0591 dl:961-965 gd:1 +ttp: b546/782 bl:2.3295 bb:1.0357 rl:2.3053 rb:1.0587 dl:930-934 gd:1 +ttp: b539/782 bl:2.3388 bb:1.0368 rl:2.3057 rb:1.0584 dl:909-912 gd:1 +ttp: b536/782 bl:2.3243 bb:1.0467 rl:2.3060 rb:1.0583 dl:899-902 gd:1 +ttp: b527/782 bl:2.3454 bb:1.0294 rl:2.3065 rb:1.0579 dl:872-875 gd:1 +ttp: b519/782 bl:2.2964 bb:1.0418 rl:2.3064 rb:1.0577 dl:850-852 gd:1 +ttp: b512/782 bl:2.3030 bb:1.0636 rl:2.3063 rb:1.0577 dl:829-832 gd:1 +ttp: b504/782 bl:2.3207 bb:1.0353 rl:2.3065 rb:1.0575 dl:807-809 gd:1 +ttp: b496/782 bl:2.4205 bb:1.0479 rl:2.3078 rb:1.0574 dl:785-788 gd:1 +ttp: b489/782 bl:2.3888 bb:1.0747 rl:2.3086 rb:1.0576 dl:769-771 gd:1 +ttp: b482/782 bl:2.3266 bb:1.0460 rl:2.3088 rb:1.0574 dl:752-754 gd:1 +ttp: b474/782 bl:2.3406 bb:1.0717 rl:2.3091 rb:1.0576 dl:733-735 gd:1 +ttp: b466/782 bl:2.3875 bb:1.0293 rl:2.3099 rb:1.0573 dl:714-717 gd:1 +ttp: b458/782 bl:2.2041 bb:1.0222 rl:2.3089 rb:1.0570 dl:697-700 gd:1 +ttp: b450/782 bl:2.3571 bb:1.0332 rl:2.3093 rb:1.0567 dl:680-682 gd:1 +ttp: b428/782 bl:2.3049 bb:1.0503 rl:2.3093 rb:1.0567 dl:636-638 gd:1 +ttp: b420/782 bl:2.3597 bb:1.0534 rl:2.3097 rb:1.0567 dl:620-622 gd:1 +ttp: b412/782 bl:2.3329 bb:1.0460 rl:2.3099 rb:1.0566 dl:605-607 gd:1 +ttp: b404/782 bl:2.3656 bb:1.0594 rl:2.3103 rb:1.0566 dl:590-592 gd:1 +ttp: b397/782 bl:2.3518 bb:1.0430 rl:2.3106 rb:1.0565 dl:577-579 gd:1 +ttp: b388/782 bl:2.3094 bb:1.0415 rl:2.3106 rb:1.0564 dl:561-562 gd:1 +ttp: b380/782 bl:2.3565 bb:1.0867 rl:2.3109 rb:1.0566 dl:547-549 gd:1 +ttp: b372/782 bl:2.3305 bb:1.0468 rl:2.3110 rb:1.0565 dl:533-535 gd:1 +ttp: b364/782 bl:2.3490 bb:1.0621 rl:2.3113 rb:1.0566 dl:521-522 gd:1 +ttp: b356/782 bl:2.3413 bb:1.0543 rl:2.3115 rb:1.0565 dl:506-508 gd:1 +ttp: b348/782 bl:2.3588 bb:1.0580 rl:2.3117 rb:1.0566 dl:494-495 gd:1 +ttp: b340/782 bl:2.4575 bb:1.0803 rl:2.3126 rb:1.0567 dl:482-483 gd:1 +ttp: b332/782 bl:2.3072 bb:1.0444 rl:2.3126 rb:1.0566 dl:469-471 gd:1 +ttp: b324/782 bl:2.3182 bb:1.0838 rl:2.3126 rb:1.0568 dl:458-459 gd:1 +ttp: b316/782 bl:2.3683 bb:1.0804 rl:2.3129 rb:1.0569 dl:445-446 gd:1 +ttp: b308/782 bl:2.4097 bb:1.0929 rl:2.3134 rb:1.0571 dl:433-435 gd:1 +ttp: b299/782 bl:2.3326 bb:1.1077 rl:2.3135 rb:1.0573 dl:420-421 gd:1 +ttp: b291/782 bl:2.2654 bb:1.0129 rl:2.3133 rb:1.0571 dl:407-409 gd:1 +ttp: b283/782 bl:2.3702 bb:1.1270 rl:2.3135 rb:1.0574 dl:396-398 gd:1 +ttp: b274/782 bl:2.3021 bb:1.0701 rl:2.3135 rb:1.0575 dl:384-385 gd:1 +ttp: b266/782 bl:2.3742 bb:1.1046 rl:2.3137 rb:1.0577 dl:374-375 gd:1 +ttp: b258/782 bl:2.4397 bb:1.0947 rl:2.3143 rb:1.0578 dl:364-365 gd:1 +ttp: b250/782 bl:2.3184 bb:1.0749 rl:2.3143 rb:1.0579 dl:354-355 gd:1 +ttp: b242/782 bl:2.3767 bb:1.1002 rl:2.3145 rb:1.0581 dl:344-345 gd:1 +ttp: b232/782 bl:2.3078 bb:1.0877 rl:2.3145 rb:1.0582 dl:331-333 gd:1 +ttp: b223/782 bl:2.3291 bb:1.1245 rl:2.3146 rb:1.0584 dl:321-322 gd:1 +ttp: b215/782 bl:2.3919 bb:1.0964 rl:2.3148 rb:1.0585 dl:312-313 gd:1 +ttp: b207/782 bl:2.3487 bb:1.1287 rl:2.3149 rb:1.0588 dl:303-304 gd:1 +ttp: b199/782 bl:2.4358 bb:1.1460 rl:2.3153 rb:1.0591 dl:295-296 gd:1 +ttp: b191/782 bl:2.4133 bb:1.0979 rl:2.3157 rb:1.0592 dl:285-286 gd:1 +ttp: b183/782 bl:2.3270 bb:1.0717 rl:2.3157 rb:1.0592 dl:277-278 gd:1 +ttp: b175/782 bl:2.3892 bb:1.1545 rl:2.3159 rb:1.0595 dl:269-270 gd:1 +ttp: b168/782 bl:2.4579 bb:1.1891 rl:2.3163 rb:1.0598 dl:263-263 gd:1 +ttp: b160/782 bl:2.3851 bb:1.1138 rl:2.3165 rb:1.0600 dl:255-255 gd:1 +ttp: b151/782 bl:2.4622 bb:1.1382 rl:2.3169 rb:1.0602 dl:246-247 gd:1 +ttp: b142/782 bl:2.3797 bb:1.1077 rl:2.3171 rb:1.0603 dl:237-238 gd:1 +ttp: b134/782 bl:2.4262 bb:1.1377 rl:2.3174 rb:1.0605 dl:230-231 gd:1 +ttp: b126/782 bl:2.3935 bb:1.1402 rl:2.3175 rb:1.0607 dl:222-223 gd:1 +ttp: b118/782 bl:2.4628 bb:1.1241 rl:2.3179 rb:1.0608 dl:215-216 gd:1 +ttp: b109/782 bl:2.4960 bb:1.1895 rl:2.3183 rb:1.0611 dl:207-208 gd:1 +ttp: b103/782 bl:2.4622 bb:1.1852 rl:2.3186 rb:1.0614 dl:202-202 gd:1 +ttp: b94/782 bl:2.5631 bb:1.2111 rl:2.3191 rb:1.0617 dl:193-194 gd:1 +ttp: b81/782 bl:2.4744 bb:1.1230 rl:2.3194 rb:1.0618 dl:182-183 gd:1 +ttp: b74/782 bl:2.4824 bb:1.1520 rl:2.3197 rb:1.0620 dl:175-176 gd:1 +ttp: b66/782 bl:2.6397 bb:1.2353 rl:2.3203 rb:1.0623 dl:169-169 gd:1 +ttp: b57/782 bl:2.4635 bb:1.1601 rl:2.3206 rb:1.0625 dl:160-161 gd:1 +ttp: b48/782 bl:2.5261 bb:1.2179 rl:2.3209 rb:1.0627 dl:151-152 gd:1 +ttp: b39/782 bl:2.4423 bb:1.1822 rl:2.3211 rb:1.0629 dl:142-143 gd:1 +ttp: b32/782 bl:2.6105 bb:1.2172 rl:2.3215 rb:1.0631 dl:135-136 gd:1 +ttp: b23/782 bl:2.5967 bb:1.2196 rl:2.3219 rb:1.0633 dl:126-127 gd:1 +ttp: b15/782 bl:2.6562 bb:1.2336 rl:2.3223 rb:1.0635 dl:115-117 gd:1 +ttp: b5/782 bl:2.7223 bb:1.2386 rl:2.3227 rb:1.0637 dl:96-99 gd:1 +quantized_ttt_phased val_loss:2.32262534 val_bpb:1.06134915 eval_time:395545ms +total_eval_time:395.5s +=== Done: SEED=1234 === diff --git a/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_seed314.log b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_seed314.log new file mode 100644 index 0000000000..9b5c27486f --- /dev/null +++ b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_seed314.log @@ -0,0 +1,951 @@ +=== Starting training: SEED=314 RUN_ID=awq_seed314_20260429_052651 === +=== Artifact dir: /workspace/pargolf-work/runs/awq_seed314_20260429_052651 === +=== NPROC=8 === +W0429 05:26:52.995000 181313 torch/distributed/run.py:803] +W0429 05:26:52.995000 181313 torch/distributed/run.py:803] ***************************************** +W0429 05:26:52.995000 181313 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0429 05:26:52.995000 181313 torch/distributed/run.py:803] ***************************************** +Hyperparameters: + adam_eps: 1e-08 + adam_wd: 0.02 + artifact_dir: /workspace/pargolf-work/runs/awq_seed314_20260429_052651 + attn_clip_sigmas: 13.0 + attn_out_gate_enabled: False + attn_out_gate_src: proj + awq_lite_bits: 8 + awq_lite_enabled: True + awq_lite_group_size: 64 + awq_lite_group_top_k: 1 + beta1: 0.9 + beta2: 0.99 + caseops_enabled: True + compressor: pergroup + data_dir: ./data/ + datasets_dir: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved + distributed: True + ema_decay: 0.9965 + embed_bits: 7 + embed_clip_sigmas: 14.0 + embed_lr: 0.6 + embed_wd: 0.085 + enable_looping_at: 0.35 + eval_seq_len: 2048 + eval_stride: 64 + fused_ce_enabled: True + gate_window: 12 + gated_attn_enabled: False + gated_attn_init_std: 0.01 + gated_attn_quant_gate: True + global_ttt_batch_seqs: 32 + global_ttt_chunk_tokens: 32768 + global_ttt_epochs: 1 + global_ttt_grad_clip: 1.0 + global_ttt_lr: 0.001 + global_ttt_momentum: 0.9 + global_ttt_respect_doc_boundaries: True + global_ttt_warmup_chunks: 0 + global_ttt_warmup_start_lr: 0.0 + gptq_calibration_batches: 16 + gptq_reserve_seconds: 0.5 + grad_accum_steps: 1 + grad_clip_norm: 0.3 + is_main_process: True + iterations: 20000 + ln_scale: True + local_rank: 0 + logfile: /workspace/pargolf-work/runs/awq_seed314_20260429_052651/awq_seed314_20260429_052651.txt + logit_softcap: 30.0 + loop_end: 5 + loop_start: 3 + lqer_asym_enabled: True + lqer_asym_group: 64 + lqer_enabled: True + lqer_factor_bits: 4 + lqer_gain_select: False + lqer_rank: 4 + lqer_scope: all + lqer_top_k: 3 + matrix_bits: 6 + matrix_clip_sigmas: 12.85 + matrix_lr: 0.026 + max_wallclock_seconds: 600.0 + min_lr: 0.1 + mlp_clip_sigmas: 11.5 + mlp_mult: 4.0 + model_dim: 512 + model_path: /workspace/pargolf-work/runs/awq_seed314_20260429_052651/final_model.pt + muon_backend_steps: 5 + muon_momentum: 0.97 + muon_momentum_warmup_start: 0.92 + muon_momentum_warmup_steps: 1500 + muon_row_normalize: True + muon_wd: 0.095 + num_heads: 8 + num_kv_heads: 4 + num_layers: 11 + num_loops: 2 + parallel_final_lane: mean + parallel_start_layer: 8 + phased_ttt_num_phases: 3 + phased_ttt_prefix_docs: 2500 + qk_gain_init: 5.0 + quantized_model_path: /workspace/pargolf-work/runs/awq_seed314_20260429_052651/final_model.int6.ptz + rank: 0 + rope_base: 10000.0 + rope_dims: 16 + rope_train_seq_len: 2048 + rope_yarn: False + run_id: awq_seed314_20260429_052651 + scalar_lr: 0.02 + seed: 314 + skip_gates_enabled: True + smear_gate_enabled: True + sparse_attn_gate_enabled: True + sparse_attn_gate_init_std: 0.0 + sparse_attn_gate_scale: 0.5 + tie_embeddings: True + tied_embed_init_std: 0.005 + tied_embed_lr: 0.03 + tokenizer_path: /workspace/pargolf/records/track_10min_16mb/2026-04-27_SP8192_LQER_SparseGate_BOSSmearFix_9HpStack_1.0611/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model + train_batch_tokens: 786432 + train_files: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/fineweb_train_*.bin + train_log_every: 500 + train_seq_len: 2048 + ttt_batch_size: 64 + ttt_beta1: 0.0 + ttt_beta2: 0.99 + ttt_chunk_size: 48 + ttt_enabled: True + ttt_eval_batches: + ttt_eval_seq_len: 2048 + ttt_grad_steps: 1 + ttt_k_lora: True + ttt_lora_lr: 0.0001 + ttt_lora_rank: 80 + ttt_mlp_lora: True + ttt_o_lora: True + ttt_optimizer: adam + ttt_weight_decay: 0.5 + val_batch_tokens: 524288 + val_bytes_files: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/fineweb_val_bytes_*.bin + val_doc_fraction: 1.0 + val_files: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/fineweb_val_*.bin + val_loss_every: 4000 + vocab_size: 8192 + warmdown_frac: 0.85 + warmup_steps: 20 + world_size: 8 + xsa_last_n: 11 +train_shards: 80 +val_tokens: 47851520 +model_params:35945671 +gptq:reserving 0s, effective=599500ms +warmup_cu_buckets:64,128,192,256 iters_each:3 +warmup_step: 1/20 +warmup_step: 2/20 +warmup_step: 3/20 +warmup_step: 4/20 +warmup_step: 5/20 +warmup_step: 6/20 +warmup_step: 10/20 +warmup_step: 20/20 +loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +loop_warmup_step: 1/20 +loop_warmup_step: 2/20 +loop_warmup_step: 3/20 +loop_warmup_step: 4/20 +loop_warmup_step: 5/20 +loop_warmup_step: 6/20 +loop_warmup_step: 10/20 +loop_warmup_step: 20/20 +0/20000 val_loss: 8.9980 val_bpb: 4.1115 +1/20000 train_loss: 9.0022 train_time: 0.0m tok/s: 17511015 +2/20000 train_loss: 12.6318 train_time: 0.0m tok/s: 11153491 +3/20000 train_loss: 10.3697 train_time: 0.0m tok/s: 9879286 +4/20000 train_loss: 9.1871 train_time: 0.0m tok/s: 9333995 +5/20000 train_loss: 7.9130 train_time: 0.0m tok/s: 9041301 +500/20000 train_loss: 2.6141 train_time: 0.8m tok/s: 8313359 +1000/20000 train_loss: 2.7425 train_time: 1.6m tok/s: 8306581 +1500/20000 train_loss: 2.7117 train_time: 2.4m tok/s: 8292448 +2000/20000 train_loss: 2.6242 train_time: 3.2m tok/s: 8294631 +layer_loop:enabled step:2212 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +2500/20000 train_loss: 2.4674 train_time: 4.2m tok/s: 7868142 +3000/20000 train_loss: 2.4906 train_time: 5.4m tok/s: 7328089 +3500/20000 train_loss: 2.6190 train_time: 6.5m tok/s: 7029380 +4000/20000 train_loss: 2.3655 train_time: 7.7m tok/s: 6804880 +4000/20000 val_loss: 2.4283 val_bpb: 1.1095 +4500/20000 train_loss: 2.5321 train_time: 8.9m tok/s: 6653369 +4986/20000 val_loss: 2.3549 val_bpb: 1.0760 +stopping_early: wallclock_cap train_time: 599550ms step: 4986/20000 +peak memory allocated: 41707 MiB reserved: 47048 MiB +ema:applying EMA weights +diagnostic pre-quantization post-ema val_loss:2.32934097 val_bpb:1.06434881 eval_time:7976ms +Serialized model: 135417533 bytes +Code size (uncompressed): 168543 bytes +Code size (compressed): 33787 bytes +GPTQ:collecting Hessians from calibration data... +GPTQ:collected 67 Hessians in 4.0s +Quantized weights: + gate_int8_row: blocks.attn.attn_gate_w + gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight + gptq (int6)+lqer_asym: blocks.mlp.fc.weight + gptq (int7)+awqgrpint8+lqer_asym: tok_emb.weight + passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, parallel_post_lambdas, parallel_resid_lambdas, skip_gates, skip_weights, smear_gate.weight, smear_lambda +Serialize: per-group lrzip compression... +Serialize: per-group compression done in 117.9s +Serialized model quantized+pergroup: 15942682 bytes +Total submission size quantized+pergroup: 15976469 bytes +Deserialize: per-group lrzip decompression... +Deserialize: decompression done in 19.3s +diagnostic quantized val_loss:2.34773220 val_bpb:1.07275234 eval_time:12677ms +Deserialize: per-group lrzip decompression... +Deserialize: decompression done in 19.0s +ttt_lora:warming up compile (random tokens, no val data) +ttt_lora:compile warmup done (90.3s) + +beginning TTT eval timer +ttt_phased: total_docs:50000 prefix_docs:2500 suffix_docs:47500 num_phases:3 boundaries:[833, 1666, 2500] +ttp: b777/782 bl:2.3105 bb:1.0828 rl:2.3105 rb:1.0828 dl:8452-9229 gd:0 +ttp: b772/782 bl:2.3255 bb:1.0960 rl:2.3165 rb:1.0881 dl:5762-6095 gd:0 +ttp: b767/782 bl:2.2678 bb:1.0731 rl:2.3046 rb:1.0844 dl:4681-4858 gd:0 +ttp: b762/782 bl:2.3513 bb:1.0888 rl:2.3127 rb:1.0852 dl:4032-4142 gd:0 +ttpp: phase:1/3 pd:1296 gd:833 t:169.0s +tttg: c1/131 lr:0.001000 t:0.4s +tttg: c2/131 lr:0.001000 t:0.5s +tttg: c3/131 lr:0.000999 t:0.5s +tttg: c4/131 lr:0.000999 t:0.6s +tttg: c5/131 lr:0.000998 t:0.7s +tttg: c6/131 lr:0.000996 t:0.7s +tttg: c7/131 lr:0.000995 t:0.8s +tttg: c8/131 lr:0.000993 t:0.9s +tttg: c9/131 lr:0.000991 t:0.9s +tttg: c10/131 lr:0.000988 t:1.0s +tttg: c11/131 lr:0.000985 t:1.1s +tttg: c12/131 lr:0.000982 t:1.1s +tttg: c13/131 lr:0.000979 t:1.2s +tttg: c14/131 lr:0.000976 t:1.2s +tttg: c15/131 lr:0.000972 t:1.3s +tttg: c16/131 lr:0.000968 t:1.4s +tttg: c17/131 lr:0.000963 t:1.4s +tttg: c18/131 lr:0.000958 t:1.5s +tttg: c19/131 lr:0.000953 t:1.6s +tttg: c20/131 lr:0.000948 t:1.6s +tttg: c21/131 lr:0.000943 t:1.7s +tttg: c22/131 lr:0.000937 t:1.8s +tttg: c23/131 lr:0.000931 t:1.8s +tttg: c24/131 lr:0.000925 t:1.9s +tttg: c25/131 lr:0.000918 t:2.0s +tttg: c26/131 lr:0.000911 t:2.0s +tttg: c27/131 lr:0.000905 t:2.1s +tttg: c28/131 lr:0.000897 t:2.2s +tttg: c29/131 lr:0.000890 t:2.2s +tttg: c30/131 lr:0.000882 t:2.3s +tttg: c31/131 lr:0.000874 t:2.3s +tttg: c32/131 lr:0.000866 t:2.4s +tttg: c33/131 lr:0.000858 t:2.5s +tttg: c34/131 lr:0.000849 t:2.5s +tttg: c35/131 lr:0.000841 t:2.6s +tttg: c36/131 lr:0.000832 t:2.7s +tttg: c37/131 lr:0.000822 t:2.7s +tttg: c38/131 lr:0.000813 t:2.8s +tttg: c39/131 lr:0.000804 t:2.9s +tttg: c40/131 lr:0.000794 t:2.9s +tttg: c41/131 lr:0.000784 t:3.0s +tttg: c42/131 lr:0.000774 t:3.1s +tttg: c43/131 lr:0.000764 t:3.1s +tttg: c44/131 lr:0.000753 t:3.2s +tttg: c45/131 lr:0.000743 t:3.3s +tttg: c46/131 lr:0.000732 t:3.3s +tttg: c47/131 lr:0.000722 t:3.4s +tttg: c48/131 lr:0.000711 t:3.5s +tttg: c49/131 lr:0.000700 t:3.5s +tttg: c50/131 lr:0.000689 t:3.6s +tttg: c51/131 lr:0.000677 t:3.7s +tttg: c52/131 lr:0.000666 t:3.7s +tttg: c53/131 lr:0.000655 t:3.8s +tttg: c54/131 lr:0.000643 t:3.8s +tttg: c55/131 lr:0.000631 t:3.9s +tttg: c56/131 lr:0.000620 t:4.0s +tttg: c57/131 lr:0.000608 t:4.0s +tttg: c58/131 lr:0.000596 t:4.1s +tttg: c59/131 lr:0.000584 t:4.2s +tttg: c60/131 lr:0.000572 t:4.2s +tttg: c61/131 lr:0.000560 t:4.3s +tttg: c62/131 lr:0.000548 t:4.4s +tttg: c63/131 lr:0.000536 t:4.4s +tttg: c64/131 lr:0.000524 t:4.5s +tttg: c65/131 lr:0.000512 t:4.6s +tttg: c66/131 lr:0.000500 t:4.6s +tttg: c67/131 lr:0.000488 t:4.7s +tttg: c68/131 lr:0.000476 t:4.8s +tttg: c69/131 lr:0.000464 t:4.8s +tttg: c70/131 lr:0.000452 t:4.9s +tttg: c71/131 lr:0.000440 t:4.9s +tttg: c72/131 lr:0.000428 t:5.0s +tttg: c73/131 lr:0.000416 t:5.1s +tttg: c74/131 lr:0.000404 t:5.1s +tttg: c75/131 lr:0.000392 t:5.2s +tttg: c76/131 lr:0.000380 t:5.3s +tttg: c77/131 lr:0.000369 t:5.3s +tttg: c78/131 lr:0.000357 t:5.4s +tttg: c79/131 lr:0.000345 t:5.5s +tttg: c80/131 lr:0.000334 t:5.5s +tttg: c81/131 lr:0.000323 t:5.6s +tttg: c82/131 lr:0.000311 t:5.7s +tttg: c83/131 lr:0.000300 t:5.7s +tttg: c84/131 lr:0.000289 t:5.8s +tttg: c85/131 lr:0.000278 t:5.8s +tttg: c86/131 lr:0.000268 t:5.9s +tttg: c87/131 lr:0.000257 t:6.0s +tttg: c88/131 lr:0.000247 t:6.0s +tttg: c89/131 lr:0.000236 t:6.1s +tttg: c90/131 lr:0.000226 t:6.2s +tttg: c91/131 lr:0.000216 t:6.2s +tttg: c92/131 lr:0.000206 t:6.3s +tttg: c93/131 lr:0.000196 t:6.4s +tttg: c94/131 lr:0.000187 t:6.4s +tttg: c95/131 lr:0.000178 t:6.5s +tttg: c96/131 lr:0.000168 t:6.6s +tttg: c97/131 lr:0.000159 t:6.6s +tttg: c98/131 lr:0.000151 t:6.7s +tttg: c99/131 lr:0.000142 t:6.8s +tttg: c100/131 lr:0.000134 t:6.8s +tttg: c101/131 lr:0.000126 t:6.9s +tttg: c102/131 lr:0.000118 t:6.9s +tttg: c103/131 lr:0.000110 t:7.0s +tttg: c104/131 lr:0.000103 t:7.1s +tttg: c105/131 lr:0.000095 t:7.1s +tttg: c106/131 lr:0.000089 t:7.2s +tttg: c107/131 lr:0.000082 t:7.3s +tttg: c108/131 lr:0.000075 t:7.3s +tttg: c109/131 lr:0.000069 t:7.4s +tttg: c110/131 lr:0.000063 t:7.5s +tttg: c111/131 lr:0.000057 t:7.5s +tttg: c112/131 lr:0.000052 t:7.6s +tttg: c113/131 lr:0.000047 t:7.7s +tttg: c114/131 lr:0.000042 t:7.7s +tttg: c115/131 lr:0.000037 t:7.8s +tttg: c116/131 lr:0.000032 t:7.9s +tttg: c117/131 lr:0.000028 t:7.9s +tttg: c118/131 lr:0.000024 t:8.0s +tttg: c119/131 lr:0.000021 t:8.0s +tttg: c120/131 lr:0.000018 t:8.1s +tttg: c121/131 lr:0.000015 t:8.2s +tttg: c122/131 lr:0.000012 t:8.2s +tttg: c123/131 lr:0.000009 t:8.3s +tttg: c124/131 lr:0.000007 t:8.4s +tttg: c125/131 lr:0.000005 t:8.4s +tttg: c126/131 lr:0.000004 t:8.5s +tttg: c127/131 lr:0.000002 t:8.6s +tttg: c128/131 lr:0.000001 t:8.6s +tttg: c129/131 lr:0.000001 t:8.7s +tttg: c130/131 lr:0.000000 t:8.8s +ttpr: phase:1/3 t:179.3s +ttp: b754/782 bl:2.2853 bb:1.0571 rl:2.3093 rb:1.0817 dl:3345-3397 gd:0 +ttp: b752/782 bl:2.3282 bb:1.0702 rl:2.3113 rb:1.0804 dl:3222-3283 gd:0 +ttpp: phase:2/3 pd:2128 gd:1666 t:245.5s +tttg: c1/219 lr:0.001000 t:0.1s +tttg: c2/219 lr:0.001000 t:0.1s +tttg: c3/219 lr:0.001000 t:0.2s +tttg: c4/219 lr:0.001000 t:0.3s +tttg: c5/219 lr:0.000999 t:0.3s +tttg: c6/219 lr:0.000999 t:0.4s +tttg: c7/219 lr:0.000998 t:0.5s +tttg: c8/219 lr:0.000997 t:0.5s +tttg: c9/219 lr:0.000997 t:0.6s +tttg: c10/219 lr:0.000996 t:0.7s +tttg: c11/219 lr:0.000995 t:0.7s +tttg: c12/219 lr:0.000994 t:0.8s +tttg: c13/219 lr:0.000993 t:0.9s +tttg: c14/219 lr:0.000991 t:0.9s +tttg: c15/219 lr:0.000990 t:1.0s +tttg: c16/219 lr:0.000988 t:1.0s +tttg: c17/219 lr:0.000987 t:1.1s +tttg: c18/219 lr:0.000985 t:1.2s +tttg: c19/219 lr:0.000983 t:1.3s +tttg: c20/219 lr:0.000981 t:1.3s +tttg: c21/219 lr:0.000979 t:1.4s +tttg: c22/219 lr:0.000977 t:1.4s +tttg: c23/219 lr:0.000975 t:1.5s +tttg: c24/219 lr:0.000973 t:1.6s +tttg: c25/219 lr:0.000970 t:1.6s +tttg: c26/219 lr:0.000968 t:1.7s +tttg: c27/219 lr:0.000965 t:1.8s +tttg: c28/219 lr:0.000963 t:1.8s +tttg: c29/219 lr:0.000960 t:1.9s +tttg: c30/219 lr:0.000957 t:2.0s +tttg: c31/219 lr:0.000954 t:2.0s +tttg: c32/219 lr:0.000951 t:2.1s +tttg: c33/219 lr:0.000948 t:2.2s +tttg: c34/219 lr:0.000945 t:2.2s +tttg: c35/219 lr:0.000941 t:2.3s +tttg: c36/219 lr:0.000938 t:2.4s +tttg: c37/219 lr:0.000934 t:2.4s +tttg: c38/219 lr:0.000931 t:2.5s +tttg: c39/219 lr:0.000927 t:2.6s +tttg: c40/219 lr:0.000923 t:2.6s +tttg: c41/219 lr:0.000919 t:2.7s +tttg: c42/219 lr:0.000915 t:2.7s +tttg: c43/219 lr:0.000911 t:2.8s +tttg: c44/219 lr:0.000907 t:2.9s +tttg: c45/219 lr:0.000903 t:2.9s +tttg: c46/219 lr:0.000898 t:3.0s +tttg: c47/219 lr:0.000894 t:3.1s +tttg: c48/219 lr:0.000890 t:3.1s +tttg: c49/219 lr:0.000885 t:3.2s +tttg: c50/219 lr:0.000880 t:3.3s +tttg: c51/219 lr:0.000876 t:3.3s +tttg: c52/219 lr:0.000871 t:3.4s +tttg: c53/219 lr:0.000866 t:3.5s +tttg: c54/219 lr:0.000861 t:3.5s +tttg: c55/219 lr:0.000856 t:3.6s +tttg: c56/219 lr:0.000851 t:3.7s +tttg: c57/219 lr:0.000846 t:3.7s +tttg: c58/219 lr:0.000841 t:3.8s +tttg: c59/219 lr:0.000835 t:3.9s +tttg: c60/219 lr:0.000830 t:3.9s +tttg: c61/219 lr:0.000824 t:4.0s +tttg: c62/219 lr:0.000819 t:4.0s +tttg: c63/219 lr:0.000813 t:4.1s +tttg: c64/219 lr:0.000808 t:4.2s +tttg: c65/219 lr:0.000802 t:4.2s +tttg: c66/219 lr:0.000796 t:4.3s +tttg: c67/219 lr:0.000790 t:4.4s +tttg: c68/219 lr:0.000784 t:4.4s +tttg: c69/219 lr:0.000779 t:4.5s +tttg: c70/219 lr:0.000773 t:4.6s +tttg: c71/219 lr:0.000766 t:4.6s +tttg: c72/219 lr:0.000760 t:4.7s +tttg: c73/219 lr:0.000754 t:4.8s +tttg: c74/219 lr:0.000748 t:4.8s +tttg: c75/219 lr:0.000742 t:4.9s +tttg: c76/219 lr:0.000735 t:5.0s +tttg: c77/219 lr:0.000729 t:5.0s +tttg: c78/219 lr:0.000722 t:5.1s +tttg: c79/219 lr:0.000716 t:5.2s +tttg: c80/219 lr:0.000709 t:5.2s +tttg: c81/219 lr:0.000703 t:5.3s +tttg: c82/219 lr:0.000696 t:5.4s +tttg: c83/219 lr:0.000690 t:5.4s +tttg: c84/219 lr:0.000683 t:5.5s +tttg: c85/219 lr:0.000676 t:5.6s +tttg: c86/219 lr:0.000670 t:5.6s +tttg: c87/219 lr:0.000663 t:5.7s +tttg: c88/219 lr:0.000656 t:5.7s +tttg: c89/219 lr:0.000649 t:5.8s +tttg: c90/219 lr:0.000642 t:5.9s +tttg: c91/219 lr:0.000635 t:5.9s +tttg: c92/219 lr:0.000628 t:6.0s +tttg: c93/219 lr:0.000621 t:6.1s +tttg: c94/219 lr:0.000614 t:6.1s +tttg: c95/219 lr:0.000607 t:6.2s +tttg: c96/219 lr:0.000600 t:6.3s +tttg: c97/219 lr:0.000593 t:6.3s +tttg: c98/219 lr:0.000586 t:6.4s +tttg: c99/219 lr:0.000579 t:6.5s +tttg: c100/219 lr:0.000572 t:6.5s +tttg: c101/219 lr:0.000565 t:6.6s +tttg: c102/219 lr:0.000558 t:6.7s +tttg: c103/219 lr:0.000550 t:6.7s +tttg: c104/219 lr:0.000543 t:6.8s +tttg: c105/219 lr:0.000536 t:6.9s +tttg: c106/219 lr:0.000529 t:6.9s +tttg: c107/219 lr:0.000522 t:7.0s +tttg: c108/219 lr:0.000514 t:7.1s +tttg: c109/219 lr:0.000507 t:7.1s +tttg: c110/219 lr:0.000500 t:7.2s +tttg: c111/219 lr:0.000493 t:7.3s +tttg: c112/219 lr:0.000486 t:7.3s +tttg: c113/219 lr:0.000478 t:7.4s +tttg: c114/219 lr:0.000471 t:7.4s +tttg: c115/219 lr:0.000464 t:7.5s +tttg: c116/219 lr:0.000457 t:7.6s +tttg: c117/219 lr:0.000450 t:7.6s +tttg: c118/219 lr:0.000442 t:7.7s +tttg: c119/219 lr:0.000435 t:7.8s +tttg: c120/219 lr:0.000428 t:7.8s +tttg: c121/219 lr:0.000421 t:7.9s +tttg: c122/219 lr:0.000414 t:8.0s +tttg: c123/219 lr:0.000407 t:8.0s +tttg: c124/219 lr:0.000400 t:8.1s +tttg: c125/219 lr:0.000393 t:8.2s +tttg: c126/219 lr:0.000386 t:8.2s +tttg: c127/219 lr:0.000379 t:8.3s +tttg: c128/219 lr:0.000372 t:8.4s +tttg: c129/219 lr:0.000365 t:8.4s +tttg: c130/219 lr:0.000358 t:8.5s +tttg: c131/219 lr:0.000351 t:8.6s +tttg: c132/219 lr:0.000344 t:8.6s +tttg: c133/219 lr:0.000337 t:8.7s +tttg: c134/219 lr:0.000330 t:8.8s +tttg: c135/219 lr:0.000324 t:8.8s +tttg: c136/219 lr:0.000317 t:8.9s +tttg: c137/219 lr:0.000310 t:8.9s +tttg: c138/219 lr:0.000304 t:9.0s +tttg: c139/219 lr:0.000297 t:9.1s +tttg: c140/219 lr:0.000291 t:9.1s +tttg: c141/219 lr:0.000284 t:9.2s +tttg: c142/219 lr:0.000278 t:9.3s +tttg: c143/219 lr:0.000271 t:9.3s +tttg: c144/219 lr:0.000265 t:9.4s +tttg: c145/219 lr:0.000258 t:9.5s +tttg: c146/219 lr:0.000252 t:9.5s +tttg: c147/219 lr:0.000246 t:9.6s +tttg: c148/219 lr:0.000240 t:9.7s +tttg: c149/219 lr:0.000234 t:9.7s +tttg: c150/219 lr:0.000227 t:9.8s +tttg: c151/219 lr:0.000221 t:9.9s +tttg: c152/219 lr:0.000216 t:9.9s +tttg: c153/219 lr:0.000210 t:10.0s +tttg: c154/219 lr:0.000204 t:10.1s +tttg: c155/219 lr:0.000198 t:10.1s +tttg: c156/219 lr:0.000192 t:10.2s +tttg: c157/219 lr:0.000187 t:10.3s +tttg: c158/219 lr:0.000181 t:10.3s +tttg: c159/219 lr:0.000176 t:10.4s +tttg: c160/219 lr:0.000170 t:10.5s +tttg: c161/219 lr:0.000165 t:10.5s +tttg: c162/219 lr:0.000159 t:10.6s +tttg: c163/219 lr:0.000154 t:10.6s +tttg: c164/219 lr:0.000149 t:10.7s +tttg: c165/219 lr:0.000144 t:10.8s +tttg: c166/219 lr:0.000139 t:10.8s +tttg: c167/219 lr:0.000134 t:10.9s +tttg: c168/219 lr:0.000129 t:11.0s +tttg: c169/219 lr:0.000124 t:11.0s +tttg: c170/219 lr:0.000120 t:11.1s +tttg: c171/219 lr:0.000115 t:11.2s +tttg: c172/219 lr:0.000110 t:11.2s +tttg: c173/219 lr:0.000106 t:11.3s +tttg: c174/219 lr:0.000102 t:11.4s +tttg: c175/219 lr:0.000097 t:11.4s +tttg: c176/219 lr:0.000093 t:11.5s +tttg: c177/219 lr:0.000089 t:11.6s +tttg: c178/219 lr:0.000085 t:11.6s +tttg: c179/219 lr:0.000081 t:11.7s +tttg: c180/219 lr:0.000077 t:11.8s +tttg: c181/219 lr:0.000073 t:11.8s +tttg: c182/219 lr:0.000069 t:11.9s +tttg: c183/219 lr:0.000066 t:11.9s +tttg: c184/219 lr:0.000062 t:12.0s +tttg: c185/219 lr:0.000059 t:12.1s +tttg: c186/219 lr:0.000055 t:12.1s +tttg: c187/219 lr:0.000052 t:12.2s +tttg: c188/219 lr:0.000049 t:12.3s +tttg: c189/219 lr:0.000046 t:12.3s +tttg: c190/219 lr:0.000043 t:12.4s +tttg: c191/219 lr:0.000040 t:12.5s +tttg: c192/219 lr:0.000037 t:12.5s +tttg: c193/219 lr:0.000035 t:12.6s +tttg: c194/219 lr:0.000032 t:12.7s +tttg: c195/219 lr:0.000030 t:12.7s +tttg: c196/219 lr:0.000027 t:12.8s +tttg: c197/219 lr:0.000025 t:12.9s +tttg: c198/219 lr:0.000023 t:12.9s +tttg: c199/219 lr:0.000021 t:13.0s +tttg: c200/219 lr:0.000019 t:13.1s +tttg: c201/219 lr:0.000017 t:13.1s +tttg: c202/219 lr:0.000015 t:13.2s +tttg: c203/219 lr:0.000013 t:13.3s +tttg: c204/219 lr:0.000012 t:13.3s +tttg: c205/219 lr:0.000010 t:13.4s +tttg: c206/219 lr:0.000009 t:13.5s +tttg: c207/219 lr:0.000007 t:13.5s +tttg: c208/219 lr:0.000006 t:13.6s +tttg: c209/219 lr:0.000005 t:13.6s +tttg: c210/219 lr:0.000004 t:13.7s +tttg: c211/219 lr:0.000003 t:13.8s +tttg: c212/219 lr:0.000003 t:13.8s +tttg: c213/219 lr:0.000002 t:13.9s +tttg: c214/219 lr:0.000001 t:14.0s +tttg: c215/219 lr:0.000001 t:14.0s +tttg: c216/219 lr:0.000000 t:14.1s +tttg: c217/219 lr:0.000000 t:14.2s +tttg: c218/219 lr:0.000000 t:14.2s +ttpr: phase:2/3 t:261.3s +ttp: b741/782 bl:2.3211 bb:1.0409 rl:2.3121 rb:1.0770 dl:2686-2730 gd:0 +ttp: b740/782 bl:2.2613 bb:1.0381 rl:2.3083 rb:1.0741 dl:2653-2686 gd:0 +ttpp: phase:3/3 pd:2960 gd:2500 t:276.4s +tttg: c1/289 lr:0.001000 t:0.1s +tttg: c2/289 lr:0.001000 t:0.1s +tttg: c3/289 lr:0.001000 t:0.2s +tttg: c4/289 lr:0.001000 t:0.3s +tttg: c5/289 lr:0.001000 t:0.3s +tttg: c6/289 lr:0.000999 t:0.4s +tttg: c7/289 lr:0.000999 t:0.5s +tttg: c8/289 lr:0.000999 t:0.5s +tttg: c9/289 lr:0.000998 t:0.6s +tttg: c10/289 lr:0.000998 t:0.7s +tttg: c11/289 lr:0.000997 t:0.7s +tttg: c12/289 lr:0.000996 t:0.8s +tttg: c13/289 lr:0.000996 t:0.8s +tttg: c14/289 lr:0.000995 t:0.9s +tttg: c15/289 lr:0.000994 t:1.0s +tttg: c16/289 lr:0.000993 t:1.0s +tttg: c17/289 lr:0.000992 t:1.1s +tttg: c18/289 lr:0.000991 t:1.2s +tttg: c19/289 lr:0.000990 t:1.2s +tttg: c20/289 lr:0.000989 t:1.3s +tttg: c21/289 lr:0.000988 t:1.4s +tttg: c22/289 lr:0.000987 t:1.4s +tttg: c23/289 lr:0.000986 t:1.5s +tttg: c24/289 lr:0.000984 t:1.6s +tttg: c25/289 lr:0.000983 t:1.6s +tttg: c26/289 lr:0.000982 t:1.7s +tttg: c27/289 lr:0.000980 t:1.8s +tttg: c28/289 lr:0.000978 t:1.8s +tttg: c29/289 lr:0.000977 t:1.9s +tttg: c30/289 lr:0.000975 t:2.0s +tttg: c31/289 lr:0.000973 t:2.0s +tttg: c32/289 lr:0.000972 t:2.1s +tttg: c33/289 lr:0.000970 t:2.1s +tttg: c34/289 lr:0.000968 t:2.2s +tttg: c35/289 lr:0.000966 t:2.3s +tttg: c36/289 lr:0.000964 t:2.3s +tttg: c37/289 lr:0.000962 t:2.4s +tttg: c38/289 lr:0.000960 t:2.5s +tttg: c39/289 lr:0.000958 t:2.5s +tttg: c40/289 lr:0.000955 t:2.6s +tttg: c41/289 lr:0.000953 t:2.7s +tttg: c42/289 lr:0.000951 t:2.7s +tttg: c43/289 lr:0.000948 t:2.8s +tttg: c44/289 lr:0.000946 t:2.9s +tttg: c45/289 lr:0.000944 t:2.9s +tttg: c46/289 lr:0.000941 t:3.0s +tttg: c47/289 lr:0.000938 t:3.1s +tttg: c48/289 lr:0.000936 t:3.1s +tttg: c49/289 lr:0.000933 t:3.2s +tttg: c50/289 lr:0.000930 t:3.3s +tttg: c51/289 lr:0.000927 t:3.3s +tttg: c52/289 lr:0.000925 t:3.4s +tttg: c53/289 lr:0.000922 t:3.5s +tttg: c54/289 lr:0.000919 t:3.5s +tttg: c55/289 lr:0.000916 t:3.6s +tttg: c56/289 lr:0.000913 t:3.7s +tttg: c57/289 lr:0.000910 t:3.7s +tttg: c58/289 lr:0.000906 t:3.8s +tttg: c59/289 lr:0.000903 t:3.8s +tttg: c60/289 lr:0.000900 t:3.9s +tttg: c61/289 lr:0.000897 t:4.0s +tttg: c62/289 lr:0.000893 t:4.0s +tttg: c63/289 lr:0.000890 t:4.1s +tttg: c64/289 lr:0.000887 t:4.2s +tttg: c65/289 lr:0.000883 t:4.2s +tttg: c66/289 lr:0.000879 t:4.3s +tttg: c67/289 lr:0.000876 t:4.4s +tttg: c68/289 lr:0.000872 t:4.4s +tttg: c69/289 lr:0.000869 t:4.5s +tttg: c70/289 lr:0.000865 t:4.6s +tttg: c71/289 lr:0.000861 t:4.6s +tttg: c72/289 lr:0.000857 t:4.7s +tttg: c73/289 lr:0.000854 t:4.8s +tttg: c74/289 lr:0.000850 t:4.8s +tttg: c75/289 lr:0.000846 t:4.9s +tttg: c76/289 lr:0.000842 t:5.0s +tttg: c77/289 lr:0.000838 t:5.0s +tttg: c78/289 lr:0.000834 t:5.1s +tttg: c79/289 lr:0.000830 t:5.2s +tttg: c80/289 lr:0.000826 t:5.2s +tttg: c81/289 lr:0.000821 t:5.3s +tttg: c82/289 lr:0.000817 t:5.4s +tttg: c83/289 lr:0.000813 t:5.4s +tttg: c84/289 lr:0.000809 t:5.5s +tttg: c85/289 lr:0.000804 t:5.6s +tttg: c86/289 lr:0.000800 t:5.6s +tttg: c87/289 lr:0.000796 t:5.7s +tttg: c88/289 lr:0.000791 t:5.7s +tttg: c89/289 lr:0.000787 t:5.8s +tttg: c90/289 lr:0.000782 t:5.9s +tttg: c91/289 lr:0.000778 t:5.9s +tttg: c92/289 lr:0.000773 t:6.0s +tttg: c93/289 lr:0.000769 t:6.1s +tttg: c94/289 lr:0.000764 t:6.1s +tttg: c95/289 lr:0.000759 t:6.2s +tttg: c96/289 lr:0.000755 t:6.3s +tttg: c97/289 lr:0.000750 t:6.3s +tttg: c98/289 lr:0.000745 t:6.4s +tttg: c99/289 lr:0.000740 t:6.5s +tttg: c100/289 lr:0.000736 t:6.5s +tttg: c101/289 lr:0.000731 t:6.6s +tttg: c102/289 lr:0.000726 t:6.7s +tttg: c103/289 lr:0.000721 t:6.7s +tttg: c104/289 lr:0.000716 t:6.8s +tttg: c105/289 lr:0.000711 t:6.9s +tttg: c106/289 lr:0.000706 t:6.9s +tttg: c107/289 lr:0.000701 t:7.0s +tttg: c108/289 lr:0.000696 t:7.1s +tttg: c109/289 lr:0.000691 t:7.1s +tttg: c110/289 lr:0.000686 t:7.2s +tttg: c111/289 lr:0.000681 t:7.2s +tttg: c112/289 lr:0.000676 t:7.3s +tttg: c113/289 lr:0.000671 t:7.4s +tttg: c114/289 lr:0.000666 t:7.4s +tttg: c115/289 lr:0.000661 t:7.5s +tttg: c116/289 lr:0.000656 t:7.6s +tttg: c117/289 lr:0.000650 t:7.6s +tttg: c118/289 lr:0.000645 t:7.7s +tttg: c119/289 lr:0.000640 t:7.8s +tttg: c120/289 lr:0.000635 t:7.8s +tttg: c121/289 lr:0.000629 t:7.9s +tttg: c122/289 lr:0.000624 t:8.0s +tttg: c123/289 lr:0.000619 t:8.0s +tttg: c124/289 lr:0.000614 t:8.1s +tttg: c125/289 lr:0.000608 t:8.2s +tttg: c126/289 lr:0.000603 t:8.2s +tttg: c127/289 lr:0.000598 t:8.3s +tttg: c128/289 lr:0.000592 t:8.3s +tttg: c129/289 lr:0.000587 t:8.4s +tttg: c130/289 lr:0.000581 t:8.5s +tttg: c131/289 lr:0.000576 t:8.5s +tttg: c132/289 lr:0.000571 t:8.6s +tttg: c133/289 lr:0.000565 t:8.7s +tttg: c134/289 lr:0.000560 t:8.7s +tttg: c135/289 lr:0.000554 t:8.8s +tttg: c136/289 lr:0.000549 t:8.9s +tttg: c137/289 lr:0.000544 t:8.9s +tttg: c138/289 lr:0.000538 t:9.0s +tttg: c139/289 lr:0.000533 t:9.1s +tttg: c140/289 lr:0.000527 t:9.1s +tttg: c141/289 lr:0.000522 t:9.2s +tttg: c142/289 lr:0.000516 t:9.3s +tttg: c143/289 lr:0.000511 t:9.3s +tttg: c144/289 lr:0.000505 t:9.4s +tttg: c145/289 lr:0.000500 t:9.4s +tttg: c146/289 lr:0.000495 t:9.5s +tttg: c147/289 lr:0.000489 t:9.6s +tttg: c148/289 lr:0.000484 t:9.6s +tttg: c149/289 lr:0.000478 t:9.7s +tttg: c150/289 lr:0.000473 t:9.8s +tttg: c151/289 lr:0.000467 t:9.8s +tttg: c152/289 lr:0.000462 t:9.9s +tttg: c153/289 lr:0.000456 t:10.0s +tttg: c154/289 lr:0.000451 t:10.0s +tttg: c155/289 lr:0.000446 t:10.1s +tttg: c156/289 lr:0.000440 t:10.2s +tttg: c157/289 lr:0.000435 t:10.2s +tttg: c158/289 lr:0.000429 t:10.3s +tttg: c159/289 lr:0.000424 t:10.4s +tttg: c160/289 lr:0.000419 t:10.4s +tttg: c161/289 lr:0.000413 t:10.5s +tttg: c162/289 lr:0.000408 t:10.6s +tttg: c163/289 lr:0.000402 t:10.6s +tttg: c164/289 lr:0.000397 t:10.7s +tttg: c165/289 lr:0.000392 t:10.8s +tttg: c166/289 lr:0.000386 t:10.8s +tttg: c167/289 lr:0.000381 t:10.9s +tttg: c168/289 lr:0.000376 t:10.9s +tttg: c169/289 lr:0.000371 t:11.0s +tttg: c170/289 lr:0.000365 t:11.1s +tttg: c171/289 lr:0.000360 t:11.1s +tttg: c172/289 lr:0.000355 t:11.2s +tttg: c173/289 lr:0.000350 t:11.3s +tttg: c174/289 lr:0.000344 t:11.3s +tttg: c175/289 lr:0.000339 t:11.4s +tttg: c176/289 lr:0.000334 t:11.5s +tttg: c177/289 lr:0.000329 t:11.5s +tttg: c178/289 lr:0.000324 t:11.6s +tttg: c179/289 lr:0.000319 t:11.7s +tttg: c180/289 lr:0.000314 t:11.7s +tttg: c181/289 lr:0.000309 t:11.8s +tttg: c182/289 lr:0.000304 t:11.9s +tttg: c183/289 lr:0.000299 t:11.9s +tttg: c184/289 lr:0.000294 t:12.0s +tttg: c185/289 lr:0.000289 t:12.0s +tttg: c186/289 lr:0.000284 t:12.1s +tttg: c187/289 lr:0.000279 t:12.2s +tttg: c188/289 lr:0.000274 t:12.2s +tttg: c189/289 lr:0.000269 t:12.3s +tttg: c190/289 lr:0.000264 t:12.4s +tttg: c191/289 lr:0.000260 t:12.4s +tttg: c192/289 lr:0.000255 t:12.5s +tttg: c193/289 lr:0.000250 t:12.6s +tttg: c194/289 lr:0.000245 t:12.6s +tttg: c195/289 lr:0.000241 t:12.7s +tttg: c196/289 lr:0.000236 t:12.8s +tttg: c197/289 lr:0.000231 t:12.8s +tttg: c198/289 lr:0.000227 t:12.9s +tttg: c199/289 lr:0.000222 t:13.0s +tttg: c200/289 lr:0.000218 t:13.0s +tttg: c201/289 lr:0.000213 t:13.1s +tttg: c202/289 lr:0.000209 t:13.1s +tttg: c203/289 lr:0.000204 t:13.2s +tttg: c204/289 lr:0.000200 t:13.3s +tttg: c205/289 lr:0.000196 t:13.3s +tttg: c206/289 lr:0.000191 t:13.4s +tttg: c207/289 lr:0.000187 t:13.5s +tttg: c208/289 lr:0.000183 t:13.5s +tttg: c209/289 lr:0.000179 t:13.6s +tttg: c210/289 lr:0.000174 t:13.7s +tttg: c211/289 lr:0.000170 t:13.7s +tttg: c212/289 lr:0.000166 t:13.8s +tttg: c213/289 lr:0.000162 t:13.9s +tttg: c214/289 lr:0.000158 t:13.9s +tttg: c215/289 lr:0.000154 t:14.0s +tttg: c216/289 lr:0.000150 t:14.1s +tttg: c217/289 lr:0.000146 t:14.1s +tttg: c218/289 lr:0.000143 t:14.2s +tttg: c219/289 lr:0.000139 t:14.3s +tttg: c220/289 lr:0.000135 t:14.3s +tttg: c221/289 lr:0.000131 t:14.4s +tttg: c222/289 lr:0.000128 t:14.4s +tttg: c223/289 lr:0.000124 t:14.5s +tttg: c224/289 lr:0.000121 t:14.6s +tttg: c225/289 lr:0.000117 t:14.6s +tttg: c226/289 lr:0.000113 t:14.7s +tttg: c227/289 lr:0.000110 t:14.8s +tttg: c228/289 lr:0.000107 t:14.8s +tttg: c229/289 lr:0.000103 t:14.9s +tttg: c230/289 lr:0.000100 t:15.0s +tttg: c231/289 lr:0.000097 t:15.0s +tttg: c232/289 lr:0.000094 t:15.1s +tttg: c233/289 lr:0.000090 t:15.2s +tttg: c234/289 lr:0.000087 t:15.2s +tttg: c235/289 lr:0.000084 t:15.3s +tttg: c236/289 lr:0.000081 t:15.4s +tttg: c237/289 lr:0.000078 t:15.4s +tttg: c238/289 lr:0.000075 t:15.5s +tttg: c239/289 lr:0.000073 t:15.6s +tttg: c240/289 lr:0.000070 t:15.6s +tttg: c241/289 lr:0.000067 t:15.7s +tttg: c242/289 lr:0.000064 t:15.7s +tttg: c243/289 lr:0.000062 t:15.8s +tttg: c244/289 lr:0.000059 t:15.9s +tttg: c245/289 lr:0.000056 t:15.9s +tttg: c246/289 lr:0.000054 t:16.0s +tttg: c247/289 lr:0.000052 t:16.1s +tttg: c248/289 lr:0.000049 t:16.1s +tttg: c249/289 lr:0.000047 t:16.2s +tttg: c250/289 lr:0.000045 t:16.3s +tttg: c251/289 lr:0.000042 t:16.3s +tttg: c252/289 lr:0.000040 t:16.4s +tttg: c253/289 lr:0.000038 t:16.5s +tttg: c254/289 lr:0.000036 t:16.5s +tttg: c255/289 lr:0.000034 t:16.6s +tttg: c256/289 lr:0.000032 t:16.7s +tttg: c257/289 lr:0.000030 t:16.7s +tttg: c258/289 lr:0.000028 t:16.8s +tttg: c259/289 lr:0.000027 t:16.8s +tttg: c260/289 lr:0.000025 t:16.9s +tttg: c261/289 lr:0.000023 t:17.0s +tttg: c262/289 lr:0.000022 t:17.0s +tttg: c263/289 lr:0.000020 t:17.1s +tttg: c264/289 lr:0.000018 t:17.2s +tttg: c265/289 lr:0.000017 t:17.2s +tttg: c266/289 lr:0.000016 t:17.3s +tttg: c267/289 lr:0.000014 t:17.4s +tttg: c268/289 lr:0.000013 t:17.4s +tttg: c269/289 lr:0.000012 t:17.5s +tttg: c270/289 lr:0.000011 t:17.6s +tttg: c271/289 lr:0.000010 t:17.6s +tttg: c272/289 lr:0.000009 t:17.7s +tttg: c273/289 lr:0.000008 t:17.8s +tttg: c274/289 lr:0.000007 t:17.8s +tttg: c275/289 lr:0.000006 t:17.9s +tttg: c276/289 lr:0.000005 t:18.0s +tttg: c277/289 lr:0.000004 t:18.0s +tttg: c278/289 lr:0.000004 t:18.1s +tttg: c279/289 lr:0.000003 t:18.1s +tttg: c280/289 lr:0.000002 t:18.2s +tttg: c281/289 lr:0.000002 t:18.3s +tttg: c282/289 lr:0.000001 t:18.3s +tttg: c283/289 lr:0.000001 t:18.4s +tttg: c284/289 lr:0.000001 t:18.5s +tttg: c285/289 lr:0.000000 t:18.5s +tttg: c286/289 lr:0.000000 t:18.6s +tttg: c287/289 lr:0.000000 t:18.7s +tttg: c288/289 lr:0.000000 t:18.7s +ttpr: phase:3/3 t:296.6s +ttp: b728/782 bl:2.3606 bb:1.0807 rl:2.3115 rb:1.0745 dl:2306-2324 gd:1 +ttp: b726/782 bl:2.3326 bb:1.0375 rl:2.3127 rb:1.0723 dl:2254-2276 gd:1 +ttp: b717/782 bl:2.2580 bb:1.0339 rl:2.3100 rb:1.0704 dl:2070-2088 gd:1 +ttp: b706/782 bl:2.4012 bb:1.0739 rl:2.3139 rb:1.0705 dl:1898-1910 gd:1 +ttp: b703/782 bl:2.3401 bb:1.0294 rl:2.3150 rb:1.0688 dl:1859-1872 gd:1 +ttp: b692/782 bl:2.2899 bb:1.0280 rl:2.3141 rb:1.0673 dl:1737-1746 gd:1 +ttp: b682/782 bl:2.3397 bb:1.0558 rl:2.3149 rb:1.0669 dl:1638-1646 gd:1 +ttp: b674/782 bl:2.4033 bb:1.0885 rl:2.3177 rb:1.0676 dl:1571-1578 gd:1 +ttp: b671/782 bl:2.3048 bb:1.0455 rl:2.3173 rb:1.0669 dl:1544-1552 gd:1 +ttp: b656/782 bl:2.3281 bb:1.1106 rl:2.3176 rb:1.0680 dl:1439-1445 gd:1 +ttp: b648/782 bl:2.2805 bb:1.0063 rl:2.3166 rb:1.0664 dl:1387-1392 gd:1 +ttp: b647/782 bl:2.2755 bb:1.0327 rl:2.3156 rb:1.0656 dl:1382-1387 gd:1 +ttp: b638/782 bl:2.3405 bb:1.0664 rl:2.3162 rb:1.0656 dl:1325-1331 gd:1 +ttp: b629/782 bl:2.3489 bb:1.0108 rl:2.3169 rb:1.0643 dl:1276-1280 gd:1 +ttp: b621/782 bl:2.2958 bb:1.0484 rl:2.3165 rb:1.0640 dl:1231-1237 gd:1 +ttp: b612/782 bl:2.2368 bb:1.0134 rl:2.3149 rb:1.0630 dl:1186-1190 gd:1 +ttp: b605/782 bl:2.2439 bb:1.0233 rl:2.3136 rb:1.0623 dl:1154-1159 gd:1 +ttp: b593/782 bl:2.2896 bb:1.0107 rl:2.3132 rb:1.0614 dl:1103-1107 gd:1 +ttp: b584/782 bl:2.2947 bb:1.0375 rl:2.3129 rb:1.0610 dl:1064-1069 gd:1 +ttp: b577/782 bl:2.2899 bb:1.0307 rl:2.3126 rb:1.0605 dl:1037-1041 gd:1 +ttp: b571/782 bl:2.2957 bb:1.0042 rl:2.3123 rb:1.0596 dl:1014-1017 gd:1 +ttp: b563/782 bl:2.2658 bb:1.0182 rl:2.3116 rb:1.0590 dl:987-990 gd:1 +ttp: b555/782 bl:2.3192 bb:1.0234 rl:2.3117 rb:1.0585 dl:959-961 gd:1 +ttp: b552/782 bl:2.2736 bb:1.0185 rl:2.3112 rb:1.0579 dl:949-952 gd:1 +ttp: b543/782 bl:2.3310 bb:1.0553 rl:2.3115 rb:1.0579 dl:921-924 gd:1 +ttp: b532/782 bl:2.3892 bb:1.0670 rl:2.3124 rb:1.0580 dl:887-889 gd:1 +ttp: b522/782 bl:2.3063 bb:1.0343 rl:2.3124 rb:1.0577 dl:858-860 gd:1 +ttp: b517/782 bl:2.3466 bb:1.0240 rl:2.3128 rb:1.0573 dl:843-846 gd:1 +ttp: b509/782 bl:2.3592 bb:1.0358 rl:2.3133 rb:1.0571 dl:820-823 gd:1 +ttp: b501/782 bl:2.3798 bb:1.0513 rl:2.3140 rb:1.0570 dl:799-802 gd:1 +ttp: b482/782 bl:2.3266 bb:1.0460 rl:2.3141 rb:1.0569 dl:752-754 gd:1 +ttp: b474/782 bl:2.3411 bb:1.0719 rl:2.3144 rb:1.0570 dl:733-735 gd:1 +ttp: b466/782 bl:2.3837 bb:1.0277 rl:2.3150 rb:1.0567 dl:714-717 gd:1 +ttp: b460/782 bl:2.2468 bb:1.0511 rl:2.3144 rb:1.0567 dl:701-703 gd:1 +ttp: b451/782 bl:2.4033 bb:1.0875 rl:2.3152 rb:1.0570 dl:682-685 gd:1 +ttp: b444/782 bl:2.3092 bb:1.0639 rl:2.3151 rb:1.0570 dl:668-670 gd:1 +ttp: b437/782 bl:2.2901 bb:1.0537 rl:2.3149 rb:1.0570 dl:653-655 gd:1 +ttp: b429/782 bl:2.2417 bb:1.0224 rl:2.3143 rb:1.0567 dl:638-640 gd:1 +ttp: b420/782 bl:2.3591 bb:1.0531 rl:2.3147 rb:1.0567 dl:620-622 gd:1 +ttp: b414/782 bl:2.2000 bb:1.0072 rl:2.3138 rb:1.0563 dl:609-611 gd:1 +ttp: b406/782 bl:2.3147 bb:1.0659 rl:2.3138 rb:1.0564 dl:593-595 gd:1 +ttp: b398/782 bl:2.2466 bb:1.0032 rl:2.3133 rb:1.0560 dl:579-581 gd:1 +ttp: b389/782 bl:2.2845 bb:1.0820 rl:2.3132 rb:1.0562 dl:563-564 gd:1 +ttp: b382/782 bl:2.2943 bb:1.0840 rl:2.3130 rb:1.0564 dl:550-552 gd:1 +ttp: b375/782 bl:2.4036 bb:1.0720 rl:2.3136 rb:1.0565 dl:538-540 gd:1 +ttp: b367/782 bl:2.2905 bb:1.0809 rl:2.3135 rb:1.0566 dl:525-527 gd:1 +ttp: b360/782 bl:2.3028 bb:1.0773 rl:2.3134 rb:1.0567 dl:513-515 gd:1 +ttp: b352/782 bl:2.4202 bb:1.0952 rl:2.3140 rb:1.0570 dl:499-501 gd:1 +ttp: b344/782 bl:2.3865 bb:1.0635 rl:2.3144 rb:1.0570 dl:488-489 gd:1 +ttp: b336/782 bl:2.4077 bb:1.0850 rl:2.3149 rb:1.0571 dl:476-477 gd:1 +ttp: b329/782 bl:2.2821 bb:1.0813 rl:2.3148 rb:1.0573 dl:465-466 gd:1 +ttp: b321/782 bl:2.3504 bb:1.0730 rl:2.3149 rb:1.0573 dl:453-455 gd:1 +ttp: b312/782 bl:2.3117 bb:1.0530 rl:2.3149 rb:1.0573 dl:439-440 gd:1 +ttp: b304/782 bl:2.3322 bb:1.0697 rl:2.3150 rb:1.0574 dl:427-429 gd:1 +ttp: b297/782 bl:2.4001 bb:1.0845 rl:2.3154 rb:1.0575 dl:417-418 gd:1 +ttp: b289/782 bl:2.3218 bb:1.0798 rl:2.3154 rb:1.0576 dl:405-406 gd:1 +ttp: b281/782 bl:2.2966 bb:1.0887 rl:2.3154 rb:1.0577 dl:394-395 gd:1 +ttp: b273/782 bl:2.3353 bb:1.0765 rl:2.3154 rb:1.0578 dl:383-384 gd:1 +ttp: b267/782 bl:2.4144 bb:1.1411 rl:2.3158 rb:1.0582 dl:375-376 gd:1 +ttp: b257/782 bl:2.4429 bb:1.1112 rl:2.3163 rb:1.0584 dl:362-364 gd:1 +ttp: b248/782 bl:2.4542 bb:1.1845 rl:2.3169 rb:1.0588 dl:351-352 gd:1 +ttp: b240/782 bl:2.3083 bb:1.0595 rl:2.3168 rb:1.0588 dl:341-342 gd:1 +ttp: b231/782 bl:2.3041 bb:1.0824 rl:2.3168 rb:1.0589 dl:330-331 gd:1 +ttp: b223/782 bl:2.3302 bb:1.1250 rl:2.3168 rb:1.0591 dl:321-322 gd:1 +ttp: b215/782 bl:2.3893 bb:1.0952 rl:2.3171 rb:1.0592 dl:312-313 gd:1 +ttp: b209/782 bl:2.4126 bb:1.1285 rl:2.3174 rb:1.0595 dl:305-306 gd:1 +ttp: b201/782 bl:2.2992 bb:1.0968 rl:2.3173 rb:1.0596 dl:297-298 gd:1 +ttp: b193/782 bl:2.3502 bb:1.1270 rl:2.3174 rb:1.0598 dl:288-289 gd:1 +ttp: b187/782 bl:2.4523 bb:1.1332 rl:2.3178 rb:1.0600 dl:281-282 gd:1 +ttp: b179/782 bl:2.3667 bb:1.1283 rl:2.3180 rb:1.0602 dl:273-274 gd:1 +ttp: b169/782 bl:2.3768 bb:1.1171 rl:2.3181 rb:1.0603 dl:263-264 gd:1 +ttp: b161/782 bl:2.3561 bb:1.1341 rl:2.3183 rb:1.0605 dl:256-256 gd:1 +ttp: b153/782 bl:2.2630 bb:1.0468 rl:2.3181 rb:1.0605 dl:248-249 gd:1 +ttp: b145/782 bl:2.5282 bb:1.1686 rl:2.3186 rb:1.0608 dl:240-241 gd:1 +ttp: b139/782 bl:2.4284 bb:1.1313 rl:2.3189 rb:1.0609 dl:234-235 gd:1 +ttp: b131/782 bl:2.3913 bb:1.1546 rl:2.3191 rb:1.0611 dl:227-228 gd:1 +ttp: b121/782 bl:2.4273 bb:1.1078 rl:2.3193 rb:1.0612 dl:218-219 gd:1 +ttp: b113/782 bl:2.5532 bb:1.1352 rl:2.3198 rb:1.0614 dl:210-211 gd:1 +ttp: b104/782 bl:2.4982 bb:1.1794 rl:2.3202 rb:1.0616 dl:202-203 gd:1 +ttp: b94/782 bl:2.5565 bb:1.2080 rl:2.3207 rb:1.0619 dl:193-194 gd:1 +ttp: b86/782 bl:2.4547 bb:1.1325 rl:2.3209 rb:1.0621 dl:186-187 gd:1 +ttp: b78/782 bl:2.5403 bb:1.1893 rl:2.3213 rb:1.0623 dl:179-180 gd:1 +ttp: b71/782 bl:2.4561 bb:1.1738 rl:2.3216 rb:1.0625 dl:173-173 gd:1 +ttp: b63/782 bl:2.5243 bb:1.2041 rl:2.3219 rb:1.0627 dl:166-166 gd:1 +ttp: b54/782 bl:2.4737 bb:1.2134 rl:2.3222 rb:1.0629 dl:157-158 gd:1 +ttp: b46/782 bl:2.5455 bb:1.2154 rl:2.3225 rb:1.0632 dl:149-150 gd:1 +ttp: b38/782 bl:2.6083 bb:1.1963 rl:2.3229 rb:1.0633 dl:141-142 gd:1 +ttp: b30/782 bl:2.5904 bb:1.2630 rl:2.3233 rb:1.0636 dl:133-134 gd:1 +ttp: b22/782 bl:2.5695 bb:1.2028 rl:2.3236 rb:1.0638 dl:124-126 gd:1 +ttp: b14/782 bl:2.5889 bb:1.1818 rl:2.3239 rb:1.0639 dl:114-115 gd:1 +ttp: b6/782 bl:2.7171 bb:1.2114 rl:2.3243 rb:1.0641 dl:99-101 gd:1 +quantized_ttt_phased val_loss:2.31977393 val_bpb:1.06004616 eval_time:395770ms +total_eval_time:395.8s +=== Done: SEED=314 === diff --git a/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_seed42.log b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_seed42.log new file mode 100644 index 0000000000..8415915fc6 --- /dev/null +++ b/records/track_10min_16mb/2020-04-29_AWQ_lite_mixedprecision_GPTQ/train_seed42.log @@ -0,0 +1,950 @@ +=== Starting training: SEED=42 RUN_ID=awq_seed42_20260429_050241 === +=== Artifact dir: /workspace/pargolf-work/runs/awq_seed42_20260429_050241 === +=== NPROC=8 === +W0429 05:02:42.809000 90934 torch/distributed/run.py:803] +W0429 05:02:42.809000 90934 torch/distributed/run.py:803] ***************************************** +W0429 05:02:42.809000 90934 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0429 05:02:42.809000 90934 torch/distributed/run.py:803] ***************************************** +Hyperparameters: + adam_eps: 1e-08 + adam_wd: 0.02 + artifact_dir: /workspace/pargolf-work/runs/awq_seed42_20260429_050241 + attn_clip_sigmas: 13.0 + attn_out_gate_enabled: False + attn_out_gate_src: proj + awq_lite_bits: 8 + awq_lite_enabled: True + awq_lite_group_size: 64 + awq_lite_group_top_k: 1 + beta1: 0.9 + beta2: 0.99 + caseops_enabled: True + compressor: pergroup + data_dir: ./data/ + datasets_dir: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved + distributed: True + ema_decay: 0.9965 + embed_bits: 7 + embed_clip_sigmas: 14.0 + embed_lr: 0.6 + embed_wd: 0.085 + enable_looping_at: 0.35 + eval_seq_len: 2048 + eval_stride: 64 + fused_ce_enabled: True + gate_window: 12 + gated_attn_enabled: False + gated_attn_init_std: 0.01 + gated_attn_quant_gate: True + global_ttt_batch_seqs: 32 + global_ttt_chunk_tokens: 32768 + global_ttt_epochs: 1 + global_ttt_grad_clip: 1.0 + global_ttt_lr: 0.001 + global_ttt_momentum: 0.9 + global_ttt_respect_doc_boundaries: True + global_ttt_warmup_chunks: 0 + global_ttt_warmup_start_lr: 0.0 + gptq_calibration_batches: 16 + gptq_reserve_seconds: 0.5 + grad_accum_steps: 1 + grad_clip_norm: 0.3 + is_main_process: True + iterations: 20000 + ln_scale: True + local_rank: 0 + logfile: /workspace/pargolf-work/runs/awq_seed42_20260429_050241/awq_seed42_20260429_050241.txt + logit_softcap: 30.0 + loop_end: 5 + loop_start: 3 + lqer_asym_enabled: True + lqer_asym_group: 64 + lqer_enabled: True + lqer_factor_bits: 4 + lqer_gain_select: False + lqer_rank: 4 + lqer_scope: all + lqer_top_k: 3 + matrix_bits: 6 + matrix_clip_sigmas: 12.85 + matrix_lr: 0.026 + max_wallclock_seconds: 600.0 + min_lr: 0.1 + mlp_clip_sigmas: 11.5 + mlp_mult: 4.0 + model_dim: 512 + model_path: /workspace/pargolf-work/runs/awq_seed42_20260429_050241/final_model.pt + muon_backend_steps: 5 + muon_momentum: 0.97 + muon_momentum_warmup_start: 0.92 + muon_momentum_warmup_steps: 1500 + muon_row_normalize: True + muon_wd: 0.095 + num_heads: 8 + num_kv_heads: 4 + num_layers: 11 + num_loops: 2 + parallel_final_lane: mean + parallel_start_layer: 8 + phased_ttt_num_phases: 3 + phased_ttt_prefix_docs: 2500 + qk_gain_init: 5.0 + quantized_model_path: /workspace/pargolf-work/runs/awq_seed42_20260429_050241/final_model.int6.ptz + rank: 0 + rope_base: 10000.0 + rope_dims: 16 + rope_train_seq_len: 2048 + rope_yarn: False + run_id: awq_seed42_20260429_050241 + scalar_lr: 0.02 + seed: 42 + skip_gates_enabled: True + smear_gate_enabled: True + sparse_attn_gate_enabled: True + sparse_attn_gate_init_std: 0.0 + sparse_attn_gate_scale: 0.5 + tie_embeddings: True + tied_embed_init_std: 0.005 + tied_embed_lr: 0.03 + tokenizer_path: /workspace/pargolf/records/track_10min_16mb/2026-04-27_SP8192_LQER_SparseGate_BOSSmearFix_9HpStack_1.0611/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model + train_batch_tokens: 786432 + train_files: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/fineweb_train_*.bin + train_log_every: 500 + train_seq_len: 2048 + ttt_batch_size: 64 + ttt_beta1: 0.0 + ttt_beta2: 0.99 + ttt_chunk_size: 48 + ttt_enabled: True + ttt_eval_batches: + ttt_eval_seq_len: 2048 + ttt_grad_steps: 1 + ttt_k_lora: True + ttt_lora_lr: 0.0001 + ttt_lora_rank: 80 + ttt_mlp_lora: True + ttt_o_lora: True + ttt_optimizer: adam + ttt_weight_decay: 0.5 + val_batch_tokens: 524288 + val_bytes_files: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/fineweb_val_bytes_*.bin + val_doc_fraction: 1.0 + val_files: /workspace/pargolf-work/data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/fineweb_val_*.bin + val_loss_every: 4000 + vocab_size: 8192 + warmdown_frac: 0.85 + warmup_steps: 20 + world_size: 8 + xsa_last_n: 11 +train_shards: 80 +val_tokens: 47851520 +model_params:35945671 +gptq:reserving 0s, effective=599500ms +warmup_cu_buckets:64,128,192,256 iters_each:3 +warmup_step: 1/20 +warmup_step: 2/20 +warmup_step: 3/20 +warmup_step: 4/20 +warmup_step: 5/20 +warmup_step: 6/20 +warmup_step: 10/20 +warmup_step: 20/20 +loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +loop_warmup_step: 1/20 +loop_warmup_step: 2/20 +loop_warmup_step: 3/20 +loop_warmup_step: 4/20 +loop_warmup_step: 5/20 +loop_warmup_step: 6/20 +loop_warmup_step: 10/20 +loop_warmup_step: 20/20 +0/20000 val_loss: 9.0076 val_bpb: 4.1159 +1/20000 train_loss: 9.0109 train_time: 0.0m tok/s: 18061930 +2/20000 train_loss: 12.6608 train_time: 0.0m tok/s: 11392388 +3/20000 train_loss: 10.3312 train_time: 0.0m tok/s: 9969099 +4/20000 train_loss: 9.1770 train_time: 0.0m tok/s: 9361768 +5/20000 train_loss: 7.9874 train_time: 0.0m tok/s: 9058402 +500/20000 train_loss: 2.6266 train_time: 0.8m tok/s: 8346539 +1000/20000 train_loss: 2.7466 train_time: 1.6m tok/s: 8318344 +1500/20000 train_loss: 2.7187 train_time: 2.4m tok/s: 8309959 +2000/20000 train_loss: 2.6271 train_time: 3.2m tok/s: 8305645 +layer_loop:enabled step:2215 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +2500/20000 train_loss: 2.4717 train_time: 4.2m tok/s: 7880578 +3000/20000 train_loss: 2.4940 train_time: 5.4m tok/s: 7336864 +3500/20000 train_loss: 2.6240 train_time: 6.5m tok/s: 7036254 +4000/20000 train_loss: 2.3701 train_time: 7.7m tok/s: 6811025 +4000/20000 val_loss: 2.4309 val_bpb: 1.1108 +4500/20000 train_loss: 2.5361 train_time: 8.9m tok/s: 6659009 +4989/20000 val_loss: 2.3567 val_bpb: 1.0768 +stopping_early: wallclock_cap train_time: 599599ms step: 4989/20000 +peak memory allocated: 41707 MiB reserved: 47048 MiB +ema:applying EMA weights +diagnostic pre-quantization post-ema val_loss:2.33105143 val_bpb:1.06513037 eval_time:8185ms +Serialized model: 135417533 bytes +Code size (uncompressed): 168543 bytes +Code size (compressed): 33787 bytes +GPTQ:collecting Hessians from calibration data... +GPTQ:collected 67 Hessians in 4.0s +Quantized weights: + gate_int8_row: blocks.attn.attn_gate_w + gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight + gptq (int6)+lqer_asym: blocks.mlp.fc.weight + gptq (int7)+awqgrpint8+lqer_asym: tok_emb.weight + passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, parallel_post_lambdas, parallel_resid_lambdas, skip_gates, skip_weights, smear_gate.weight, smear_lambda +Serialize: per-group lrzip compression... +Serialize: per-group compression done in 110.0s +Serialized model quantized+pergroup: 15944716 bytes +Total submission size quantized+pergroup: 15978503 bytes +Deserialize: per-group lrzip decompression... +Deserialize: decompression done in 19.8s +diagnostic quantized val_loss:2.34961250 val_bpb:1.07361151 eval_time:12561ms +Deserialize: per-group lrzip decompression... +Deserialize: decompression done in 19.4s +ttt_lora:warming up compile (random tokens, no val data) +ttt_lora:compile warmup done (91.7s) + +beginning TTT eval timer +ttt_phased: total_docs:50000 prefix_docs:2500 suffix_docs:47500 num_phases:3 boundaries:[833, 1666, 2500] +ttp: b776/782 bl:2.2579 bb:1.0705 rl:2.2579 rb:1.0705 dl:7534-8350 gd:0 +ttp: b773/782 bl:2.1991 bb:1.0356 rl:2.2318 rb:1.0550 dl:6104-6447 gd:0 +ttp: b769/782 bl:2.3288 bb:1.0841 rl:2.2578 rb:1.0629 dl:5097-5309 gd:0 +ttp: b763/782 bl:2.4204 bb:1.0998 rl:2.2868 rb:1.0697 dl:4142-4283 gd:0 +ttpp: phase:1/3 pd:1296 gd:833 t:168.3s +tttg: c1/131 lr:0.001000 t:0.4s +tttg: c2/131 lr:0.001000 t:0.5s +tttg: c3/131 lr:0.000999 t:0.6s +tttg: c4/131 lr:0.000999 t:0.6s +tttg: c5/131 lr:0.000998 t:0.7s +tttg: c6/131 lr:0.000996 t:0.8s +tttg: c7/131 lr:0.000995 t:0.9s +tttg: c8/131 lr:0.000993 t:0.9s +tttg: c9/131 lr:0.000991 t:1.0s +tttg: c10/131 lr:0.000988 t:1.1s +tttg: c11/131 lr:0.000985 t:1.2s +tttg: c12/131 lr:0.000982 t:1.2s +tttg: c13/131 lr:0.000979 t:1.3s +tttg: c14/131 lr:0.000976 t:1.4s +tttg: c15/131 lr:0.000972 t:1.4s +tttg: c16/131 lr:0.000968 t:1.5s +tttg: c17/131 lr:0.000963 t:1.6s +tttg: c18/131 lr:0.000958 t:1.6s +tttg: c19/131 lr:0.000953 t:1.7s +tttg: c20/131 lr:0.000948 t:1.8s +tttg: c21/131 lr:0.000943 t:1.8s +tttg: c22/131 lr:0.000937 t:1.9s +tttg: c23/131 lr:0.000931 t:2.0s +tttg: c24/131 lr:0.000925 t:2.0s +tttg: c25/131 lr:0.000918 t:2.1s +tttg: c26/131 lr:0.000911 t:2.1s +tttg: c27/131 lr:0.000905 t:2.2s +tttg: c28/131 lr:0.000897 t:2.3s +tttg: c29/131 lr:0.000890 t:2.3s +tttg: c30/131 lr:0.000882 t:2.4s +tttg: c31/131 lr:0.000874 t:2.5s +tttg: c32/131 lr:0.000866 t:2.5s +tttg: c33/131 lr:0.000858 t:2.6s +tttg: c34/131 lr:0.000849 t:2.7s +tttg: c35/131 lr:0.000841 t:2.7s +tttg: c36/131 lr:0.000832 t:2.8s +tttg: c37/131 lr:0.000822 t:2.9s +tttg: c38/131 lr:0.000813 t:2.9s +tttg: c39/131 lr:0.000804 t:3.0s +tttg: c40/131 lr:0.000794 t:3.0s +tttg: c41/131 lr:0.000784 t:3.1s +tttg: c42/131 lr:0.000774 t:3.2s +tttg: c43/131 lr:0.000764 t:3.2s +tttg: c44/131 lr:0.000753 t:3.3s +tttg: c45/131 lr:0.000743 t:3.4s +tttg: c46/131 lr:0.000732 t:3.4s +tttg: c47/131 lr:0.000722 t:3.5s +tttg: c48/131 lr:0.000711 t:3.6s +tttg: c49/131 lr:0.000700 t:3.6s +tttg: c50/131 lr:0.000689 t:3.7s +tttg: c51/131 lr:0.000677 t:3.8s +tttg: c52/131 lr:0.000666 t:3.8s +tttg: c53/131 lr:0.000655 t:3.9s +tttg: c54/131 lr:0.000643 t:3.9s +tttg: c55/131 lr:0.000631 t:4.0s +tttg: c56/131 lr:0.000620 t:4.1s +tttg: c57/131 lr:0.000608 t:4.1s +tttg: c58/131 lr:0.000596 t:4.2s +tttg: c59/131 lr:0.000584 t:4.3s +tttg: c60/131 lr:0.000572 t:4.3s +tttg: c61/131 lr:0.000560 t:4.4s +tttg: c62/131 lr:0.000548 t:4.5s +tttg: c63/131 lr:0.000536 t:4.5s +tttg: c64/131 lr:0.000524 t:4.6s +tttg: c65/131 lr:0.000512 t:4.7s +tttg: c66/131 lr:0.000500 t:4.7s +tttg: c67/131 lr:0.000488 t:4.8s +tttg: c68/131 lr:0.000476 t:4.9s +tttg: c69/131 lr:0.000464 t:4.9s +tttg: c70/131 lr:0.000452 t:5.0s +tttg: c71/131 lr:0.000440 t:5.0s +tttg: c72/131 lr:0.000428 t:5.1s +tttg: c73/131 lr:0.000416 t:5.2s +tttg: c74/131 lr:0.000404 t:5.2s +tttg: c75/131 lr:0.000392 t:5.3s +tttg: c76/131 lr:0.000380 t:5.4s +tttg: c77/131 lr:0.000369 t:5.4s +tttg: c78/131 lr:0.000357 t:5.5s +tttg: c79/131 lr:0.000345 t:5.6s +tttg: c80/131 lr:0.000334 t:5.6s +tttg: c81/131 lr:0.000323 t:5.7s +tttg: c82/131 lr:0.000311 t:5.8s +tttg: c83/131 lr:0.000300 t:5.8s +tttg: c84/131 lr:0.000289 t:5.9s +tttg: c85/131 lr:0.000278 t:6.0s +tttg: c86/131 lr:0.000268 t:6.0s +tttg: c87/131 lr:0.000257 t:6.1s +tttg: c88/131 lr:0.000247 t:6.2s +tttg: c89/131 lr:0.000236 t:6.2s +tttg: c90/131 lr:0.000226 t:6.3s +tttg: c91/131 lr:0.000216 t:6.3s +tttg: c92/131 lr:0.000206 t:6.4s +tttg: c93/131 lr:0.000196 t:6.5s +tttg: c94/131 lr:0.000187 t:6.5s +tttg: c95/131 lr:0.000178 t:6.6s +tttg: c96/131 lr:0.000168 t:6.7s +tttg: c97/131 lr:0.000159 t:6.7s +tttg: c98/131 lr:0.000151 t:6.8s +tttg: c99/131 lr:0.000142 t:6.9s +tttg: c100/131 lr:0.000134 t:6.9s +tttg: c101/131 lr:0.000126 t:7.0s +tttg: c102/131 lr:0.000118 t:7.1s +tttg: c103/131 lr:0.000110 t:7.1s +tttg: c104/131 lr:0.000103 t:7.2s +tttg: c105/131 lr:0.000095 t:7.3s +tttg: c106/131 lr:0.000089 t:7.3s +tttg: c107/131 lr:0.000082 t:7.4s +tttg: c108/131 lr:0.000075 t:7.4s +tttg: c109/131 lr:0.000069 t:7.5s +tttg: c110/131 lr:0.000063 t:7.6s +tttg: c111/131 lr:0.000057 t:7.6s +tttg: c112/131 lr:0.000052 t:7.7s +tttg: c113/131 lr:0.000047 t:7.8s +tttg: c114/131 lr:0.000042 t:7.8s +tttg: c115/131 lr:0.000037 t:7.9s +tttg: c116/131 lr:0.000032 t:8.0s +tttg: c117/131 lr:0.000028 t:8.0s +tttg: c118/131 lr:0.000024 t:8.1s +tttg: c119/131 lr:0.000021 t:8.2s +tttg: c120/131 lr:0.000018 t:8.2s +tttg: c121/131 lr:0.000015 t:8.3s +tttg: c122/131 lr:0.000012 t:8.3s +tttg: c123/131 lr:0.000009 t:8.4s +tttg: c124/131 lr:0.000007 t:8.5s +tttg: c125/131 lr:0.000005 t:8.5s +tttg: c126/131 lr:0.000004 t:8.6s +tttg: c127/131 lr:0.000002 t:8.7s +tttg: c128/131 lr:0.000001 t:8.7s +tttg: c129/131 lr:0.000001 t:8.8s +tttg: c130/131 lr:0.000000 t:8.9s +ttpr: phase:1/3 t:178.7s +ttp: b754/782 bl:2.2878 bb:1.0582 rl:2.2869 rb:1.0682 dl:3345-3397 gd:0 +ttp: b750/782 bl:2.3855 bb:1.0719 rl:2.2972 rb:1.0686 dl:3090-3149 gd:0 +ttpp: phase:2/3 pd:2128 gd:1666 t:244.4s +tttg: c1/219 lr:0.001000 t:0.1s +tttg: c2/219 lr:0.001000 t:0.1s +tttg: c3/219 lr:0.001000 t:0.2s +tttg: c4/219 lr:0.001000 t:0.3s +tttg: c5/219 lr:0.000999 t:0.3s +tttg: c6/219 lr:0.000999 t:0.4s +tttg: c7/219 lr:0.000998 t:0.5s +tttg: c8/219 lr:0.000997 t:0.5s +tttg: c9/219 lr:0.000997 t:0.6s +tttg: c10/219 lr:0.000996 t:0.7s +tttg: c11/219 lr:0.000995 t:0.7s +tttg: c12/219 lr:0.000994 t:0.8s +tttg: c13/219 lr:0.000993 t:0.9s +tttg: c14/219 lr:0.000991 t:0.9s +tttg: c15/219 lr:0.000990 t:1.0s +tttg: c16/219 lr:0.000988 t:1.0s +tttg: c17/219 lr:0.000987 t:1.1s +tttg: c18/219 lr:0.000985 t:1.2s +tttg: c19/219 lr:0.000983 t:1.3s +tttg: c20/219 lr:0.000981 t:1.3s +tttg: c21/219 lr:0.000979 t:1.4s +tttg: c22/219 lr:0.000977 t:1.4s +tttg: c23/219 lr:0.000975 t:1.5s +tttg: c24/219 lr:0.000973 t:1.6s +tttg: c25/219 lr:0.000970 t:1.6s +tttg: c26/219 lr:0.000968 t:1.7s +tttg: c27/219 lr:0.000965 t:1.8s +tttg: c28/219 lr:0.000963 t:1.8s +tttg: c29/219 lr:0.000960 t:1.9s +tttg: c30/219 lr:0.000957 t:2.0s +tttg: c31/219 lr:0.000954 t:2.0s +tttg: c32/219 lr:0.000951 t:2.1s +tttg: c33/219 lr:0.000948 t:2.2s +tttg: c34/219 lr:0.000945 t:2.2s +tttg: c35/219 lr:0.000941 t:2.3s +tttg: c36/219 lr:0.000938 t:2.4s +tttg: c37/219 lr:0.000934 t:2.4s +tttg: c38/219 lr:0.000931 t:2.5s +tttg: c39/219 lr:0.000927 t:2.6s +tttg: c40/219 lr:0.000923 t:2.6s +tttg: c41/219 lr:0.000919 t:2.7s +tttg: c42/219 lr:0.000915 t:2.8s +tttg: c43/219 lr:0.000911 t:2.8s +tttg: c44/219 lr:0.000907 t:2.9s +tttg: c45/219 lr:0.000903 t:3.0s +tttg: c46/219 lr:0.000898 t:3.0s +tttg: c47/219 lr:0.000894 t:3.1s +tttg: c48/219 lr:0.000890 t:3.1s +tttg: c49/219 lr:0.000885 t:3.2s +tttg: c50/219 lr:0.000880 t:3.3s +tttg: c51/219 lr:0.000876 t:3.4s +tttg: c52/219 lr:0.000871 t:3.4s +tttg: c53/219 lr:0.000866 t:3.5s +tttg: c54/219 lr:0.000861 t:3.6s +tttg: c55/219 lr:0.000856 t:3.6s +tttg: c56/219 lr:0.000851 t:3.7s +tttg: c57/219 lr:0.000846 t:3.8s +tttg: c58/219 lr:0.000841 t:3.8s +tttg: c59/219 lr:0.000835 t:3.9s +tttg: c60/219 lr:0.000830 t:4.0s +tttg: c61/219 lr:0.000824 t:4.0s +tttg: c62/219 lr:0.000819 t:4.1s +tttg: c63/219 lr:0.000813 t:4.1s +tttg: c64/219 lr:0.000808 t:4.2s +tttg: c65/219 lr:0.000802 t:4.3s +tttg: c66/219 lr:0.000796 t:4.3s +tttg: c67/219 lr:0.000790 t:4.4s +tttg: c68/219 lr:0.000784 t:4.5s +tttg: c69/219 lr:0.000779 t:4.5s +tttg: c70/219 lr:0.000773 t:4.6s +tttg: c71/219 lr:0.000766 t:4.7s +tttg: c72/219 lr:0.000760 t:4.7s +tttg: c73/219 lr:0.000754 t:4.8s +tttg: c74/219 lr:0.000748 t:4.9s +tttg: c75/219 lr:0.000742 t:4.9s +tttg: c76/219 lr:0.000735 t:5.0s +tttg: c77/219 lr:0.000729 t:5.1s +tttg: c78/219 lr:0.000722 t:5.1s +tttg: c79/219 lr:0.000716 t:5.2s +tttg: c80/219 lr:0.000709 t:5.2s +tttg: c81/219 lr:0.000703 t:5.3s +tttg: c82/219 lr:0.000696 t:5.4s +tttg: c83/219 lr:0.000690 t:5.4s +tttg: c84/219 lr:0.000683 t:5.5s +tttg: c85/219 lr:0.000676 t:5.6s +tttg: c86/219 lr:0.000670 t:5.6s +tttg: c87/219 lr:0.000663 t:5.7s +tttg: c88/219 lr:0.000656 t:5.8s +tttg: c89/219 lr:0.000649 t:5.8s +tttg: c90/219 lr:0.000642 t:5.9s +tttg: c91/219 lr:0.000635 t:6.0s +tttg: c92/219 lr:0.000628 t:6.0s +tttg: c93/219 lr:0.000621 t:6.1s +tttg: c94/219 lr:0.000614 t:6.2s +tttg: c95/219 lr:0.000607 t:6.2s +tttg: c96/219 lr:0.000600 t:6.3s +tttg: c97/219 lr:0.000593 t:6.3s +tttg: c98/219 lr:0.000586 t:6.4s +tttg: c99/219 lr:0.000579 t:6.5s +tttg: c100/219 lr:0.000572 t:6.5s +tttg: c101/219 lr:0.000565 t:6.6s +tttg: c102/219 lr:0.000558 t:6.7s +tttg: c103/219 lr:0.000550 t:6.7s +tttg: c104/219 lr:0.000543 t:6.8s +tttg: c105/219 lr:0.000536 t:6.9s +tttg: c106/219 lr:0.000529 t:6.9s +tttg: c107/219 lr:0.000522 t:7.0s +tttg: c108/219 lr:0.000514 t:7.1s +tttg: c109/219 lr:0.000507 t:7.1s +tttg: c110/219 lr:0.000500 t:7.2s +tttg: c111/219 lr:0.000493 t:7.3s +tttg: c112/219 lr:0.000486 t:7.3s +tttg: c113/219 lr:0.000478 t:7.4s +tttg: c114/219 lr:0.000471 t:7.5s +tttg: c115/219 lr:0.000464 t:7.5s +tttg: c116/219 lr:0.000457 t:7.6s +tttg: c117/219 lr:0.000450 t:7.7s +tttg: c118/219 lr:0.000442 t:7.7s +tttg: c119/219 lr:0.000435 t:7.8s +tttg: c120/219 lr:0.000428 t:7.9s +tttg: c121/219 lr:0.000421 t:7.9s +tttg: c122/219 lr:0.000414 t:8.0s +tttg: c123/219 lr:0.000407 t:8.1s +tttg: c124/219 lr:0.000400 t:8.1s +tttg: c125/219 lr:0.000393 t:8.2s +tttg: c126/219 lr:0.000386 t:8.3s +tttg: c127/219 lr:0.000379 t:8.3s +tttg: c128/219 lr:0.000372 t:8.4s +tttg: c129/219 lr:0.000365 t:8.5s +tttg: c130/219 lr:0.000358 t:8.5s +tttg: c131/219 lr:0.000351 t:8.6s +tttg: c132/219 lr:0.000344 t:8.7s +tttg: c133/219 lr:0.000337 t:8.7s +tttg: c134/219 lr:0.000330 t:8.8s +tttg: c135/219 lr:0.000324 t:8.9s +tttg: c136/219 lr:0.000317 t:8.9s +tttg: c137/219 lr:0.000310 t:9.0s +tttg: c138/219 lr:0.000304 t:9.0s +tttg: c139/219 lr:0.000297 t:9.1s +tttg: c140/219 lr:0.000291 t:9.2s +tttg: c141/219 lr:0.000284 t:9.2s +tttg: c142/219 lr:0.000278 t:9.3s +tttg: c143/219 lr:0.000271 t:9.4s +tttg: c144/219 lr:0.000265 t:9.4s +tttg: c145/219 lr:0.000258 t:9.5s +tttg: c146/219 lr:0.000252 t:9.6s +tttg: c147/219 lr:0.000246 t:9.6s +tttg: c148/219 lr:0.000240 t:9.7s +tttg: c149/219 lr:0.000234 t:9.8s +tttg: c150/219 lr:0.000227 t:9.8s +tttg: c151/219 lr:0.000221 t:9.9s +tttg: c152/219 lr:0.000216 t:10.0s +tttg: c153/219 lr:0.000210 t:10.0s +tttg: c154/219 lr:0.000204 t:10.1s +tttg: c155/219 lr:0.000198 t:10.2s +tttg: c156/219 lr:0.000192 t:10.2s +tttg: c157/219 lr:0.000187 t:10.3s +tttg: c158/219 lr:0.000181 t:10.3s +tttg: c159/219 lr:0.000176 t:10.4s +tttg: c160/219 lr:0.000170 t:10.5s +tttg: c161/219 lr:0.000165 t:10.5s +tttg: c162/219 lr:0.000159 t:10.6s +tttg: c163/219 lr:0.000154 t:10.7s +tttg: c164/219 lr:0.000149 t:10.7s +tttg: c165/219 lr:0.000144 t:10.8s +tttg: c166/219 lr:0.000139 t:10.9s +tttg: c167/219 lr:0.000134 t:10.9s +tttg: c168/219 lr:0.000129 t:11.0s +tttg: c169/219 lr:0.000124 t:11.1s +tttg: c170/219 lr:0.000120 t:11.1s +tttg: c171/219 lr:0.000115 t:11.2s +tttg: c172/219 lr:0.000110 t:11.3s +tttg: c173/219 lr:0.000106 t:11.3s +tttg: c174/219 lr:0.000102 t:11.4s +tttg: c175/219 lr:0.000097 t:11.5s +tttg: c176/219 lr:0.000093 t:11.5s +tttg: c177/219 lr:0.000089 t:11.6s +tttg: c178/219 lr:0.000085 t:11.6s +tttg: c179/219 lr:0.000081 t:11.7s +tttg: c180/219 lr:0.000077 t:11.8s +tttg: c181/219 lr:0.000073 t:11.8s +tttg: c182/219 lr:0.000069 t:11.9s +tttg: c183/219 lr:0.000066 t:12.0s +tttg: c184/219 lr:0.000062 t:12.0s +tttg: c185/219 lr:0.000059 t:12.1s +tttg: c186/219 lr:0.000055 t:12.2s +tttg: c187/219 lr:0.000052 t:12.2s +tttg: c188/219 lr:0.000049 t:12.3s +tttg: c189/219 lr:0.000046 t:12.4s +tttg: c190/219 lr:0.000043 t:12.4s +tttg: c191/219 lr:0.000040 t:12.5s +tttg: c192/219 lr:0.000037 t:12.6s +tttg: c193/219 lr:0.000035 t:12.6s +tttg: c194/219 lr:0.000032 t:12.7s +tttg: c195/219 lr:0.000030 t:12.8s +tttg: c196/219 lr:0.000027 t:12.8s +tttg: c197/219 lr:0.000025 t:12.9s +tttg: c198/219 lr:0.000023 t:13.0s +tttg: c199/219 lr:0.000021 t:13.0s +tttg: c200/219 lr:0.000019 t:13.1s +tttg: c201/219 lr:0.000017 t:13.1s +tttg: c202/219 lr:0.000015 t:13.2s +tttg: c203/219 lr:0.000013 t:13.3s +tttg: c204/219 lr:0.000012 t:13.3s +tttg: c205/219 lr:0.000010 t:13.4s +tttg: c206/219 lr:0.000009 t:13.5s +tttg: c207/219 lr:0.000007 t:13.5s +tttg: c208/219 lr:0.000006 t:13.6s +tttg: c209/219 lr:0.000005 t:13.7s +tttg: c210/219 lr:0.000004 t:13.7s +tttg: c211/219 lr:0.000003 t:13.8s +tttg: c212/219 lr:0.000003 t:13.9s +tttg: c213/219 lr:0.000002 t:13.9s +tttg: c214/219 lr:0.000001 t:14.0s +tttg: c215/219 lr:0.000001 t:14.1s +tttg: c216/219 lr:0.000000 t:14.1s +tttg: c217/219 lr:0.000000 t:14.2s +tttg: c218/219 lr:0.000000 t:14.3s +ttpr: phase:2/3 t:260.2s +ttp: b741/782 bl:2.3222 bb:1.0414 rl:2.2992 rb:1.0663 dl:2686-2730 gd:0 +ttp: b740/782 bl:2.2612 bb:1.0380 rl:2.2964 rb:1.0641 dl:2653-2686 gd:0 +ttpp: phase:3/3 pd:2960 gd:2500 t:275.3s +tttg: c1/289 lr:0.001000 t:0.1s +tttg: c2/289 lr:0.001000 t:0.1s +tttg: c3/289 lr:0.001000 t:0.2s +tttg: c4/289 lr:0.001000 t:0.3s +tttg: c5/289 lr:0.001000 t:0.3s +tttg: c6/289 lr:0.000999 t:0.4s +tttg: c7/289 lr:0.000999 t:0.5s +tttg: c8/289 lr:0.000999 t:0.5s +tttg: c9/289 lr:0.000998 t:0.6s +tttg: c10/289 lr:0.000998 t:0.7s +tttg: c11/289 lr:0.000997 t:0.7s +tttg: c12/289 lr:0.000996 t:0.8s +tttg: c13/289 lr:0.000996 t:0.9s +tttg: c14/289 lr:0.000995 t:0.9s +tttg: c15/289 lr:0.000994 t:1.0s +tttg: c16/289 lr:0.000993 t:1.0s +tttg: c17/289 lr:0.000992 t:1.1s +tttg: c18/289 lr:0.000991 t:1.2s +tttg: c19/289 lr:0.000990 t:1.2s +tttg: c20/289 lr:0.000989 t:1.3s +tttg: c21/289 lr:0.000988 t:1.4s +tttg: c22/289 lr:0.000987 t:1.4s +tttg: c23/289 lr:0.000986 t:1.5s +tttg: c24/289 lr:0.000984 t:1.6s +tttg: c25/289 lr:0.000983 t:1.6s +tttg: c26/289 lr:0.000982 t:1.7s +tttg: c27/289 lr:0.000980 t:1.8s +tttg: c28/289 lr:0.000978 t:1.8s +tttg: c29/289 lr:0.000977 t:1.9s +tttg: c30/289 lr:0.000975 t:2.0s +tttg: c31/289 lr:0.000973 t:2.0s +tttg: c32/289 lr:0.000972 t:2.1s +tttg: c33/289 lr:0.000970 t:2.2s +tttg: c34/289 lr:0.000968 t:2.2s +tttg: c35/289 lr:0.000966 t:2.3s +tttg: c36/289 lr:0.000964 t:2.4s +tttg: c37/289 lr:0.000962 t:2.4s +tttg: c38/289 lr:0.000960 t:2.5s +tttg: c39/289 lr:0.000958 t:2.5s +tttg: c40/289 lr:0.000955 t:2.6s +tttg: c41/289 lr:0.000953 t:2.7s +tttg: c42/289 lr:0.000951 t:2.7s +tttg: c43/289 lr:0.000948 t:2.8s +tttg: c44/289 lr:0.000946 t:2.9s +tttg: c45/289 lr:0.000944 t:2.9s +tttg: c46/289 lr:0.000941 t:3.0s +tttg: c47/289 lr:0.000938 t:3.1s +tttg: c48/289 lr:0.000936 t:3.1s +tttg: c49/289 lr:0.000933 t:3.2s +tttg: c50/289 lr:0.000930 t:3.3s +tttg: c51/289 lr:0.000927 t:3.3s +tttg: c52/289 lr:0.000925 t:3.4s +tttg: c53/289 lr:0.000922 t:3.5s +tttg: c54/289 lr:0.000919 t:3.5s +tttg: c55/289 lr:0.000916 t:3.6s +tttg: c56/289 lr:0.000913 t:3.7s +tttg: c57/289 lr:0.000910 t:3.7s +tttg: c58/289 lr:0.000906 t:3.8s +tttg: c59/289 lr:0.000903 t:3.9s +tttg: c60/289 lr:0.000900 t:3.9s +tttg: c61/289 lr:0.000897 t:4.0s +tttg: c62/289 lr:0.000893 t:4.1s +tttg: c63/289 lr:0.000890 t:4.1s +tttg: c64/289 lr:0.000887 t:4.2s +tttg: c65/289 lr:0.000883 t:4.3s +tttg: c66/289 lr:0.000879 t:4.3s +tttg: c67/289 lr:0.000876 t:4.4s +tttg: c68/289 lr:0.000872 t:4.4s +tttg: c69/289 lr:0.000869 t:4.5s +tttg: c70/289 lr:0.000865 t:4.6s +tttg: c71/289 lr:0.000861 t:4.6s +tttg: c72/289 lr:0.000857 t:4.7s +tttg: c73/289 lr:0.000854 t:4.8s +tttg: c74/289 lr:0.000850 t:4.8s +tttg: c75/289 lr:0.000846 t:4.9s +tttg: c76/289 lr:0.000842 t:5.0s +tttg: c77/289 lr:0.000838 t:5.0s +tttg: c78/289 lr:0.000834 t:5.1s +tttg: c79/289 lr:0.000830 t:5.2s +tttg: c80/289 lr:0.000826 t:5.2s +tttg: c81/289 lr:0.000821 t:5.3s +tttg: c82/289 lr:0.000817 t:5.4s +tttg: c83/289 lr:0.000813 t:5.4s +tttg: c84/289 lr:0.000809 t:5.5s +tttg: c85/289 lr:0.000804 t:5.6s +tttg: c86/289 lr:0.000800 t:5.6s +tttg: c87/289 lr:0.000796 t:5.7s +tttg: c88/289 lr:0.000791 t:5.7s +tttg: c89/289 lr:0.000787 t:5.8s +tttg: c90/289 lr:0.000782 t:5.9s +tttg: c91/289 lr:0.000778 t:5.9s +tttg: c92/289 lr:0.000773 t:6.0s +tttg: c93/289 lr:0.000769 t:6.1s +tttg: c94/289 lr:0.000764 t:6.1s +tttg: c95/289 lr:0.000759 t:6.2s +tttg: c96/289 lr:0.000755 t:6.3s +tttg: c97/289 lr:0.000750 t:6.3s +tttg: c98/289 lr:0.000745 t:6.4s +tttg: c99/289 lr:0.000740 t:6.5s +tttg: c100/289 lr:0.000736 t:6.5s +tttg: c101/289 lr:0.000731 t:6.6s +tttg: c102/289 lr:0.000726 t:6.7s +tttg: c103/289 lr:0.000721 t:6.7s +tttg: c104/289 lr:0.000716 t:6.8s +tttg: c105/289 lr:0.000711 t:6.9s +tttg: c106/289 lr:0.000706 t:6.9s +tttg: c107/289 lr:0.000701 t:7.0s +tttg: c108/289 lr:0.000696 t:7.0s +tttg: c109/289 lr:0.000691 t:7.1s +tttg: c110/289 lr:0.000686 t:7.2s +tttg: c111/289 lr:0.000681 t:7.3s +tttg: c112/289 lr:0.000676 t:7.3s +tttg: c113/289 lr:0.000671 t:7.4s +tttg: c114/289 lr:0.000666 t:7.4s +tttg: c115/289 lr:0.000661 t:7.5s +tttg: c116/289 lr:0.000656 t:7.6s +tttg: c117/289 lr:0.000650 t:7.6s +tttg: c118/289 lr:0.000645 t:7.7s +tttg: c119/289 lr:0.000640 t:7.8s +tttg: c120/289 lr:0.000635 t:7.8s +tttg: c121/289 lr:0.000629 t:7.9s +tttg: c122/289 lr:0.000624 t:8.0s +tttg: c123/289 lr:0.000619 t:8.0s +tttg: c124/289 lr:0.000614 t:8.1s +tttg: c125/289 lr:0.000608 t:8.2s +tttg: c126/289 lr:0.000603 t:8.2s +tttg: c127/289 lr:0.000598 t:8.3s +tttg: c128/289 lr:0.000592 t:8.4s +tttg: c129/289 lr:0.000587 t:8.4s +tttg: c130/289 lr:0.000581 t:8.5s +tttg: c131/289 lr:0.000576 t:8.6s +tttg: c132/289 lr:0.000571 t:8.6s +tttg: c133/289 lr:0.000565 t:8.7s +tttg: c134/289 lr:0.000560 t:8.7s +tttg: c135/289 lr:0.000554 t:8.8s +tttg: c136/289 lr:0.000549 t:8.9s +tttg: c137/289 lr:0.000544 t:8.9s +tttg: c138/289 lr:0.000538 t:9.0s +tttg: c139/289 lr:0.000533 t:9.1s +tttg: c140/289 lr:0.000527 t:9.1s +tttg: c141/289 lr:0.000522 t:9.2s +tttg: c142/289 lr:0.000516 t:9.3s +tttg: c143/289 lr:0.000511 t:9.3s +tttg: c144/289 lr:0.000505 t:9.4s +tttg: c145/289 lr:0.000500 t:9.5s +tttg: c146/289 lr:0.000495 t:9.5s +tttg: c147/289 lr:0.000489 t:9.6s +tttg: c148/289 lr:0.000484 t:9.7s +tttg: c149/289 lr:0.000478 t:9.7s +tttg: c150/289 lr:0.000473 t:9.8s +tttg: c151/289 lr:0.000467 t:9.8s +tttg: c152/289 lr:0.000462 t:9.9s +tttg: c153/289 lr:0.000456 t:10.0s +tttg: c154/289 lr:0.000451 t:10.0s +tttg: c155/289 lr:0.000446 t:10.1s +tttg: c156/289 lr:0.000440 t:10.2s +tttg: c157/289 lr:0.000435 t:10.2s +tttg: c158/289 lr:0.000429 t:10.3s +tttg: c159/289 lr:0.000424 t:10.4s +tttg: c160/289 lr:0.000419 t:10.4s +tttg: c161/289 lr:0.000413 t:10.5s +tttg: c162/289 lr:0.000408 t:10.6s +tttg: c163/289 lr:0.000402 t:10.6s +tttg: c164/289 lr:0.000397 t:10.7s +tttg: c165/289 lr:0.000392 t:10.8s +tttg: c166/289 lr:0.000386 t:10.8s +tttg: c167/289 lr:0.000381 t:10.9s +tttg: c168/289 lr:0.000376 t:11.0s +tttg: c169/289 lr:0.000371 t:11.0s +tttg: c170/289 lr:0.000365 t:11.1s +tttg: c171/289 lr:0.000360 t:11.2s +tttg: c172/289 lr:0.000355 t:11.2s +tttg: c173/289 lr:0.000350 t:11.3s +tttg: c174/289 lr:0.000344 t:11.3s +tttg: c175/289 lr:0.000339 t:11.4s +tttg: c176/289 lr:0.000334 t:11.5s +tttg: c177/289 lr:0.000329 t:11.5s +tttg: c178/289 lr:0.000324 t:11.6s +tttg: c179/289 lr:0.000319 t:11.7s +tttg: c180/289 lr:0.000314 t:11.7s +tttg: c181/289 lr:0.000309 t:11.8s +tttg: c182/289 lr:0.000304 t:11.9s +tttg: c183/289 lr:0.000299 t:11.9s +tttg: c184/289 lr:0.000294 t:12.0s +tttg: c185/289 lr:0.000289 t:12.1s +tttg: c186/289 lr:0.000284 t:12.1s +tttg: c187/289 lr:0.000279 t:12.2s +tttg: c188/289 lr:0.000274 t:12.3s +tttg: c189/289 lr:0.000269 t:12.3s +tttg: c190/289 lr:0.000264 t:12.4s +tttg: c191/289 lr:0.000260 t:12.5s +tttg: c192/289 lr:0.000255 t:12.5s +tttg: c193/289 lr:0.000250 t:12.6s +tttg: c194/289 lr:0.000245 t:12.7s +tttg: c195/289 lr:0.000241 t:12.7s +tttg: c196/289 lr:0.000236 t:12.8s +tttg: c197/289 lr:0.000231 t:12.9s +tttg: c198/289 lr:0.000227 t:12.9s +tttg: c199/289 lr:0.000222 t:13.0s +tttg: c200/289 lr:0.000218 t:13.0s +tttg: c201/289 lr:0.000213 t:13.1s +tttg: c202/289 lr:0.000209 t:13.2s +tttg: c203/289 lr:0.000204 t:13.2s +tttg: c204/289 lr:0.000200 t:13.3s +tttg: c205/289 lr:0.000196 t:13.4s +tttg: c206/289 lr:0.000191 t:13.4s +tttg: c207/289 lr:0.000187 t:13.5s +tttg: c208/289 lr:0.000183 t:13.6s +tttg: c209/289 lr:0.000179 t:13.6s +tttg: c210/289 lr:0.000174 t:13.7s +tttg: c211/289 lr:0.000170 t:13.8s +tttg: c212/289 lr:0.000166 t:13.8s +tttg: c213/289 lr:0.000162 t:13.9s +tttg: c214/289 lr:0.000158 t:14.0s +tttg: c215/289 lr:0.000154 t:14.0s +tttg: c216/289 lr:0.000150 t:14.1s +tttg: c217/289 lr:0.000146 t:14.1s +tttg: c218/289 lr:0.000143 t:14.2s +tttg: c219/289 lr:0.000139 t:14.3s +tttg: c220/289 lr:0.000135 t:14.3s +tttg: c221/289 lr:0.000131 t:14.4s +tttg: c222/289 lr:0.000128 t:14.5s +tttg: c223/289 lr:0.000124 t:14.5s +tttg: c224/289 lr:0.000121 t:14.6s +tttg: c225/289 lr:0.000117 t:14.7s +tttg: c226/289 lr:0.000113 t:14.7s +tttg: c227/289 lr:0.000110 t:14.8s +tttg: c228/289 lr:0.000107 t:14.9s +tttg: c229/289 lr:0.000103 t:14.9s +tttg: c230/289 lr:0.000100 t:15.0s +tttg: c231/289 lr:0.000097 t:15.1s +tttg: c232/289 lr:0.000094 t:15.1s +tttg: c233/289 lr:0.000090 t:15.2s +tttg: c234/289 lr:0.000087 t:15.3s +tttg: c235/289 lr:0.000084 t:15.3s +tttg: c236/289 lr:0.000081 t:15.4s +tttg: c237/289 lr:0.000078 t:15.4s +tttg: c238/289 lr:0.000075 t:15.5s +tttg: c239/289 lr:0.000073 t:15.6s +tttg: c240/289 lr:0.000070 t:15.6s +tttg: c241/289 lr:0.000067 t:15.7s +tttg: c242/289 lr:0.000064 t:15.8s +tttg: c243/289 lr:0.000062 t:15.8s +tttg: c244/289 lr:0.000059 t:15.9s +tttg: c245/289 lr:0.000056 t:16.0s +tttg: c246/289 lr:0.000054 t:16.0s +tttg: c247/289 lr:0.000052 t:16.1s +tttg: c248/289 lr:0.000049 t:16.2s +tttg: c249/289 lr:0.000047 t:16.2s +tttg: c250/289 lr:0.000045 t:16.3s +tttg: c251/289 lr:0.000042 t:16.4s +tttg: c252/289 lr:0.000040 t:16.4s +tttg: c253/289 lr:0.000038 t:16.5s +tttg: c254/289 lr:0.000036 t:16.6s +tttg: c255/289 lr:0.000034 t:16.6s +tttg: c256/289 lr:0.000032 t:16.7s +tttg: c257/289 lr:0.000030 t:16.7s +tttg: c258/289 lr:0.000028 t:16.8s +tttg: c259/289 lr:0.000027 t:16.9s +tttg: c260/289 lr:0.000025 t:16.9s +tttg: c261/289 lr:0.000023 t:17.0s +tttg: c262/289 lr:0.000022 t:17.1s +tttg: c263/289 lr:0.000020 t:17.1s +tttg: c264/289 lr:0.000018 t:17.2s +tttg: c265/289 lr:0.000017 t:17.3s +tttg: c266/289 lr:0.000016 t:17.3s +tttg: c267/289 lr:0.000014 t:17.4s +tttg: c268/289 lr:0.000013 t:17.5s +tttg: c269/289 lr:0.000012 t:17.5s +tttg: c270/289 lr:0.000011 t:17.6s +tttg: c271/289 lr:0.000010 t:17.7s +tttg: c272/289 lr:0.000009 t:17.7s +tttg: c273/289 lr:0.000008 t:17.8s +tttg: c274/289 lr:0.000007 t:17.9s +tttg: c275/289 lr:0.000006 t:17.9s +tttg: c276/289 lr:0.000005 t:18.0s +tttg: c277/289 lr:0.000004 t:18.1s +tttg: c278/289 lr:0.000004 t:18.1s +tttg: c279/289 lr:0.000003 t:18.2s +tttg: c280/289 lr:0.000002 t:18.3s +tttg: c281/289 lr:0.000002 t:18.3s +tttg: c282/289 lr:0.000001 t:18.4s +tttg: c283/289 lr:0.000001 t:18.5s +tttg: c284/289 lr:0.000001 t:18.5s +tttg: c285/289 lr:0.000000 t:18.6s +tttg: c286/289 lr:0.000000 t:18.6s +tttg: c287/289 lr:0.000000 t:18.7s +tttg: c288/289 lr:0.000000 t:18.8s +ttpr: phase:3/3 t:295.6s +ttp: b728/782 bl:2.3646 bb:1.0825 rl:2.3005 rb:1.0653 dl:2306-2324 gd:1 +ttp: b726/782 bl:2.3332 bb:1.0378 rl:2.3024 rb:1.0637 dl:2254-2276 gd:1 +ttp: b718/782 bl:2.2929 bb:1.0291 rl:2.3019 rb:1.0619 dl:2089-2106 gd:1 +ttp: b708/782 bl:2.3094 bb:1.0330 rl:2.3022 rb:1.0606 dl:1924-1937 gd:1 +ttp: b698/782 bl:2.2507 bb:1.0302 rl:2.3002 rb:1.0594 dl:1803-1814 gd:1 +ttp: b694/782 bl:2.3095 bb:1.0561 rl:2.3006 rb:1.0593 dl:1758-1769 gd:1 +ttp: b683/782 bl:2.2743 bb:1.0587 rl:2.2997 rb:1.0592 dl:1646-1657 gd:1 +ttp: b676/782 bl:2.3351 bb:1.0503 rl:2.3008 rb:1.0590 dl:1586-1595 gd:1 +ttp: b664/782 bl:2.3405 bb:1.0272 rl:2.3019 rb:1.0580 dl:1493-1499 gd:1 +ttp: b656/782 bl:2.3266 bb:1.1099 rl:2.3026 rb:1.0593 dl:1439-1445 gd:1 +ttp: b649/782 bl:2.2869 bb:1.0168 rl:2.3022 rb:1.0582 dl:1392-1398 gd:1 +ttp: b640/782 bl:2.3097 bb:1.0522 rl:2.3024 rb:1.0581 dl:1337-1343 gd:1 +ttp: b639/782 bl:2.3156 bb:1.0341 rl:2.3027 rb:1.0575 dl:1331-1337 gd:1 +ttp: b630/782 bl:2.3257 bb:1.0404 rl:2.3032 rb:1.0571 dl:1280-1285 gd:1 +ttp: b622/782 bl:2.2625 bb:1.0335 rl:2.3023 rb:1.0567 dl:1237-1243 gd:1 +ttp: b612/782 bl:2.2380 bb:1.0139 rl:2.3011 rb:1.0558 dl:1186-1190 gd:1 +ttp: b605/782 bl:2.2518 bb:1.0269 rl:2.3002 rb:1.0553 dl:1154-1159 gd:1 +ttp: b593/782 bl:2.2939 bb:1.0126 rl:2.3001 rb:1.0545 dl:1103-1107 gd:1 +ttp: b584/782 bl:2.3000 bb:1.0399 rl:2.3001 rb:1.0543 dl:1064-1069 gd:1 +ttp: b576/782 bl:2.3787 bb:1.0941 rl:2.3013 rb:1.0549 dl:1033-1037 gd:1 +ttp: b570/782 bl:2.3526 bb:1.0538 rl:2.3021 rb:1.0549 dl:1010-1014 gd:1 +ttp: b561/782 bl:2.2486 bb:1.0143 rl:2.3013 rb:1.0543 dl:979-983 gd:1 +ttp: b554/782 bl:2.4295 bb:1.0937 rl:2.3031 rb:1.0548 dl:955-959 gd:1 +ttp: b550/782 bl:2.3654 bb:1.0583 rl:2.3039 rb:1.0549 dl:943-946 gd:1 +ttp: b542/782 bl:2.3289 bb:1.0400 rl:2.3043 rb:1.0547 dl:918-921 gd:1 +ttp: b533/782 bl:2.3751 bb:1.0685 rl:2.3051 rb:1.0549 dl:890-892 gd:1 +ttp: b525/782 bl:2.3540 bb:1.0202 rl:2.3057 rb:1.0544 dl:866-869 gd:1 +ttp: b515/782 bl:2.3456 bb:1.0445 rl:2.3062 rb:1.0543 dl:838-841 gd:1 +ttp: b507/782 bl:2.2990 bb:1.0294 rl:2.3061 rb:1.0540 dl:814-817 gd:1 +ttp: b502/782 bl:2.3205 bb:1.0283 rl:2.3062 rb:1.0537 dl:802-804 gd:1 +ttp: b495/782 bl:2.3096 bb:1.0317 rl:2.3063 rb:1.0535 dl:783-785 gd:1 +ttp: b488/782 bl:2.2943 bb:1.0096 rl:2.3062 rb:1.0531 dl:766-769 gd:1 +ttp: b483/782 bl:2.2585 bb:1.0304 rl:2.3057 rb:1.0528 dl:754-756 gd:1 +ttp: b475/782 bl:2.3657 bb:1.0557 rl:2.3063 rb:1.0529 dl:735-737 gd:1 +ttp: b467/782 bl:2.3490 bb:1.0528 rl:2.3067 rb:1.0529 dl:717-719 gd:1 +ttp: b442/782 bl:2.2637 bb:1.0331 rl:2.3063 rb:1.0527 dl:664-666 gd:1 +ttp: b435/782 bl:2.3146 bb:1.0223 rl:2.3064 rb:1.0524 dl:648-651 gd:1 +ttp: b428/782 bl:2.3059 bb:1.0507 rl:2.3064 rb:1.0524 dl:636-638 gd:1 +ttp: b421/782 bl:2.2927 bb:1.0038 rl:2.3063 rb:1.0520 dl:622-624 gd:1 +ttp: b415/782 bl:2.2854 bb:1.0586 rl:2.3061 rb:1.0521 dl:611-613 gd:1 +ttp: b406/782 bl:2.3052 bb:1.0615 rl:2.3061 rb:1.0522 dl:593-595 gd:1 +ttp: b398/782 bl:2.2450 bb:1.0025 rl:2.3057 rb:1.0518 dl:579-581 gd:1 +ttp: b390/782 bl:2.3508 bb:1.0591 rl:2.3060 rb:1.0519 dl:564-566 gd:1 +ttp: b382/782 bl:2.2946 bb:1.0841 rl:2.3059 rb:1.0521 dl:550-552 gd:1 +ttp: b374/782 bl:2.3055 bb:1.0394 rl:2.3059 rb:1.0520 dl:537-538 gd:1 +ttp: b366/782 bl:2.3324 bb:1.0685 rl:2.3061 rb:1.0521 dl:524-525 gd:1 +ttp: b358/782 bl:2.4081 bb:1.0807 rl:2.3067 rb:1.0523 dl:510-512 gd:1 +ttp: b350/782 bl:2.3254 bb:1.0568 rl:2.3068 rb:1.0523 dl:497-498 gd:1 +ttp: b341/782 bl:2.2983 bb:1.0765 rl:2.3067 rb:1.0524 dl:483-485 gd:1 +ttp: b332/782 bl:2.3008 bb:1.0415 rl:2.3067 rb:1.0524 dl:469-471 gd:1 +ttp: b325/782 bl:2.3480 bb:1.0800 rl:2.3069 rb:1.0525 dl:459-461 gd:1 +ttp: b318/782 bl:2.3393 bb:1.0691 rl:2.3071 rb:1.0526 dl:448-450 gd:1 +ttp: b311/782 bl:2.3483 bb:1.0824 rl:2.3073 rb:1.0527 dl:438-439 gd:1 +ttp: b303/782 bl:2.3933 bb:1.0916 rl:2.3077 rb:1.0529 dl:426-427 gd:1 +ttp: b295/782 bl:2.2652 bb:1.0627 rl:2.3075 rb:1.0530 dl:414-415 gd:1 +ttp: b287/782 bl:2.4058 bb:1.0961 rl:2.3079 rb:1.0531 dl:402-403 gd:1 +ttp: b279/782 bl:2.3198 bb:1.0961 rl:2.3080 rb:1.0533 dl:391-392 gd:1 +ttp: b271/782 bl:2.3787 bb:1.1267 rl:2.3083 rb:1.0536 dl:380-382 gd:1 +ttp: b263/782 bl:2.3945 bb:1.0832 rl:2.3086 rb:1.0537 dl:370-371 gd:1 +ttp: b255/782 bl:2.3686 bb:1.0924 rl:2.3089 rb:1.0539 dl:360-361 gd:1 +ttp: b247/782 bl:2.3481 bb:1.0930 rl:2.3090 rb:1.0540 dl:350-351 gd:1 +ttp: b239/782 bl:2.3827 bb:1.1064 rl:2.3093 rb:1.0542 dl:340-341 gd:1 +ttp: b231/782 bl:2.3032 bb:1.0820 rl:2.3093 rb:1.0543 dl:330-331 gd:1 +ttp: b223/782 bl:2.3331 bb:1.1264 rl:2.3093 rb:1.0546 dl:321-322 gd:1 +ttp: b215/782 bl:2.3911 bb:1.0960 rl:2.3096 rb:1.0547 dl:312-313 gd:1 +ttp: b207/782 bl:2.3590 bb:1.1337 rl:2.3098 rb:1.0549 dl:303-304 gd:1 +ttp: b199/782 bl:2.4295 bb:1.1431 rl:2.3102 rb:1.0552 dl:295-296 gd:1 +ttp: b191/782 bl:2.4232 bb:1.1024 rl:2.3105 rb:1.0554 dl:285-286 gd:1 +ttp: b183/782 bl:2.3185 bb:1.0678 rl:2.3105 rb:1.0554 dl:277-278 gd:1 +ttp: b175/782 bl:2.3931 bb:1.1563 rl:2.3108 rb:1.0557 dl:269-270 gd:1 +ttp: b166/782 bl:2.4733 bb:1.1056 rl:2.3112 rb:1.0558 dl:260-262 gd:1 +ttp: b158/782 bl:2.3438 bb:1.1082 rl:2.3113 rb:1.0559 dl:253-254 gd:1 +ttp: b150/782 bl:2.3297 bb:1.1062 rl:2.3113 rb:1.0561 dl:245-246 gd:1 +ttp: b142/782 bl:2.3805 bb:1.1081 rl:2.3115 rb:1.0562 dl:237-238 gd:1 +ttp: b133/782 bl:2.3743 bb:1.1389 rl:2.3117 rb:1.0564 dl:229-230 gd:1 +ttp: b125/782 bl:2.4792 bb:1.1423 rl:2.3120 rb:1.0566 dl:222-222 gd:1 +ttp: b115/782 bl:2.4610 bb:1.1647 rl:2.3124 rb:1.0568 dl:212-213 gd:1 +ttp: b107/782 bl:2.4470 bb:1.1719 rl:2.3127 rb:1.0570 dl:205-206 gd:1 +ttp: b98/782 bl:2.5967 bb:1.2184 rl:2.3132 rb:1.0574 dl:197-198 gd:1 +ttp: b89/782 bl:2.4945 bb:1.1527 rl:2.3136 rb:1.0575 dl:189-190 gd:1 +ttp: b82/782 bl:2.4833 bb:1.1820 rl:2.3139 rb:1.0578 dl:183-183 gd:1 +ttp: b73/782 bl:2.5372 bb:1.2453 rl:2.3143 rb:1.0581 dl:174-175 gd:1 +ttp: b64/782 bl:2.5231 bb:1.1508 rl:2.3147 rb:1.0582 dl:166-167 gd:1 +ttp: b56/782 bl:2.5337 bb:1.2145 rl:2.3150 rb:1.0585 dl:159-160 gd:1 +ttp: b48/782 bl:2.5281 bb:1.2189 rl:2.3154 rb:1.0587 dl:151-152 gd:1 +ttp: b40/782 bl:2.4877 bb:1.1524 rl:2.3156 rb:1.0589 dl:143-144 gd:1 +ttp: b33/782 bl:2.5933 bb:1.2220 rl:2.3160 rb:1.0591 dl:136-137 gd:1 +ttp: b24/782 bl:2.4569 bb:1.1588 rl:2.3162 rb:1.0592 dl:127-128 gd:1 +ttp: b16/782 bl:2.6281 bb:1.2593 rl:2.3165 rb:1.0594 dl:117-118 gd:1 +ttp: b8/782 bl:2.7817 bb:1.2912 rl:2.3170 rb:1.0597 dl:103-105 gd:1 +quantized_ttt_phased val_loss:2.32225638 val_bpb:1.06118055 eval_time:392807ms +total_eval_time:392.8s +=== Done: SEED=42 === diff --git a/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/README.md b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/README.md new file mode 100644 index 0000000000..12cfbc0460 --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/README.md @@ -0,0 +1,54 @@ +# Record: SmearGate BOS Fix + PR #1787 Base + Smear Gate + LQER Asymmetric + Phased TTT + +**val_bpb = 1.06128** | **~15.95 MB** | 8xH100 SXM + +## Result + +| Seed | Pre-TTT BPB | Post-TTT BPB | Artifact (bytes) | +|------|-------------|--------------|------------------| +| 42 | 1.07406 | **1.06128** | 15,952,086 | + +Merged SOTA (PR #1493): **1.0810 BPP**. Delta: **-0.0197 BPP**. Clears the 0.005-nat threshold. + +## Key Change: SmearGate BOS Document Boundary Fix + +Builds on PR #1797 stack (PR #1787 base + SmearGate + LQER Asymmetric) but fixes the **SmearGate cross-document leakage bug** identified by @cocohearts in PR #1797 audit. + +The bug: SmearGate 1-token causal lookback does not mask BOS positions, so the final token of document N smears into BOS of document N+1. + +The fix (applied in both forward_logits and forward_ttt): + + bos_mask = (input_ids[:, 1:] == 1).unsqueeze(-1) + g = g.masked_fill(bos_mask, 0.0) + +## Technique Stack + +| Component | Origin | +|-----------|--------| +| CaseOps bijective case transform | PR #1729 / PR #1736 | +| SparseAttnGate | PR #1787 (nprime06) | +| SmearGate + BOS fix | PR #1797 + this submission | +| LQER asymmetric rank-4 | PR #1797 | +| Phased TTT (score-first, 3 phases) | PR #1394 / PR #1736 | +| PolarNS + MIN_LR=0.1 + FusedCE | PR #1787 | +| Full Hessian GPTQ + Brotli | PR #1019 / PR #1530 | + +## Architecture + +11L x 512d x 8H/4KV, MLP 4x, LeakyReLU(0.5)^2, Partial RoPE (16/64 dims), layerwise LN scale, tied embeddings, logit softcap=30.0. Depth recurrence: layers 3-5 looped x2 (activated at frac=0.35). Parallel residuals from layer 8. XSA on all 11 layers. SmearGate window=12. + +## Compliance + +- Artifact <= 16,000,000 bytes: 15,952,086 bytes +- train_time <= 600s: 599.6s +- eval_time <= 600s: 519.5s +- Issue #1017 Conditions 1-4: All satisfied. SmearGate BOS mask ensures no cross-document leakage. + +## Credits + +- @nprime06 -- PR #1787 base stack +- @romeerp -- CaseOps transform (PR #1729) +- @dexhunter -- SmearGate + LQER (PR #1797) +- @cocohearts -- Identifying SmearGate BOS bug +- @abaybektursun -- Score-first TTT (PR #549) +- @clarkkev -- GPTQ SDClip + SP8192 (PR #1394) diff --git a/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/lossless_caps.py b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/lossless_caps.py new file mode 100644 index 0000000000..98e472f824 --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/lossless_caps.py @@ -0,0 +1,833 @@ +"""Lossless capitalization pre-encoding helpers. + +This module provides a narrow, reversible transform that only touches +ASCII capital letters `A-Z`. Each uppercase ASCII letter is rewritten as +``, where `sentinel` is a private-use Unicode +character that is escaped by doubling if it appears literally in the +input text. + +Example with the default sentinel `\\uE000`: + + "The NASA Launch" -> "\\uE000the \\uE000n\\uE000a\\uE000s\\uE000a \\uE000launch" + +The transform is intentionally simple for v1: + +- lowercase ASCII letters are unchanged +- uppercase ASCII letters become sentinel + lowercase letter +- non-ASCII characters are left untouched +- literal sentinel characters are escaped as sentinel + sentinel + +This makes the transform exactly invertible while allowing a downstream +tokenizer to reuse lowercase subwords across case variants. +""" + +from __future__ import annotations + +import json +from pathlib import Path +from typing import Callable, Iterable + +LOSSLESS_CAPS_V1 = "lossless_caps_v1" +LOSSLESS_CAPS_V2 = "lossless_caps_v2" +LOSSLESS_CAPS_V3 = "lossless_caps_v3" +LOSSLESS_CAPS_V4 = "lossless_caps_v4" +LOSSLESS_CAPS_V5 = "lossless_caps_v5" +LOSSLESS_CAPS_V6 = "lossless_caps_v6" +LOSSLESS_CAPS_V7 = "lossless_caps_v7" +LOSSLESS_CAPS_CASEOPS_V1 = "lossless_caps_caseops_v1" +IDENTITY = "identity" +DEFAULT_SENTINEL = "\uE000" +DEFAULT_V2_TITLE = "\uE001" +DEFAULT_V2_ALLCAPS = "\uE002" +DEFAULT_V2_CAPNEXT = "\uE003" +DEFAULT_V2_ESC = "\uE004" +DEFAULT_V5_TITLE_MIN_LEN = 7 +DEFAULT_V6_ALLCAPS_MIN_LEN = 3 +DEFAULT_V7_ALLCAPS_MIN_LEN = 4 + + +class LosslessCapsError(ValueError): + """Raised when a transformed string is malformed.""" + + +def _is_ascii_upper(ch: str) -> bool: + return "A" <= ch <= "Z" + + +def _is_ascii_lower(ch: str) -> bool: + return "a" <= ch <= "z" + + +def _is_ascii_alpha(ch: str) -> bool: + return _is_ascii_lower(ch) or _is_ascii_upper(ch) + + +def _validate_distinct_single_chars(*chars: str) -> None: + if any(len(ch) != 1 for ch in chars): + raise ValueError("all control characters must be exactly one character") + if len(set(chars)) != len(chars): + raise ValueError("control characters must be distinct") + + +def encode_lossless_caps_v1(text: str, *, sentinel: str = DEFAULT_SENTINEL) -> str: + """Encode ASCII capitals reversibly using a one-character sentinel.""" + if len(sentinel) != 1: + raise ValueError("sentinel must be exactly one character") + out: list[str] = [] + for ch in text: + if ch == sentinel: + out.append(sentinel) + out.append(sentinel) + elif _is_ascii_upper(ch): + out.append(sentinel) + out.append(ch.lower()) + else: + out.append(ch) + return "".join(out) + + +def decode_lossless_caps_v1(text: str, *, sentinel: str = DEFAULT_SENTINEL) -> str: + """Decode the `lossless_caps_v1` transform back to the original text.""" + if len(sentinel) != 1: + raise ValueError("sentinel must be exactly one character") + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch != sentinel: + out.append(ch) + i += 1 + continue + if i + 1 >= n: + raise LosslessCapsError("dangling capitalization sentinel at end of string") + nxt = text[i + 1] + if nxt == sentinel: + out.append(sentinel) + elif _is_ascii_lower(nxt): + out.append(nxt.upper()) + else: + raise LosslessCapsError( + f"invalid sentinel escape sequence {sentinel + nxt!r}; " + "expected doubled sentinel or sentinel + lowercase ASCII letter" + ) + i += 2 + return "".join(out) + + +def encode_lossless_caps_v2( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + capnext: str = DEFAULT_V2_CAPNEXT, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Encode ASCII word capitalization with cheap word-level markers. + + Rules over maximal ASCII alphabetic runs: + - lowercase words stay unchanged + - TitleCase words become `title + lowercase(word)` + - ALLCAPS words become `allcaps + lowercase(word)` + - mixed-case words use: + - optional `title` when the first letter is uppercase + - `capnext + lowercase(letter)` for subsequent uppercase letters + - literal control characters are escaped as `esc + literal` + """ + _validate_distinct_single_chars(title, allcaps, capnext, esc) + controls = {title, allcaps, capnext, esc} + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch in controls: + out.append(esc) + out.append(ch) + i += 1 + continue + if not _is_ascii_alpha(ch): + out.append(ch) + i += 1 + continue + + j = i + 1 + while j < n and _is_ascii_alpha(text[j]): + j += 1 + word = text[i:j] + lower_word = word.lower() + + if word.islower(): + out.append(word) + elif len(word) >= 2 and word.isupper(): + out.append(allcaps) + out.append(lower_word) + elif _is_ascii_upper(word[0]) and word[1:].islower(): + out.append(title) + out.append(lower_word) + else: + if _is_ascii_upper(word[0]): + out.append(title) + out.append(lower_word[0]) + for orig_ch, lower_ch in zip(word[1:], lower_word[1:], strict=True): + if _is_ascii_upper(orig_ch): + out.append(capnext) + out.append(lower_ch) + i = j + return "".join(out) + + +def decode_lossless_caps_v2( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + capnext: str = DEFAULT_V2_CAPNEXT, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v2` transform back to the original text.""" + _validate_distinct_single_chars(title, allcaps, capnext, esc) + out: list[str] = [] + pending_escape = False + pending_word_mode: str | None = None + active_allcaps = False + pending_capnext = False + in_ascii_word = False + + for ch in text: + if pending_escape: + if pending_word_mode is not None and not _is_ascii_alpha(ch): + raise LosslessCapsError("escaped control char cannot satisfy pending word capitalization mode") + out.append(ch) + pending_escape = False + if _is_ascii_alpha(ch): + in_ascii_word = True + else: + in_ascii_word = False + active_allcaps = False + continue + + if ch == esc: + pending_escape = True + continue + if ch == title: + if pending_word_mode is not None or in_ascii_word or pending_capnext: + raise LosslessCapsError("invalid title marker placement") + pending_word_mode = "title" + continue + if ch == allcaps: + if pending_word_mode is not None or in_ascii_word or pending_capnext: + raise LosslessCapsError("invalid allcaps marker placement") + pending_word_mode = "allcaps" + continue + if ch == capnext: + if pending_capnext: + raise LosslessCapsError("duplicate capnext marker") + pending_capnext = True + continue + + if _is_ascii_alpha(ch): + at_word_start = not in_ascii_word + if at_word_start: + if pending_word_mode == "allcaps": + out.append(ch.upper()) + active_allcaps = True + elif pending_word_mode == "title": + out.append(ch.upper()) + elif pending_capnext: + out.append(ch.upper()) + else: + out.append(ch) + pending_word_mode = None + pending_capnext = False + in_ascii_word = True + continue + + if pending_word_mode is not None: + raise LosslessCapsError("word capitalization marker leaked into the middle of a word") + if active_allcaps: + out.append(ch.upper()) + elif pending_capnext: + out.append(ch.upper()) + else: + out.append(ch) + pending_capnext = False + continue + + if pending_word_mode is not None or pending_capnext: + raise LosslessCapsError("capitalization marker not followed by an ASCII letter") + out.append(ch) + in_ascii_word = False + active_allcaps = False + + if pending_escape: + raise LosslessCapsError("dangling escape marker at end of string") + if pending_word_mode is not None or pending_capnext: + raise LosslessCapsError("dangling capitalization marker at end of string") + return "".join(out) + + +def encode_lossless_caps_v3( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Encode only common word-level capitalization patterns. + + Rules over maximal ASCII alphabetic runs: + - lowercase words stay unchanged + - TitleCase words become `title + lowercase(word)` + - ALLCAPS words become `allcaps + lowercase(word)` + - all other mixed-case words are left unchanged + - literal control characters are escaped as `esc + literal` + """ + _validate_distinct_single_chars(title, allcaps, esc) + controls = {title, allcaps, esc} + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch in controls: + out.append(esc) + out.append(ch) + i += 1 + continue + if not _is_ascii_alpha(ch): + out.append(ch) + i += 1 + continue + + j = i + 1 + while j < n and _is_ascii_alpha(text[j]): + j += 1 + word = text[i:j] + + if word.islower(): + out.append(word) + elif len(word) >= 2 and word.isupper(): + out.append(allcaps) + out.append(word.lower()) + elif _is_ascii_upper(word[0]) and word[1:].islower(): + out.append(title) + out.append(word.lower()) + else: + out.append(word) + i = j + return "".join(out) + + +def decode_lossless_caps_v3( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v3` transform back to the original text.""" + _validate_distinct_single_chars(title, allcaps, esc) + out: list[str] = [] + pending_escape = False + pending_word_mode: str | None = None + active_allcaps = False + in_ascii_word = False + + for ch in text: + if pending_escape: + if pending_word_mode is not None and not _is_ascii_alpha(ch): + raise LosslessCapsError("escaped control char cannot satisfy pending word capitalization mode") + out.append(ch) + pending_escape = False + if _is_ascii_alpha(ch): + in_ascii_word = True + else: + in_ascii_word = False + active_allcaps = False + continue + + if ch == esc: + pending_escape = True + continue + if ch == title: + if pending_word_mode is not None or in_ascii_word: + raise LosslessCapsError("invalid title marker placement") + pending_word_mode = "title" + continue + if ch == allcaps: + if pending_word_mode is not None or in_ascii_word: + raise LosslessCapsError("invalid allcaps marker placement") + pending_word_mode = "allcaps" + continue + + if _is_ascii_alpha(ch): + at_word_start = not in_ascii_word + if at_word_start: + if pending_word_mode == "allcaps": + out.append(ch.upper()) + active_allcaps = True + elif pending_word_mode == "title": + out.append(ch.upper()) + else: + out.append(ch) + pending_word_mode = None + in_ascii_word = True + continue + + if pending_word_mode is not None: + raise LosslessCapsError("word capitalization marker leaked into the middle of a word") + out.append(ch.upper() if active_allcaps else ch) + continue + + if pending_word_mode is not None: + raise LosslessCapsError("capitalization marker not followed by an ASCII letter") + out.append(ch) + in_ascii_word = False + active_allcaps = False + + if pending_escape: + raise LosslessCapsError("dangling escape marker at end of string") + if pending_word_mode is not None: + raise LosslessCapsError("dangling capitalization marker at end of string") + return "".join(out) + + +def encode_lossless_caps_v4( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Encode only ALLCAPS ASCII words, leaving all other case untouched.""" + _validate_distinct_single_chars(allcaps, esc) + controls = {allcaps, esc} + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch in controls: + out.append(esc) + out.append(ch) + i += 1 + continue + if not _is_ascii_alpha(ch): + out.append(ch) + i += 1 + continue + j = i + 1 + while j < n and _is_ascii_alpha(text[j]): + j += 1 + word = text[i:j] + if len(word) >= 2 and word.isupper(): + out.append(allcaps) + out.append(word.lower()) + else: + out.append(word) + i = j + return "".join(out) + + +def decode_lossless_caps_v4( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v4` transform back to the original text.""" + _validate_distinct_single_chars(allcaps, esc) + out: list[str] = [] + pending_escape = False + pending_allcaps = False + in_ascii_word = False + active_allcaps = False + + for ch in text: + if pending_escape: + if pending_allcaps and not _is_ascii_alpha(ch): + raise LosslessCapsError("escaped control char cannot satisfy pending allcaps mode") + out.append(ch) + pending_escape = False + if _is_ascii_alpha(ch): + in_ascii_word = True + else: + in_ascii_word = False + active_allcaps = False + continue + + if ch == esc: + pending_escape = True + continue + if ch == allcaps: + if pending_allcaps or in_ascii_word: + raise LosslessCapsError("invalid allcaps marker placement") + pending_allcaps = True + continue + + if _is_ascii_alpha(ch): + if not in_ascii_word: + active_allcaps = pending_allcaps + pending_allcaps = False + in_ascii_word = True + out.append(ch.upper() if active_allcaps else ch) + continue + + if pending_allcaps: + raise LosslessCapsError("allcaps marker not followed by an ASCII letter") + out.append(ch) + in_ascii_word = False + active_allcaps = False + + if pending_escape: + raise LosslessCapsError("dangling escape marker at end of string") + if pending_allcaps: + raise LosslessCapsError("dangling allcaps marker at end of string") + return "".join(out) + + +def encode_lossless_caps_v5( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, + title_min_len: int = DEFAULT_V5_TITLE_MIN_LEN, +) -> str: + """Encode ALLCAPS words and only sufficiently long TitleCase words.""" + _validate_distinct_single_chars(title, allcaps, esc) + controls = {title, allcaps, esc} + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch in controls: + out.append(esc) + out.append(ch) + i += 1 + continue + if not _is_ascii_alpha(ch): + out.append(ch) + i += 1 + continue + j = i + 1 + while j < n and _is_ascii_alpha(text[j]): + j += 1 + word = text[i:j] + if len(word) >= 2 and word.isupper(): + out.append(allcaps) + out.append(word.lower()) + elif len(word) >= title_min_len and _is_ascii_upper(word[0]) and word[1:].islower(): + out.append(title) + out.append(word.lower()) + else: + out.append(word) + i = j + return "".join(out) + + +def decode_lossless_caps_v5( + text: str, + *, + title: str = DEFAULT_V2_TITLE, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v5` transform back to the original text.""" + return decode_lossless_caps_v3(text, title=title, allcaps=allcaps, esc=esc) + + +def encode_lossless_caps_v6( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, + allcaps_min_len: int = DEFAULT_V6_ALLCAPS_MIN_LEN, +) -> str: + """Encode only ALLCAPS words with length >= allcaps_min_len.""" + _validate_distinct_single_chars(allcaps, esc) + controls = {allcaps, esc} + out: list[str] = [] + i = 0 + n = len(text) + while i < n: + ch = text[i] + if ch in controls: + out.append(esc) + out.append(ch) + i += 1 + continue + if not _is_ascii_alpha(ch): + out.append(ch) + i += 1 + continue + j = i + 1 + while j < n and _is_ascii_alpha(text[j]): + j += 1 + word = text[i:j] + if len(word) >= allcaps_min_len and word.isupper(): + out.append(allcaps) + out.append(word.lower()) + else: + out.append(word) + i = j + return "".join(out) + + +def decode_lossless_caps_v6( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v6` transform back to the original text.""" + return decode_lossless_caps_v4(text, allcaps=allcaps, esc=esc) + + +def encode_lossless_caps_v7( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, + allcaps_min_len: int = DEFAULT_V7_ALLCAPS_MIN_LEN, +) -> str: + """Encode only ALLCAPS words with length >= 4.""" + return encode_lossless_caps_v6( + text, + allcaps=allcaps, + esc=esc, + allcaps_min_len=allcaps_min_len, + ) + + +def decode_lossless_caps_v7( + text: str, + *, + allcaps: str = DEFAULT_V2_ALLCAPS, + esc: str = DEFAULT_V2_ESC, +) -> str: + """Decode the `lossless_caps_v7` transform back to the original text.""" + return decode_lossless_caps_v6(text, allcaps=allcaps, esc=esc) + + +def get_text_transform(name: str | None) -> Callable[[str], str]: + """Return the forward text transform for the given config name.""" + normalized = IDENTITY if name in {None, "", IDENTITY} else str(name) + if normalized == IDENTITY: + return lambda text: text + if normalized == LOSSLESS_CAPS_V1: + return encode_lossless_caps_v1 + if normalized == LOSSLESS_CAPS_V2: + return encode_lossless_caps_v2 + if normalized == LOSSLESS_CAPS_V3: + return encode_lossless_caps_v3 + if normalized == LOSSLESS_CAPS_V4: + return encode_lossless_caps_v4 + if normalized == LOSSLESS_CAPS_V5: + return encode_lossless_caps_v5 + if normalized == LOSSLESS_CAPS_V6: + return encode_lossless_caps_v6 + if normalized == LOSSLESS_CAPS_V7: + return encode_lossless_caps_v7 + if normalized == LOSSLESS_CAPS_CASEOPS_V1: + return encode_lossless_caps_v2 + raise ValueError(f"unsupported text_transform={name!r}") + + +def get_text_inverse_transform(name: str | None) -> Callable[[str], str]: + """Return the inverse transform for the given config name.""" + normalized = IDENTITY if name in {None, "", IDENTITY} else str(name) + if normalized == IDENTITY: + return lambda text: text + if normalized == LOSSLESS_CAPS_V1: + return decode_lossless_caps_v1 + if normalized == LOSSLESS_CAPS_V2: + return decode_lossless_caps_v2 + if normalized == LOSSLESS_CAPS_V3: + return decode_lossless_caps_v3 + if normalized == LOSSLESS_CAPS_V4: + return decode_lossless_caps_v4 + if normalized == LOSSLESS_CAPS_V5: + return decode_lossless_caps_v5 + if normalized == LOSSLESS_CAPS_V6: + return decode_lossless_caps_v6 + if normalized == LOSSLESS_CAPS_V7: + return decode_lossless_caps_v7 + if normalized == LOSSLESS_CAPS_CASEOPS_V1: + return decode_lossless_caps_v2 + raise ValueError(f"unsupported text_transform={name!r}") + + +def normalize_text_transform_name(name: str | None) -> str: + """Normalize empty/None transform names to the identity transform.""" + return IDENTITY if name in {None, "", IDENTITY} else str(name) + + +def get_text_transform_control_symbols(name: str | None) -> list[str]: + """Return reserved control symbols used by a transform, if any.""" + normalized = normalize_text_transform_name(name) + if normalized == IDENTITY: + return [] + if normalized == LOSSLESS_CAPS_V1: + return [DEFAULT_SENTINEL] + if normalized == LOSSLESS_CAPS_V2: + return [DEFAULT_V2_TITLE, DEFAULT_V2_ALLCAPS, DEFAULT_V2_CAPNEXT, DEFAULT_V2_ESC] + if normalized == LOSSLESS_CAPS_CASEOPS_V1: + return [DEFAULT_V2_TITLE, DEFAULT_V2_ALLCAPS, DEFAULT_V2_CAPNEXT, DEFAULT_V2_ESC] + if normalized in {LOSSLESS_CAPS_V3, LOSSLESS_CAPS_V5}: + return [DEFAULT_V2_TITLE, DEFAULT_V2_ALLCAPS, DEFAULT_V2_ESC] + if normalized in {LOSSLESS_CAPS_V4, LOSSLESS_CAPS_V6, LOSSLESS_CAPS_V7}: + return [DEFAULT_V2_ALLCAPS, DEFAULT_V2_ESC] + raise ValueError(f"unsupported text_transform={name!r}") + + +def infer_text_transform_from_manifest(tokenizer_path: str | Path) -> str: + """Best-effort lookup of a tokenizer's text transform from a local manifest.""" + tokenizer_path = Path(tokenizer_path).expanduser().resolve() + manifest_candidates = [ + tokenizer_path.parent.parent / "manifest.json", + tokenizer_path.parent / "manifest.json", + ] + for manifest_path in manifest_candidates: + if not manifest_path.is_file(): + continue + try: + payload = json.loads(manifest_path.read_text(encoding="utf-8")) + except (OSError, json.JSONDecodeError): + continue + tokenizers = payload.get("tokenizers") + if not isinstance(tokenizers, list): + continue + for tokenizer_meta in tokenizers: + if not isinstance(tokenizer_meta, dict): + continue + model_path = tokenizer_meta.get("model_path") or tokenizer_meta.get("path") + if not model_path: + continue + candidate = (manifest_path.parent / str(model_path)).resolve() + if candidate == tokenizer_path: + return normalize_text_transform_name(tokenizer_meta.get("text_transform")) + return IDENTITY + + +def surface_piece_original_byte_counts( + surfaces: Iterable[str], + *, + text_transform_name: str | None = None, + sentinel: str = DEFAULT_SENTINEL, +) -> list[int]: + """Return exact original UTF-8 byte counts contributed by each surface piece. + + `surfaces` must be the exact decoded text fragments emitted by SentencePiece + in order, e.g. `piece.surface` from `encode_as_immutable_proto`. + """ + normalized = normalize_text_transform_name(text_transform_name) + if normalized == IDENTITY: + return [len(surface.encode("utf-8")) for surface in surfaces] + if normalized == LOSSLESS_CAPS_V1: + if len(sentinel) != 1: + raise ValueError("sentinel must be exactly one character") + sentinel_bytes = len(sentinel.encode("utf-8")) + pending_sentinel = False + counts: list[int] = [] + for surface in surfaces: + piece_bytes = 0 + for ch in surface: + if pending_sentinel: + if ch == sentinel: + piece_bytes += sentinel_bytes + elif _is_ascii_lower(ch): + piece_bytes += 1 + else: + raise LosslessCapsError( + f"invalid continuation {ch!r} after capitalization sentinel" + ) + pending_sentinel = False + continue + if ch == sentinel: + pending_sentinel = True + else: + piece_bytes += len(ch.encode("utf-8")) + counts.append(piece_bytes) + if pending_sentinel: + raise LosslessCapsError("dangling capitalization sentinel across piece boundary") + return counts + if normalized not in {LOSSLESS_CAPS_V2, LOSSLESS_CAPS_V3, LOSSLESS_CAPS_V4, LOSSLESS_CAPS_V5, LOSSLESS_CAPS_V6, LOSSLESS_CAPS_V7, LOSSLESS_CAPS_CASEOPS_V1}: + raise ValueError(f"unsupported text_transform={text_transform_name!r}") + + title = DEFAULT_V2_TITLE + allcaps = DEFAULT_V2_ALLCAPS + capnext = DEFAULT_V2_CAPNEXT + esc = DEFAULT_V2_ESC + if normalized in {LOSSLESS_CAPS_V2, LOSSLESS_CAPS_CASEOPS_V1}: + _validate_distinct_single_chars(title, allcaps, capnext, esc) + elif normalized in {LOSSLESS_CAPS_V4, LOSSLESS_CAPS_V6, LOSSLESS_CAPS_V7}: + _validate_distinct_single_chars(allcaps, esc) + else: + _validate_distinct_single_chars(title, allcaps, esc) + pending_escape = False + pending_word_mode: str | None = None + active_allcaps = False + pending_capnext = False + in_ascii_word = False + counts: list[int] = [] + for surface in surfaces: + piece_bytes = 0 + for ch in surface: + if pending_escape: + if pending_word_mode is not None and not _is_ascii_alpha(ch): + raise LosslessCapsError("escaped control char cannot satisfy pending word capitalization mode") + piece_bytes += len(ch.encode("utf-8")) + pending_escape = False + if _is_ascii_alpha(ch): + in_ascii_word = True + else: + in_ascii_word = False + active_allcaps = False + continue + if ch == esc: + pending_escape = True + continue + if normalized in {LOSSLESS_CAPS_V2, LOSSLESS_CAPS_V3, LOSSLESS_CAPS_V5, LOSSLESS_CAPS_CASEOPS_V1} and ch == title: + if pending_word_mode is not None or in_ascii_word or pending_capnext: + raise LosslessCapsError("invalid title marker placement") + pending_word_mode = "title" + continue + if ch == allcaps: + if pending_word_mode is not None or in_ascii_word or pending_capnext: + raise LosslessCapsError("invalid allcaps marker placement") + pending_word_mode = "allcaps" + continue + if normalized in {LOSSLESS_CAPS_V2, LOSSLESS_CAPS_CASEOPS_V1} and ch == capnext: + if pending_capnext: + raise LosslessCapsError("duplicate capnext marker") + pending_capnext = True + continue + + if _is_ascii_alpha(ch): + at_word_start = not in_ascii_word + if at_word_start: + piece_bytes += 1 + active_allcaps = pending_word_mode == "allcaps" + pending_word_mode = None + pending_capnext = False + in_ascii_word = True + continue + if pending_word_mode is not None: + raise LosslessCapsError("word capitalization marker leaked into the middle of a word") + piece_bytes += 1 + pending_capnext = False + continue + + if pending_word_mode is not None or pending_capnext: + raise LosslessCapsError("capitalization marker not followed by an ASCII letter") + piece_bytes += len(ch.encode("utf-8")) + in_ascii_word = False + active_allcaps = False + counts.append(piece_bytes) + if pending_escape: + raise LosslessCapsError("dangling escape marker across piece boundary") + if pending_word_mode is not None or pending_capnext: + raise LosslessCapsError("dangling capitalization marker across piece boundary") + return counts diff --git a/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/prepare_caseops_data.py b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/prepare_caseops_data.py new file mode 100644 index 0000000000..5c3f13e69c --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/prepare_caseops_data.py @@ -0,0 +1,177 @@ +"""Prepare CaseOps-tokenized FineWeb shards + per-token byte sidecar. + +CaseOps (``lossless_caps_caseops_v1``) is a bijective, character-level text +transform that introduces four operator tokens in place of explicit +capitalization: TITLE, ALLCAPS, CAPNEXT, ESC. The transform is fully +reversible — no information is lost relative to the untransformed UTF-8 +text, so BPB stays computable on TRUE byte counts. + +Forward pipeline: + 1. Read the canonical FineWeb-10B doc stream (``docs_selected.jsonl`` + produced by ``data/download_hf_docs_and_tokenize.py`` in the root repo). + 2. Apply ``encode_lossless_caps_v2`` (the caseops_v1 alias) to each doc. + 3. Tokenize with the shipped SP model + ``tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model`` + (reserves TITLE/ALLCAPS/CAPNEXT/ESC + sentinel as user_defined_symbols). + 4. Write uint16 train/val shards (``fineweb_{train,val}_XXXXXX.bin``). + 5. For the VAL stream only, emit per-token byte sidecar shards + (``fineweb_val_bytes_XXXXXX.bin``, uint16 parallel arrays) that record + each token's ORIGINAL pre-transform UTF-8 byte count. BPB is computed + from these canonical bytes so the score is on the untransformed text + (not the transformed representation). + +Output layout — matches what ``train_gpt.py`` expects under +``DATA_DIR=./data`` with ``CASEOPS_ENABLED=1``: + + data/datasets/fineweb10B_sp8192_caseops/datasets/ + tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model + datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/ + fineweb_train_000000.bin + fineweb_train_000001.bin + ... + fineweb_val_000000.bin + fineweb_val_bytes_000000.bin + +Usage: + + python3 prepare_caseops_data.py \\ + --docs ./fineweb10B_raw/docs_selected.jsonl \\ + --out ./data/datasets/fineweb10B_sp8192_caseops/datasets \\ + --sp ./tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model + +Requirements: sentencepiece, numpy. CPU-only. Runs once; reused across seeds. +""" +from __future__ import annotations + +import argparse +import json +import pathlib +import struct +import sys + +import numpy as np +import sentencepiece as spm + +# Local import — lossless_caps.py ships next to this script. +sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent)) +from lossless_caps import ( # noqa: E402 + LOSSLESS_CAPS_CASEOPS_V1, + encode_lossless_caps_v2, + surface_piece_original_byte_counts, +) + + +SHARD_MAGIC = 20240520 +SHARD_VERSION = 1 +SHARD_TOKENS = 10_000_000 # tokens per shard — matches the main pipeline +BOS_ID = 1 # SP model's control token; train_gpt.py:_find_docs requires BOS per doc + + +def _write_shard(out_path: pathlib.Path, arr: np.ndarray) -> None: + """Write a uint16 shard in the standard header-prefixed format.""" + assert arr.dtype == np.uint16 + header = np.zeros(256, dtype=np.int32) + header[0] = SHARD_MAGIC + header[1] = SHARD_VERSION + header[2] = int(arr.size) + with out_path.open("wb") as fh: + fh.write(header.tobytes()) + fh.write(arr.tobytes()) + + +def _iter_docs(docs_path: pathlib.Path): + """Yield doc strings from a jsonl file (one json object per line).""" + with docs_path.open("r", encoding="utf-8") as fh: + for line in fh: + line = line.strip() + if not line: + continue + obj = json.loads(line) + # Support both {"text": ...} and raw strings. + yield obj["text"] if isinstance(obj, dict) else obj + + +def _token_original_byte_counts( + sp: spm.SentencePieceProcessor, + original_text: str, + transformed_text: str, +) -> np.ndarray: + """Per-token canonical (pre-transform) UTF-8 byte counts. + + Delegates to ``surface_piece_original_byte_counts`` in ``lossless_caps.py`` + — the canonical exporter used by the PR #1729 / HF-hosted CaseOps dataset. + Operator pieces (U+E001..U+E004) contribute 0 original bytes; letter pieces + contribute their pre-transform UTF-8 byte count. + """ + proto = sp.encode_as_immutable_proto(transformed_text) + byte_counts = surface_piece_original_byte_counts( + (piece.surface for piece in proto.pieces), + text_transform_name=LOSSLESS_CAPS_CASEOPS_V1, + ) + return np.asarray(list(byte_counts), dtype=np.uint16) + + +def main() -> None: + ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) + ap.add_argument("--docs", required=True, type=pathlib.Path, help="Path to docs_selected.jsonl") + ap.add_argument("--out", required=True, type=pathlib.Path, help="Output datasets dir") + ap.add_argument("--sp", required=True, type=pathlib.Path, help="Path to CaseOps SP model") + ap.add_argument("--val-docs", type=int, default=10_000, help="Validation docs count") + args = ap.parse_args() + + sp = spm.SentencePieceProcessor(model_file=str(args.sp)) + print(f"loaded sp: vocab={sp.vocab_size()}", flush=True) + + train_out = args.out / "datasets" / "fineweb10B_sp8192_lossless_caps_caseops_v1_reserved" + train_out.mkdir(parents=True, exist_ok=True) + + val_buf_tokens: list[int] = [] + val_buf_bytes: list[int] = [] + train_buf: list[int] = [] + val_written = 0 + train_written = 0 + n_docs = 0 + + for text in _iter_docs(args.docs): + transformed = encode_lossless_caps_v2(text) + token_ids = [BOS_ID] + sp.encode(transformed, out_type=int) + if n_docs < args.val_docs: + # Validation doc — also compute byte sidecar + byte_counts = _token_original_byte_counts(sp, text, transformed) + val_buf_tokens.extend(token_ids) + val_buf_bytes.append(0) # BOS contributes 0 original bytes + val_buf_bytes.extend(int(b) for b in byte_counts) + if len(val_buf_tokens) >= SHARD_TOKENS: + _write_shard(train_out / f"fineweb_val_{val_written:06d}.bin", + np.array(val_buf_tokens[:SHARD_TOKENS], dtype=np.uint16)) + _write_shard(train_out / f"fineweb_val_bytes_{val_written:06d}.bin", + np.array(val_buf_bytes[:SHARD_TOKENS], dtype=np.uint16)) + val_buf_tokens = val_buf_tokens[SHARD_TOKENS:] + val_buf_bytes = val_buf_bytes[SHARD_TOKENS:] + val_written += 1 + else: + train_buf.extend(token_ids) + if len(train_buf) >= SHARD_TOKENS: + _write_shard(train_out / f"fineweb_train_{train_written:06d}.bin", + np.array(train_buf[:SHARD_TOKENS], dtype=np.uint16)) + train_buf = train_buf[SHARD_TOKENS:] + train_written += 1 + n_docs += 1 + if n_docs % 10_000 == 0: + print(f" processed {n_docs} docs train_shards={train_written} val_shards={val_written}", flush=True) + + # Flush tail buffers into final (possibly short) shards. + if val_buf_tokens: + _write_shard(train_out / f"fineweb_val_{val_written:06d}.bin", + np.array(val_buf_tokens, dtype=np.uint16)) + _write_shard(train_out / f"fineweb_val_bytes_{val_written:06d}.bin", + np.array(val_buf_bytes, dtype=np.uint16)) + if train_buf: + _write_shard(train_out / f"fineweb_train_{train_written:06d}.bin", + np.array(train_buf, dtype=np.uint16)) + + print(f"done. docs={n_docs} train_shards={train_written + (1 if train_buf else 0)} val_shards={val_written + (1 if val_buf_tokens else 0)}") + + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/submission.json b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/submission.json new file mode 100644 index 0000000000..76b8351b61 --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/submission.json @@ -0,0 +1,30 @@ +{ + "author": "aquarious dante workman", + "github_id": "aquariouseworkman", + "name": "SmearGate BOS Fix + PR #1787 Base + LQER Asym + Phased TTT", + "date": "2026-04-27", + "track": "10min_16mb", + "val_bpb": 1.06128, + "seeds": [42], + "seed_results": { + "42": {"val_bpb": 1.06128, "artifact_bytes": 15952086} + }, + "hardware": "8xH100 80GB SXM", + "pytorch_version": "2.9.1+cu128", + "technique_summary": "PR #1787 base (CaseOps + SparseAttnGate + PolarNS + MIN_LR + FusedCE) + SmearGate with BOS document boundary fix + LQER Asymmetric rank-4 correction + Phased TTT (3 phases, score-first)", + "compliance": { + "train_under_600s": true, + "artifact_under_16mb": true, + "eval_under_600s": true, + "smeargate_bos_fix": true, + "score_first_ttt": true + }, + "attribution": { + "pr1787_base": "@nprime06 (PR #1787)", + "caseops": "@romeerp (PR #1729)", + "smeargate_lqer": "@dexhunter (PR #1797)", + "smeargate_bos_audit": "@cocohearts (PR #1797 audit)", + "ttt_framework": "@abaybektursun (PR #549)", + "gptq_sdclip": "@clarkkev (PR #1394)" + } +} diff --git a/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model new file mode 100644 index 0000000000..fffc8bb306 Binary files /dev/null and b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model differ diff --git a/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/train_gpt.py b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/train_gpt.py new file mode 100644 index 0000000000..740f060bb0 --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/train_gpt.py @@ -0,0 +1,3555 @@ +import base64, collections, copy, fcntl, glob, io, lzma, math, os +from pathlib import Path +import random, re, subprocess, sys, time, uuid, numpy as np, sentencepiece as spm, torch, torch.distributed as dist, torch.nn.functional as F +from torch import Tensor, nn +from flash_attn_interface import ( + flash_attn_func as flash_attn_3_func, + flash_attn_varlen_func, +) +from concurrent.futures import ThreadPoolExecutor +import triton +import triton.language as tl +from triton.tools.tensor_descriptor import TensorDescriptor + + +# ===== Fused softcapped cross-entropy (Triton) — training-only path ===== +# Replaces the eager +# logits_softcap = softcap * tanh(logits / softcap) +# F.cross_entropy(logits_softcap.float(), targets, reduction="mean") +# sequence with a single fused kernel that reads logits_proj once, applies +# softcap in-register, and computes (LSE, loss) in one streaming pass. The +# backward kernel mirrors the forward so there's no stored softcapped logits. +# Numerically identical to the eager path up to fp32 accumulation differences. +_FUSED_CE_LIBRARY = "pgsubmission1draft7fusedce" +_FUSED_CE_BLOCK_SIZE = 1024 +_FUSED_CE_NUM_WARPS = 4 + + +@triton.jit +def _softcapped_ce_fwd_kernel( + logits_ptr, losses_ptr, lse_ptr, targets_ptr, + stride_logits_n, stride_logits_v, + n_rows, n_cols, softcap, + block_size: tl.constexpr, +): + row_idx = tl.program_id(0).to(tl.int64) + logits_row_ptr = logits_ptr + row_idx * stride_logits_n + max_val = -float("inf") + sum_exp = 0.0 + A = 2.0 * softcap + inv_C = 2.0 / softcap + for off in range(0, n_cols, block_size): + cols = off + tl.arange(0, block_size) + mask = cols < n_cols + val = tl.load( + logits_row_ptr + cols * stride_logits_v, + mask=mask, other=-float("inf"), + ).to(tl.float32) + z = A * tl.sigmoid(val * inv_C) + z = tl.where(mask, z, -float("inf")) + curr_max = tl.max(z, axis=0) + new_max = tl.maximum(max_val, curr_max) + sum_exp = sum_exp * tl.exp(max_val - new_max) + tl.sum(tl.exp(z - new_max), axis=0) + max_val = new_max + lse = max_val + tl.log(sum_exp) + tl.store(lse_ptr + row_idx, lse) + target = tl.load(targets_ptr + row_idx).to(tl.int32) + target_val = tl.load(logits_row_ptr + target * stride_logits_v).to(tl.float32) + target_z = A * tl.sigmoid(target_val * inv_C) + tl.store(losses_ptr + row_idx, lse - target_z) + + +@triton.jit +def _softcapped_ce_bwd_kernel( + grad_logits_ptr, grad_losses_ptr, lse_ptr, logits_ptr, targets_ptr, + stride_logits_n, stride_logits_v, + stride_grad_n, stride_grad_v, + n_rows, n_cols, softcap, + block_size: tl.constexpr, +): + row_idx = tl.program_id(0).to(tl.int64) + logits_row_ptr = logits_ptr + row_idx * stride_logits_n + grad_row_ptr = grad_logits_ptr + row_idx * stride_grad_n + lse = tl.load(lse_ptr + row_idx) + grad_loss = tl.load(grad_losses_ptr + row_idx).to(tl.float32) + target = tl.load(targets_ptr + row_idx).to(tl.int32) + A = 2.0 * softcap + inv_C = 2.0 / softcap + dz_dx_scale = A * inv_C + for off in range(0, n_cols, block_size): + cols = off + tl.arange(0, block_size) + mask = cols < n_cols + val = tl.load( + logits_row_ptr + cols * stride_logits_v, + mask=mask, other=0.0, + ).to(tl.float32) + sigmoid_u = tl.sigmoid(val * inv_C) + z = A * sigmoid_u + probs = tl.exp(z - lse) + grad_z = grad_loss * (probs - tl.where(cols == target, 1.0, 0.0)) + grad_x = grad_z * (dz_dx_scale * sigmoid_u * (1.0 - sigmoid_u)) + tl.store(grad_row_ptr + cols * stride_grad_v, grad_x, mask=mask) + + +def _validate_softcapped_ce_inputs( + logits: Tensor, targets: Tensor, softcap: float, +) -> tuple[Tensor, Tensor]: + if logits.ndim != 2: + raise ValueError(f"Expected logits.ndim=2, got {logits.ndim}") + if targets.ndim != 1: + raise ValueError(f"Expected targets.ndim=1, got {targets.ndim}") + if logits.shape[0] != targets.shape[0]: + raise ValueError( + f"Expected matching rows, got logits={tuple(logits.shape)} targets={tuple(targets.shape)}" + ) + if not logits.is_cuda or not targets.is_cuda: + raise ValueError("softcapped_cross_entropy requires CUDA tensors") + if softcap <= 0.0: + raise ValueError(f"softcap must be positive, got {softcap}") + if logits.dtype not in (torch.float16, torch.bfloat16, torch.float32): + raise ValueError(f"Unsupported logits dtype: {logits.dtype}") + logits = logits.contiguous() + targets = targets.contiguous() + if targets.dtype != torch.int64: + targets = targets.to(dtype=torch.int64) + return logits, targets + + +@torch.library.custom_op(f"{_FUSED_CE_LIBRARY}::softcapped_ce", mutates_args=()) +def softcapped_ce_op(logits: Tensor, targets: Tensor, softcap: float) -> tuple[Tensor, Tensor]: + logits, targets = _validate_softcapped_ce_inputs(logits, targets, float(softcap)) + n_rows, n_cols = logits.shape + losses = torch.empty((n_rows,), device=logits.device, dtype=torch.float32) + lse = torch.empty((n_rows,), device=logits.device, dtype=torch.float32) + _softcapped_ce_fwd_kernel[(n_rows,)]( + logits, losses, lse, targets, + logits.stride(0), logits.stride(1), + n_rows, n_cols, float(softcap), + block_size=_FUSED_CE_BLOCK_SIZE, num_warps=_FUSED_CE_NUM_WARPS, + ) + return losses, lse + + +@softcapped_ce_op.register_fake +def _(logits: Tensor, targets: Tensor, softcap: float): + if logits.ndim != 2 or targets.ndim != 1: + raise ValueError("softcapped_ce fake impl expects 2D logits and 1D targets") + if logits.shape[0] != targets.shape[0]: + raise ValueError( + f"Expected matching rows, got logits={tuple(logits.shape)} targets={tuple(targets.shape)}" + ) + n_rows = logits.shape[0] + return ( + logits.new_empty((n_rows,), dtype=torch.float32), + logits.new_empty((n_rows,), dtype=torch.float32), + ) + + +@torch.library.custom_op(f"{_FUSED_CE_LIBRARY}::softcapped_ce_backward", mutates_args=()) +def softcapped_ce_backward_op( + logits: Tensor, targets: Tensor, lse: Tensor, grad_losses: Tensor, softcap: float, +) -> Tensor: + logits, targets = _validate_softcapped_ce_inputs(logits, targets, float(softcap)) + lse = lse.contiguous() + grad_losses = grad_losses.contiguous().to(dtype=torch.float32) + if lse.ndim != 1 or grad_losses.ndim != 1: + raise ValueError("Expected 1D lse and grad_losses") + if lse.shape[0] != logits.shape[0] or grad_losses.shape[0] != logits.shape[0]: + raise ValueError( + f"Expected row-aligned lse/grad_losses, got logits={tuple(logits.shape)} " + f"lse={tuple(lse.shape)} grad_losses={tuple(grad_losses.shape)}" + ) + grad_logits = torch.empty_like(logits) + n_rows, n_cols = logits.shape + _softcapped_ce_bwd_kernel[(n_rows,)]( + grad_logits, grad_losses, lse, logits, targets, + logits.stride(0), logits.stride(1), + grad_logits.stride(0), grad_logits.stride(1), + n_rows, n_cols, float(softcap), + block_size=_FUSED_CE_BLOCK_SIZE, num_warps=_FUSED_CE_NUM_WARPS, + ) + return grad_logits + + +@softcapped_ce_backward_op.register_fake +def _(logits: Tensor, targets: Tensor, lse: Tensor, grad_losses: Tensor, softcap: float): + if logits.ndim != 2 or targets.ndim != 1 or lse.ndim != 1 or grad_losses.ndim != 1: + raise ValueError("softcapped_ce_backward fake impl expects 2D logits and 1D row tensors") + if ( + logits.shape[0] != targets.shape[0] + or logits.shape[0] != lse.shape[0] + or logits.shape[0] != grad_losses.shape[0] + ): + raise ValueError("softcapped_ce_backward fake impl expects row-aligned tensors") + return logits.new_empty(logits.shape) + + +def _softcapped_ce_setup_context( + ctx: torch.autograd.function.FunctionCtx, inputs, output, +) -> None: + logits, targets, softcap = inputs + _losses, lse = output + ctx.save_for_backward(logits, targets, lse) + ctx.softcap = float(softcap) + + +def _softcapped_ce_backward( + ctx: torch.autograd.function.FunctionCtx, grad_losses: Tensor, grad_lse: "Tensor | None", +): + del grad_lse + logits, targets, lse = ctx.saved_tensors + grad_logits = torch.ops.pgsubmission1draft7fusedce.softcapped_ce_backward( + logits, targets, lse, grad_losses, ctx.softcap + ) + return grad_logits, None, None + + +softcapped_ce_op.register_autograd( + _softcapped_ce_backward, setup_context=_softcapped_ce_setup_context, +) + + +def softcapped_cross_entropy( + logits: Tensor, targets: Tensor, softcap: float, reduction: str = "mean", +) -> Tensor: + losses, _lse = torch.ops.pgsubmission1draft7fusedce.softcapped_ce( + logits, targets, float(softcap) + ) + if reduction == "none": + return losses + if reduction == "sum": + return losses.sum() + if reduction == "mean": + return losses.mean() + raise ValueError(f"Unsupported reduction={reduction!r}") + + +class Hyperparameters: + data_dir = os.environ.get("DATA_DIR", "./data/") + seed = int(os.environ.get("SEED", 1337)) + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_frac = float(os.environ.get("WARMDOWN_FRAC", 0.75)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786432)) + # Fused softcapped CE (Triton). Training-only — forward_logits eval path still uses + # eager softcap+F.cross_entropy. Default ON since validated as at-worst neutral. + fused_ce_enabled = bool(int(os.environ.get("FUSED_CE_ENABLED", "1"))) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 6e2)) + val_batch_tokens = int(os.environ.get("VAL_BATCH_TOKENS", 524288)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 8192)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 4.0)) + skip_gates_enabled = bool(int(os.environ.get("SKIP_GATES_ENABLED", "1"))) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 3e1)) + rope_base = float(os.environ.get("ROPE_BASE", 1e4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + rope_train_seq_len = int(os.environ.get("ROPE_TRAIN_SEQ_LEN", 2048)) + rope_yarn = bool(int(os.environ.get("ROPE_YARN", "0"))) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 5.0)) + num_loops = int(os.environ.get("NUM_LOOPS", 2)) + loop_start = int(os.environ.get("LOOP_START", 3)) + loop_end = int(os.environ.get("LOOP_END", 5)) + enable_looping_at = float(os.environ.get("ENABLE_LOOPING_AT", 0.35)) + parallel_start_layer = int(os.environ.get("PARALLEL_START_LAYER", 8)) + parallel_final_lane = os.environ.get("PARALLEL_FINAL_LANE", "mean") + min_lr = float(os.environ.get("MIN_LR", 0.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.026)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.97)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float( + os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92) + ) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + muon_row_normalize = bool(int(os.environ.get("MUON_ROW_NORMALIZE", "1"))) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-08)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + adam_wd = float(os.environ.get("ADAM_WD", 0.02)) + muon_wd = float(os.environ.get("MUON_WD", 0.095)) + embed_wd = float(os.environ.get("EMBED_WD", 0.085)) + ema_decay = float(os.environ.get("EMA_DECAY", 0.9965)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lora_rank = int(os.environ.get("TTT_LORA_RANK", 96)) + ttt_lora_lr = float(os.environ.get("TTT_LORA_LR", 0.0001)) + ttt_chunk_size = int(os.environ.get("TTT_CHUNK_SIZE", 48)) + ttt_eval_seq_len = int(os.environ.get("TTT_EVAL_SEQ_LEN", 2048)) + ttt_batch_size = int(os.environ.get("TTT_BATCH_SIZE", 64)) + ttt_grad_steps = int(os.environ.get("TTT_GRAD_STEPS", 1)) + ttt_weight_decay = float(os.environ.get("TTT_WEIGHT_DECAY", 1.0)) + ttt_beta1 = float(os.environ.get("TTT_BETA1", 0)) + ttt_beta2 = float(os.environ.get("TTT_BETA2", 0.999)) + ttt_k_lora = bool(int(os.environ.get("TTT_K_LORA", "1"))) + ttt_mlp_lora = bool(int(os.environ.get("TTT_MLP_LORA", "1"))) + ttt_o_lora = bool(int(os.environ.get("TTT_O_LORA", "1"))) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adam") + ttt_eval_batches = os.environ.get("TTT_EVAL_BATCHES", "") + val_doc_fraction = float(os.environ.get("VAL_DOC_FRACTION", 1.0)) + compressor = os.environ.get("COMPRESSOR", "brotli") + gptq_calibration_batches = int(os.environ.get("GPTQ_CALIBRATION_BATCHES", 16)) + gptq_reserve_seconds = float(os.environ.get("GPTQ_RESERVE_SECONDS", 4.0)) + phased_ttt_prefix_docs = int(os.environ.get("PHASED_TTT_PREFIX_DOCS", 2000)) + phased_ttt_num_phases = int(os.environ.get("PHASED_TTT_NUM_PHASES", 1)) + global_ttt_lr = float(os.environ.get("GLOBAL_TTT_LR", 0.001)) + global_ttt_momentum = float(os.environ.get("GLOBAL_TTT_MOMENTUM", 0.9)) + global_ttt_epochs = int(os.environ.get("GLOBAL_TTT_EPOCHS", 1)) + global_ttt_chunk_tokens = int(os.environ.get("GLOBAL_TTT_CHUNK_TOKENS", 32768)) + global_ttt_batch_seqs = int(os.environ.get("GLOBAL_TTT_BATCH_SEQS", 32)) + global_ttt_warmup_start_lr = float(os.environ.get("GLOBAL_TTT_WARMUP_START_LR", 0.0)) + global_ttt_warmup_chunks = int(os.environ.get("GLOBAL_TTT_WARMUP_CHUNKS", 0)) + global_ttt_grad_clip = float(os.environ.get("GLOBAL_TTT_GRAD_CLIP", 1.0)) + global_ttt_respect_doc_boundaries = bool(int(os.environ.get("GLOBAL_TTT_RESPECT_DOC_BOUNDARIES", "1"))) + matrix_bits = int(os.environ.get("MATRIX_BITS", 6)) + embed_bits = int(os.environ.get("EMBED_BITS", 8)) + matrix_clip_sigmas = float(os.environ.get("MATRIX_CLIP_SIGMAS", 12.85)) + embed_clip_sigmas = float(os.environ.get("EMBED_CLIP_SIGMAS", 2e1)) + mlp_clip_sigmas = float(os.environ.get("MLP_CLIP_SIGMAS", 10.0)) + attn_clip_sigmas = float(os.environ.get("ATTN_CLIP_SIGMAS", 13.0)) + # AttnOutGate (per-head multiplicative output gate, PR #1667 MarioPaerle). + # Zero-init weight: 2*sigmoid(0)=1 -> transparent at start. Source defaults to + # block input x ('proj'); 'q' uses raw Q projection output. + attn_out_gate_enabled = bool(int(os.environ.get("ATTN_OUT_GATE_ENABLED", "0"))) + attn_out_gate_src = os.environ.get("ATTN_OUT_GATE_SRC", "proj") + # SmearGate (input-dependent forward-1 token smear, modded-nanogpt @classiclarryd + # via PR #1667). x_t <- x_t + lam * sigmoid(W*x_t[:gate_window]) * x_{t-1}. + # lam=0 + W=0 -> transparent at init. + smear_gate_enabled = bool(int(os.environ.get("SMEAR_GATE_ENABLED", "0"))) + # Window: first GATE_WINDOW dims of the source feed the gate projection. + gate_window = int(os.environ.get("GATE_WINDOW", 12)) + # Gated Attention (Qwen, NeurIPS 2025 Best Paper, arXiv:2505.06708; + # qiuzh20/gated_attention). Per-head sigmoid gate on SDPA output, BEFORE + # out_proj. Gate input = full block input x (paper's headwise G1 variant + # driven from hidden_states). W_g shape (num_heads, dim), plain sigmoid. + # Near-zero init gives g~0.5 at step 0 (half attention output); per-block + # attn_scale (init 1.0) compensates during training. Name contains + # "attn_gate" so CONTROL_TENSOR_NAME_PATTERNS routes it to scalar AdamW. + gated_attn_enabled = bool(int(os.environ.get("GATED_ATTN_ENABLED", "0"))) + gated_attn_init_std = float(os.environ.get("GATED_ATTN_INIT_STD", 0.01)) + # Dedicated int8-per-row quantization for `attn_gate_w` tensors. These are + # small ((num_heads, dim) = (8, 512) = 4096 params) and bypass GPTQ via the + # numel<=65536 passthrough branch -> stored as fp16 (8 KB/layer, ~65 KB total + # compressed). int8-per-row cuts the raw tensor in half with negligible BPB + # impact: scales per head (8 values), symmetric quant over [-127, 127]. + # No Hessian needed (gate weights not in collect_hessians()). + gated_attn_quant_gate = bool(int(os.environ.get("GATED_ATTN_QUANT_GATE", "0"))) + # Sparse Attention Gate (modded-nanogpt-style). Keeps dense SDPA and only + # swaps the output-gate input to the first GATE_WINDOW residual dims. + # W_g: (num_heads, gate_window) = (8, 12) = 96 params/layer (~44K total), + # vs dense GatedAttn's (8, 512) = 4K/layer (~44K diff). Name "attn_gate_w" + # is shared so quant routing and int8 gate passthrough Just Work. Gate + # passthrough int8 still applies via GATED_ATTN_QUANT_GATE=1. + # Mutually exclusive with ATTN_OUT_GATE_ENABLED and GATED_ATTN_ENABLED. + sparse_attn_gate_enabled = bool(int(os.environ.get("SPARSE_ATTN_GATE_ENABLED", "0"))) + sparse_attn_gate_init_std = float(os.environ.get("SPARSE_ATTN_GATE_INIT_STD", 0.0)) + sparse_attn_gate_scale = float(os.environ.get("SPARSE_ATTN_GATE_SCALE", 1.0)) + # LQER asymmetric rank-k correction on top-K quant-error tensors (PR #1530 v2 port). + # Computes SVD of E = W_fp - W_quant, packs top-r A,B as INT2/INT4 (asym) or INTk (sym). + lqer_enabled = bool(int(os.environ.get("LQER_ENABLED", "1"))) + lqer_rank = int(os.environ.get("LQER_RANK", 4)) + lqer_top_k = int(os.environ.get("LQER_TOP_K", 3)) + lqer_factor_bits = int(os.environ.get("LQER_FACTOR_BITS", 4)) + lqer_asym_enabled = bool(int(os.environ.get("LQER_ASYM_ENABLED", "1"))) + lqer_asym_group = int(os.environ.get("LQER_ASYM_GROUP", "64")) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + is_main_process = rank == 0 + grad_accum_steps = 8 // world_size + # CaseOps integration: optional override of dataset root + tokenizer path. + # When CASEOPS_ENABLED=1, the wrapper loads a per-token byte sidecar + # (fineweb_val_bytes_*.bin, identical shard layout to val_*.bin) and uses + # it as the canonical raw-byte budget for BPB accounting. The sidecar + # REPLACES the build_sentencepiece_luts byte-counting path entirely. + caseops_enabled = bool(int(os.environ.get("CASEOPS_ENABLED", "0"))) + _default_caseops_data = os.path.join( + data_dir, + "datasets", + "fineweb10B_sp8192_caseops", + "datasets", + "datasets", + "fineweb10B_sp8192_lossless_caps_caseops_v1_reserved", + ) + _default_caseops_tok = os.path.join( + data_dir, + "datasets", + "fineweb10B_sp8192_caseops", + "datasets", + "tokenizers", + "fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model", + ) + if caseops_enabled: + datasets_dir = os.environ.get("DATA_PATH", _default_caseops_data) + tokenizer_path = os.environ.get("TOKENIZER_PATH", _default_caseops_tok) + else: + datasets_dir = os.environ.get( + "DATA_PATH", + os.path.join(data_dir, "datasets", f"fineweb10B_sp{vocab_size}"), + ) + tokenizer_path = os.environ.get( + "TOKENIZER_PATH", + os.path.join(data_dir, "tokenizers", f"fineweb_{vocab_size}_bpe.model"), + ) + train_files = os.path.join(datasets_dir, "fineweb_train_*.bin") + val_files = os.path.join(datasets_dir, "fineweb_val_*.bin") + val_bytes_files = os.path.join(datasets_dir, "fineweb_val_bytes_*.bin") + artifact_dir = os.environ.get("ARTIFACT_DIR", "") + logfile = ( + os.path.join(artifact_dir, f"{run_id}.txt") + if artifact_dir + else f"logs/{run_id}.txt" + ) + model_path = ( + os.path.join(artifact_dir, "final_model.pt") + if artifact_dir + else "final_model.pt" + ) + quantized_model_path = ( + os.path.join(artifact_dir, "final_model.int6.ptz") + if artifact_dir + else "final_model.int6.ptz" + ) + + +_logger_hparams = None + + +def set_logging_hparams(h): + global _logger_hparams + _logger_hparams = h + + +def log(msg, console=True): + if _logger_hparams is None: + print(msg) + return + if _logger_hparams.is_main_process: + if console: + print(msg) + if _logger_hparams.logfile is not None: + with open(_logger_hparams.logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + +class ValidationData: + def __init__(self, h, device): + self.sp = spm.SentencePieceProcessor(model_file=h.tokenizer_path) + if int(self.sp.vocab_size()) != h.vocab_size: + raise ValueError( + f"VOCAB_SIZE={h.vocab_size} does not match tokenizer vocab_size={int(self.sp.vocab_size())}" + ) + self.val_tokens = load_validation_tokens(h.val_files, h.eval_seq_len) + ( + self.base_bytes_lut, + self.has_leading_space_lut, + self.is_boundary_token_lut, + ) = build_sentencepiece_luts(self.sp, h.vocab_size, device) + # CaseOps: when enabled, load per-token byte sidecar and stash it as a + # CPU tensor aligned 1:1 with self.val_tokens. eval_val/eval_val_ttt + # branches use this as the canonical raw-byte budget per token. + self.caseops_enabled = bool(getattr(h, "caseops_enabled", False)) + self.val_bytes = None + if self.caseops_enabled: + self.val_bytes = load_validation_byte_sidecar( + h.val_bytes_files, h.eval_seq_len, self.val_tokens.numel() + ) + + +def build_sentencepiece_luts(sp, vocab_size, device): + sp_vocab_size = int(sp.vocab_size()) + assert ( + sp.piece_to_id("▁") != sp.unk_id() + ), "Tokenizer must have '▁' (space) as its own token for correct BPB byte counting" + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern, seq_len): + # Filter out CaseOps byte sidecar shards which share the val_*.bin glob. + files = [ + Path(p) + for p in sorted(glob.glob(pattern)) + if "_bytes_" not in Path(p).name + ] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = (tokens.numel() - 1) // seq_len * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def load_validation_byte_sidecar(pattern, seq_len, expected_len): + """Load CaseOps per-token byte sidecar(s). Same shard layout as token shards + (256 int32 header + uint16 array). Each entry = canonical raw-text byte + budget for that token in the corresponding val shard. Returns a CPU + int16 tensor sliced to match expected_len (i.e. val_tokens length).""" + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No byte sidecar files for pattern: {pattern}") + shards = [load_data_shard(file) for file in files] + # load_data_shard returns uint16 — that's exactly what the sidecar stores. + bytes_full = torch.cat(shards).contiguous() + if bytes_full.numel() < expected_len: + raise ValueError( + f"Byte sidecar too short: {bytes_full.numel()} < val_tokens {expected_len}" + ) + return bytes_full[:expected_len].to(torch.int32) + + +def load_data_shard(file): + header_bytes = 256 * np.dtype(" 0: + pos = start + while pos < end: + seg_starts.append(pos) + pos += max_doc_len + else: + seg_starts.append(start) + boundaries = seg_starts + [total_len] + padded_len = get_next_multiple_of_n(len(boundaries), bucket_size) + cu = torch.full((padded_len,), total_len, dtype=torch.int32, device=device) + cu[: len(boundaries)] = torch.tensor(boundaries, dtype=torch.int32, device=device) + seg_ends = seg_starts[1:] + [total_len] + max_seqlen = max(end - start for start, end in zip(seg_starts, seg_ends)) + return cu, max_seqlen + +class DocumentPackingLoader: + _shard_pool = ThreadPoolExecutor(1) + + def __init__(self, h, device, cu_bucket_size=64): + self.rank = h.rank + self.world_size = h.world_size + self.device = device + self.cu_bucket_size = cu_bucket_size + self.max_seq_len = h.train_seq_len + all_files = [Path(p) for p in sorted(glob.glob(h.train_files))] + if not all_files: + raise FileNotFoundError(f"No files found for pattern: {h.train_files}") + self.files = all_files + self.file_iter = iter(self.files) + self._init_shard(load_data_shard(next(self.file_iter))) + self._next_shard = self._submit_next_shard() + self._batch_pool = ThreadPoolExecutor(1) + self._next_batch = None + + def _init_shard(self, tokens): + global BOS_ID + self.tokens = tokens + self.shard_size = tokens.numel() + if BOS_ID is None: + BOS_ID = 1 + self.bos_idx = ( + (tokens == BOS_ID).nonzero(as_tuple=True)[0].to(torch.int64).cpu().numpy() + ) + self.cursor = int(self.bos_idx[0]) + + def _submit_next_shard(self): + try: + path = next(self.file_iter) + return self._shard_pool.submit(load_data_shard, path) + except StopIteration: + return None + + def _advance_shard(self): + if self._next_shard is None: + self.file_iter = iter(self.files) + self._next_shard = self._shard_pool.submit( + load_data_shard, next(self.file_iter) + ) + self._init_shard(self._next_shard.result()) + self._next_shard = self._submit_next_shard() + + def _local_doc_starts(self, local_start, total_len): + lo = np.searchsorted(self.bos_idx, local_start, side="left") + hi = np.searchsorted(self.bos_idx, local_start + total_len, side="left") + return (self.bos_idx[lo:hi] - local_start).tolist() + + def _prepare_batch(self, num_tokens_local, max_seq_len): + per_rank_span = num_tokens_local + 1 + global_span = per_rank_span * self.world_size + while self.cursor + global_span > self.shard_size: + self._advance_shard() + local_start = self.cursor + self.rank * per_rank_span + buf = self.tokens[local_start : local_start + per_rank_span] + inputs = buf[:-1].to(dtype=torch.int64).pin_memory() + targets = buf[1:].to(dtype=torch.int64).pin_memory() + starts = self._local_doc_starts(local_start, inputs.numel()) + cu_seqlens, max_seqlen = _build_cu_seqlens( + starts, inputs.numel(), inputs.device, max_seq_len, self.cu_bucket_size + ) + cu_seqlens = cu_seqlens.pin_memory() + self.cursor += global_span + return inputs, targets, cu_seqlens, max_seqlen + + def next_batch(self, global_tokens, grad_accum_steps): + num_tokens_local = global_tokens // (self.world_size * grad_accum_steps) + if self._next_batch is not None: + inputs, targets, cu_seqlens, max_seqlen = self._next_batch.result() + else: + inputs, targets, cu_seqlens, max_seqlen = self._prepare_batch( + num_tokens_local, self.max_seq_len + ) + self._next_batch = self._batch_pool.submit( + self._prepare_batch, num_tokens_local, self.max_seq_len + ) + return ( + inputs[None].to(self.device, non_blocking=True), + targets[None].to(self.device, non_blocking=True), + cu_seqlens.to(self.device, non_blocking=True), + max_seqlen, + ) + + +class ShuffledSequenceLoader: + def __init__(self, h, device): + self.world_size = h.world_size + self.seq_len = h.train_seq_len + self.device = device + all_files = [Path(p) for p in sorted(glob.glob(h.train_files))] + if not all_files: + raise FileNotFoundError(f"No files found for pattern: {h.train_files}") + self.files = all_files[h.rank :: h.world_size] + self.rng = np.random.Generator(np.random.PCG64(h.rank)) + self.num_tokens = [_read_num_tokens(f) for f in self.files] + self.start_inds = [[] for _ in self.files] + for si in range(len(self.files)): + self._reset_shard(si) + + def _reset_shard(self, si): + max_phase = min( + self.seq_len - 1, max(0, self.num_tokens[si] - self.seq_len - 1) + ) + phase = int(self.rng.integers(max_phase + 1)) if max_phase > 0 else 0 + num_sequences = (self.num_tokens[si] - 1 - phase) // self.seq_len + sequence_order = self.rng.permutation(num_sequences) + self.start_inds[si] = (phase + sequence_order * self.seq_len).tolist() + + def next_batch(self, global_tokens, grad_accum_steps): + device_tokens = global_tokens // (self.world_size * grad_accum_steps) + device_batch_size = device_tokens // self.seq_len + remaining = np.array([len(s) for s in self.start_inds], dtype=np.float64) + x = torch.empty((device_batch_size, self.seq_len), dtype=torch.int64) + y = torch.empty((device_batch_size, self.seq_len), dtype=torch.int64) + for bi in range(device_batch_size): + total = remaining.sum() + if total <= 0: + for si in range(len(self.files)): + self._reset_shard(si) + remaining = np.array( + [len(s) for s in self.start_inds], dtype=np.float64 + ) + total = remaining.sum() + probs = remaining / total + si = int(self.rng.choice(len(self.files), p=probs)) + start_ind = self.start_inds[si].pop() + remaining[si] -= 1 + mm = _get_shard_memmap(self.files[si]) + window = torch.as_tensor( + np.array(mm[start_ind : start_ind + self.seq_len + 1], dtype=np.int64) + ) + x[bi] = window[:-1] + y[bi] = window[1:] + return x.to(self.device, non_blocking=True), y.to( + self.device, non_blocking=True + ) + + +class RMSNorm(nn.Module): + def __init__(self, eps=None): + super().__init__() + self.eps = eps + + def forward(self, x): + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + def forward(self, x): + w = self.weight.to(x.dtype) + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + + +@triton.jit +def linear_leaky_relu_square_kernel( + a_desc, + b_desc, + c_desc, + aux_desc, + M, + N, + K, + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + NUM_SMS: tl.constexpr, + FORWARD: tl.constexpr, +): + dtype = tl.bfloat16 + start_pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(M, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + k_tiles = tl.cdiv(K, BLOCK_SIZE_K) + num_tiles = num_pid_m * num_pid_n + tile_id_c = start_pid - NUM_SMS + for tile_id in tl.range(start_pid, num_tiles, NUM_SMS, flatten=True): + pid_m = tile_id // num_pid_n + pid_n = tile_id % num_pid_n + offs_am = pid_m * BLOCK_SIZE_M + offs_bn = pid_n * BLOCK_SIZE_N + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for ki in range(k_tiles): + offs_k = ki * BLOCK_SIZE_K + a = a_desc.load([offs_am, offs_k]) + b = b_desc.load([offs_bn, offs_k]) + accumulator = tl.dot(a, b.T, accumulator) + tile_id_c += NUM_SMS + offs_am_c = offs_am + offs_bn_c = offs_bn + acc = tl.reshape(accumulator, (BLOCK_SIZE_M, 2, BLOCK_SIZE_N // 2)) + acc = tl.permute(acc, (0, 2, 1)) + acc0, acc1 = tl.split(acc) + c0 = acc0.to(dtype) + c1 = acc1.to(dtype) + if not FORWARD: + pre0 = aux_desc.load([offs_am_c, offs_bn_c]) + pre1 = aux_desc.load([offs_am_c, offs_bn_c + BLOCK_SIZE_N // 2]) + c0 = c0 * tl.where(pre0 > 0, 2.0 * pre0, 0.5 * pre0) + c1 = c1 * tl.where(pre1 > 0, 2.0 * pre1, 0.5 * pre1) + c_desc.store([offs_am_c, offs_bn_c], c0) + c_desc.store([offs_am_c, offs_bn_c + BLOCK_SIZE_N // 2], c1) + if FORWARD: + aux0 = tl.where(c0 > 0, c0, 0.5 * c0) + aux1 = tl.where(c1 > 0, c1, 0.5 * c1) + aux_desc.store([offs_am_c, offs_bn_c], aux0 * aux0) + aux_desc.store([offs_am_c, offs_bn_c + BLOCK_SIZE_N // 2], aux1 * aux1) + + +def linear_leaky_relu_square(a, b, aux=None): + M, K = a.shape + N, K2 = b.shape + assert K == K2 + c = torch.empty((M, N), device=a.device, dtype=a.dtype) + forward = aux is None + if aux is None: + aux = torch.empty((M, N), device=a.device, dtype=a.dtype) + num_sms = torch.cuda.get_device_properties(a.device).multi_processor_count + BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K = 128, 256, 64 + num_stages = 4 if forward else 3 + a_desc = TensorDescriptor.from_tensor(a, [BLOCK_SIZE_M, BLOCK_SIZE_K]) + b_desc = TensorDescriptor.from_tensor(b, [BLOCK_SIZE_N, BLOCK_SIZE_K]) + c_desc = TensorDescriptor.from_tensor(c, [BLOCK_SIZE_M, BLOCK_SIZE_N // 2]) + aux_desc = TensorDescriptor.from_tensor(aux, [BLOCK_SIZE_M, BLOCK_SIZE_N // 2]) + grid = lambda _meta: ( + min(num_sms, triton.cdiv(M, BLOCK_SIZE_M) * triton.cdiv(N, BLOCK_SIZE_N)), + ) + linear_leaky_relu_square_kernel[grid]( + a_desc, + b_desc, + c_desc, + aux_desc, + M, + N, + K, + BLOCK_SIZE_M=BLOCK_SIZE_M, + BLOCK_SIZE_N=BLOCK_SIZE_N, + BLOCK_SIZE_K=BLOCK_SIZE_K, + NUM_SMS=num_sms, + FORWARD=forward, + num_stages=num_stages, + num_warps=8, + ) + if forward: + return c, aux + return c + + +class FusedLinearLeakyReLUSquareFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, x, w1, w2): + x_flat = x.reshape(-1, x.shape[-1]) + pre, post = linear_leaky_relu_square(x_flat, w1) + out = F.linear(post, w2) + ctx.save_for_backward(x, w1, w2, pre, post) + return out.view(*x.shape[:-1], out.shape[-1]) + + @staticmethod + def backward(ctx, grad_output): + x, w1, w2, pre, post = ctx.saved_tensors + x_flat = x.reshape(-1, x.shape[-1]) + grad_output_flat = grad_output.reshape(-1, grad_output.shape[-1]) + dw2 = grad_output_flat.T @ post + dpre = linear_leaky_relu_square(grad_output_flat, w2.T.contiguous(), aux=pre) + dw1 = dpre.T @ x_flat + dx = dpre @ w1 + return dx.view_as(x), dw1, dw2 + + +FusedLeakyReLUSquareMLP = FusedLinearLeakyReLUSquareFunction.apply + + +class Rotary(nn.Module): + def __init__(self, dim, base=1e4, train_seq_len=1024, rope_dims=0, yarn=True): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.yarn = yarn + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / base ** ( + torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims + ) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached = None + self._sin_cached = None + + def forward(self, seq_len, device, dtype): + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached < seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if self.yarn and seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * scale ** (rd / (rd - 2)) + inv_freq = 1.0 / new_base ** ( + torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd + ) + else: + inv_freq = self.inv_freq.float().to(device) + t = torch.arange(seq_len, device=device, dtype=torch.float32) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached[:, :seq_len].to(dtype=dtype), self._sin_cached[:, :seq_len].to(dtype=dtype) + + +def apply_rotary_emb(x, cos, sin, rope_dims=0): + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * -sin + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * -sin + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__( + self, dim, num_heads, num_kv_heads, rope_base, qk_gain_init, train_seq_len, yarn=True, + attn_out_gate=False, attn_out_gate_src="proj", gate_window=12, + gated_attn=False, gated_attn_init_std=0.01, + sparse_attn_gate=False, sparse_attn_gate_init_std=0.0, sparse_attn_gate_scale=1.0, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + if int(attn_out_gate) + int(gated_attn) + int(sparse_attn_gate) > 1: + raise ValueError( + "attn_out_gate, gated_attn, and sparse_attn_gate are mutually exclusive" + ) + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + self.q_gain = nn.Parameter( + torch.full((num_heads,), qk_gain_init, dtype=torch.float32) + ) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=train_seq_len, yarn=yarn) + self.use_xsa = False + # AttnOutGate (PR #1667 MarioPaerle): per-head multiplicative gate on attention + # output. CastedLinear so restore_fp32_params casts back to fp32 for GPTQ. + # _zero_init -> 2*sigmoid(0)=1 -> transparent at init. + self.attn_out_gate = attn_out_gate + self.attn_out_gate_src = attn_out_gate_src + self.gate_window = gate_window + if attn_out_gate: + self.attn_gate_proj = CastedLinear(gate_window, num_heads, bias=False) + self.attn_gate_proj._zero_init = True + # Gated Attention (arXiv:2505.06708, Qwen, NeurIPS 2025). Per-head sigmoid + # gate on SDPA output, BEFORE out_proj. Gate projection W_g: (num_heads, dim). + # Name "attn_gate_w" contains "attn_gate" substring so it matches + # CONTROL_TENSOR_NAME_PATTERNS and routes to the scalar AdamW group. + # fp32 Parameter -> restore_fp32_params path covers it via the ndim<2 OR + # name-pattern check (name matches "attn_gate"). Cast to x.dtype on use. + self.gated_attn = gated_attn + if gated_attn: + W = torch.empty(num_heads, dim, dtype=torch.float32) + nn.init.normal_(W, mean=0.0, std=gated_attn_init_std) + self.attn_gate_w = nn.Parameter(W) + # Sparse attention head-output gate (modded-nanogpt style). Keeps dense SDPA + # and only narrows the gate input to the first gate_window residual dims. + # W_g: (num_heads, gate_window). y_{t,h} <- sigmoid(scale * W_g_h @ x_t[:gate_window]) * y_{t,h}. + # Shares attn_gate_w name with dense GatedAttn so the quant routing + # (CONTROL_TENSOR_NAME_PATTERNS / attn_gate_w int8 passthrough) is unchanged. + self.sparse_attn_gate = sparse_attn_gate + self.sparse_attn_gate_scale = sparse_attn_gate_scale + if sparse_attn_gate: + W = torch.empty(num_heads, gate_window, dtype=torch.float32) + if sparse_attn_gate_init_std > 0: + nn.init.normal_(W, mean=0.0, std=sparse_attn_gate_init_std) + else: + nn.init.zeros_(W) + self.attn_gate_w = nn.Parameter(W) + + def _xsa_efficient(self, y, v): + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + + def forward(self, x, q_w, k_w, v_w, out_w, cu_seqlens=None, max_seqlen=0): + bsz, seqlen, dim = x.shape + # q_raw kept around as a tap point for attn_out_gate_src='q' (post-projection, + # pre-reshape, pre-RoPE). + q_raw = F.linear(x, q_w.to(x.dtype)) + q = q_raw.reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if cu_seqlens is not None: + y = flash_attn_varlen_func( + q[0], + k[0], + v[0], + cu_seqlens_q=cu_seqlens, + cu_seqlens_k=cu_seqlens, + max_seqlen_q=max_seqlen, + max_seqlen_k=max_seqlen, + causal=True, + window_size=(-1, -1), + )[None] + else: + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + # AttnOutGate inlined (PR #1667). Inline + .contiguous() barrier so torch.compile + # fullgraph=True is happy (this avoids the @torch.compiler.disable trap that + # crashed gates v3). Per-head gate on (B,T,H,D) tensor: g shape [B,T,H], broadcast + # over D via [..., None]. zero-init weight -> 2*sigmoid(0)=1 -> transparent. + if self.attn_out_gate: + gate_src = q_raw if self.attn_out_gate_src == "q" else x + gate_in = gate_src[..., : self.gate_window].contiguous() + g = 2.0 * torch.sigmoid(self.attn_gate_proj(gate_in)) + y = y * g[..., None] + # Gated Attention (arXiv:2505.06708 G1). Inline + .contiguous() barrier so + # torch.compile fullgraph=True is happy. Per-head gate on (B,T,H,D): g shape + # [B,T,H], broadcast over D via [..., None]. Paper: g = sigmoid(x @ W_g.T) + # where W_g: (H, dim). .to(x.dtype) on fp32 param before broadcast with bf16. + if self.gated_attn: + x_c = x.contiguous() + g = torch.sigmoid(F.linear(x_c, self.attn_gate_w.to(x.dtype))) + y = y * g[..., None] + # Sparse head-output gate: narrower (gate_window) input, same shape g as GatedAttn. + if self.sparse_attn_gate: + gate_in = x[..., : self.gate_window].contiguous() + g = torch.sigmoid( + self.sparse_attn_gate_scale + * F.linear(gate_in, self.attn_gate_w.to(x.dtype)) + ) + y = y * g[..., None] + y = y.reshape(bsz, seqlen, dim) + self._last_proj_input = y.detach() if getattr(self, "_calib", False) else None + return F.linear(y, out_w.to(x.dtype)) + + +class MLP(nn.Module): + def __init__(self, dim, mlp_mult): + super().__init__() + self.use_fused = True + + def forward(self, x, up_w, down_w): + if self.training and self.use_fused: + return FusedLeakyReLUSquareMLP(x, up_w.to(x.dtype), down_w.to(x.dtype)) + hidden = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5).square() + self._last_down_input = hidden.detach() if getattr(self, "_calib", False) else None + return F.linear(hidden, down_w.to(x.dtype)) + + +class Block(nn.Module): + def __init__( + self, + dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + train_seq_len, + layer_idx=0, + ln_scale=False, + yarn=True, + attn_out_gate=False, + attn_out_gate_src="proj", + gate_window=12, + gated_attn=False, + gated_attn_init_std=0.01, + sparse_attn_gate=False, + sparse_attn_gate_init_std=0.0, + sparse_attn_gate_scale=1.0, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention( + dim, num_heads, num_kv_heads, rope_base, qk_gain_init, train_seq_len, yarn=yarn, + attn_out_gate=attn_out_gate, attn_out_gate_src=attn_out_gate_src, gate_window=gate_window, + gated_attn=gated_attn, gated_attn_init_std=gated_attn_init_std, + sparse_attn_gate=sparse_attn_gate, + sparse_attn_gate_init_std=sparse_attn_gate_init_std, + sparse_attn_gate_scale=sparse_attn_gate_scale, + ) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter( + torch.stack((torch.ones(dim), torch.zeros(dim))).float() + ) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + + def forward(self, x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=None, max_seqlen=0): + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn( + self.attn_norm(x_in) * self.ln_scale_factor, + q_w, k_w, v_w, out_w, + cu_seqlens=cu_seqlens, + max_seqlen=max_seqlen, + ) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[ + None, None, : + ] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + return x_out + +class GPT(nn.Module): + def __init__(self, h): + super().__init__() + if h.logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {h.logit_softcap}") + self.tie_embeddings = h.tie_embeddings + self.tied_embed_init_std = h.tied_embed_init_std + self.logit_softcap = h.logit_softcap + self.fused_ce_enabled = bool(h.fused_ce_enabled) + self.tok_emb = nn.Embedding(h.vocab_size, h.model_dim) + self.num_layers = h.num_layers + head_dim = h.model_dim // h.num_heads + kv_dim = h.num_kv_heads * head_dim + hidden_dim = int(h.mlp_mult * h.model_dim) + self.qo_bank = nn.Parameter(torch.empty(2 * h.num_layers, h.model_dim, h.model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * h.num_layers, kv_dim, h.model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(h.num_layers, hidden_dim, h.model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(h.num_layers, h.model_dim, hidden_dim)) + self.num_encoder_layers = h.num_layers // 2 + self.num_decoder_layers = h.num_layers - self.num_encoder_layers + self.blocks = nn.ModuleList( + [ + Block( + h.model_dim, + h.num_heads, + h.num_kv_heads, + h.mlp_mult, + h.rope_base, + h.qk_gain_init, + h.train_seq_len, + layer_idx=i, + ln_scale=h.ln_scale, + yarn=h.rope_yarn, + attn_out_gate=h.attn_out_gate_enabled, + attn_out_gate_src=h.attn_out_gate_src, + gate_window=h.gate_window, + gated_attn=h.gated_attn_enabled, + gated_attn_init_std=h.gated_attn_init_std, + sparse_attn_gate=h.sparse_attn_gate_enabled, + sparse_attn_gate_init_std=h.sparse_attn_gate_init_std, + sparse_attn_gate_scale=h.sparse_attn_gate_scale, + ) + for i in range(h.num_layers) + ] + ) + if h.rope_dims > 0: + head_dim = h.model_dim // h.num_heads + for block in self.blocks: + block.attn.rope_dims = h.rope_dims + block.attn.rotary = Rotary( + head_dim, + base=h.rope_base, + train_seq_len=h.train_seq_len, + rope_dims=h.rope_dims, + yarn=h.rope_yarn, + ) + self.final_norm = RMSNorm() + self.lm_head = ( + None + if h.tie_embeddings + else CastedLinear(h.model_dim, h.vocab_size, bias=False) + ) + if self.lm_head is not None: + self.lm_head._zero_init = True + if h.xsa_last_n > 0: + for i in range(max(0, h.num_layers - h.xsa_last_n), h.num_layers): + self.blocks[i].attn.use_xsa = True + self.looping_active = False + if h.num_loops > 0: + loop_seg = list(range(h.loop_start, h.loop_end + 1)) + all_indices = list(range(h.loop_start)) + for _ in range(h.num_loops + 1): + all_indices.extend(loop_seg) + all_indices.extend(range(h.loop_end + 1, h.num_layers)) + num_enc = len(all_indices) // 2 + self.encoder_indices = all_indices[:num_enc] + self.decoder_indices = all_indices[num_enc:] + else: + self.encoder_indices = list(range(self.num_encoder_layers)) + self.decoder_indices = list(range(self.num_encoder_layers, h.num_layers)) + self.num_skip_weights = min( + len(self.encoder_indices), len(self.decoder_indices) + ) + self.skip_weights = nn.Parameter( + torch.ones(self.num_skip_weights, h.model_dim, dtype=torch.float32) + ) + self.skip_gates = ( + nn.Parameter( + torch.zeros(self.num_skip_weights, h.model_dim, dtype=torch.float32) + ) + if h.skip_gates_enabled + else None + ) + self.parallel_start_layer = h.parallel_start_layer + self.parallel_final_lane = h.parallel_final_lane.lower() + self.parallel_post_lambdas = nn.Parameter( + torch.ones(h.num_layers, 2, 2, dtype=torch.float32) + ) + self.parallel_resid_lambdas = nn.Parameter( + torch.full((h.num_layers, 2), 1.1, dtype=torch.float32) + ) + # SmearGate (PR #1667 / modded-nanogpt @classiclarryd): + # x_t <- x_t + lam * sigmoid(W * x_t[:gate_window]) * x_{t-1}. + # Per-token forward-1 smear of the embedding lane. W zero-init + lam=0 -> + # transparent at init. Uses CastedLinear so restore_fp32_params handles dtype. + self.smear_gate_enabled = h.smear_gate_enabled + if self.smear_gate_enabled: + self.smear_window = h.gate_window + self.smear_gate = CastedLinear(self.smear_window, 1, bias=False) + self.smear_gate._zero_init = True + self.smear_lambda = nn.Parameter(torch.zeros(1, dtype=torch.float32)) + self._init_weights() + + def _init_weights(self): + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) + nn.init.zeros_(self.qo_bank.data[n + i]) + self.qo_bank.data[n + i].mul_(proj_scale) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) + for i in range(n): + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) + nn.init.zeros_(self.mlp_down_bank.data[i]) + self.mlp_down_bank.data[i].mul_(proj_scale) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif ( + module.weight.ndim == 2 + and module.weight.shape[0] >= 64 + and module.weight.shape[1] >= 64 + ): + nn.init.orthogonal_(module.weight, gain=1.0) + + def _bank_weights(self, i): + n = self.num_layers + return ( + self.qo_bank[i], + self.kv_bank[i], + self.kv_bank[n + i], + self.qo_bank[n + i], + self.mlp_up_bank[i], + self.mlp_down_bank[i], + ) + + def _parallel_block( + self, block_idx, lane0, lane1, x0, + q_w, k_w, v_w, out_w, up_w, down_w, + cu_seqlens=None, max_seqlen=0, + ): + block = self.blocks[block_idx] + mix = block.resid_mix.to(dtype=lane0.dtype) + attn_read = mix[0][None, None, :] * lane0 + mix[1][None, None, :] * x0 + attn_out = block.attn( + block.attn_norm(attn_read) * block.ln_scale_factor, + q_w, k_w, v_w, out_w, + cu_seqlens=cu_seqlens, max_seqlen=max_seqlen, + ) + attn_out = block.attn_scale.to(dtype=attn_out.dtype)[None, None, :] * attn_out + mlp_read = lane1 + mlp_out = block.mlp_scale.to(dtype=lane1.dtype)[None, None, :] * block.mlp( + block.mlp_norm(mlp_read) * block.ln_scale_factor, up_w, down_w + ) + attn_resid = self.parallel_resid_lambdas[block_idx, 0].to(dtype=lane0.dtype) + attn_post = self.parallel_post_lambdas[block_idx, 0].to(dtype=lane0.dtype) + mlp_resid = self.parallel_resid_lambdas[block_idx, 1].to(dtype=lane0.dtype) + mlp_post = self.parallel_post_lambdas[block_idx, 1].to(dtype=lane0.dtype) + lane0 = attn_resid * lane0 + attn_post[0] * attn_out + mlp_post[0] * mlp_out + lane1 = mlp_resid * lane1 + attn_post[1] * attn_out + mlp_post[1] * mlp_out + return lane0, lane1 + + def _final_parallel_hidden(self, lane0, lane1): + if self.parallel_final_lane == "mlp": + return lane1 + if self.parallel_final_lane == "attn": + return lane0 + return 0.5 * (lane0 + lane1) + + def _forward_hidden(self, input_ids, cu_seqlens=None, max_seqlen=0): + """Run the encoder/decoder stack to the final RMSNorm; returns pre-projection hidden. + Shared by eval (softcap+projection via forward_logits) and train (fused CE path).""" + x = self.tok_emb(input_ids) + # SmearGate (PR #1667). Inline gate compute with .contiguous() on the slice fed + # to the projection so torch.compile fullgraph is happy. lam=0 + W=0 -> identity + # at init. This block runs unconditionally on the smear path; the cat keeps + # position 0 untouched so causality holds. + if self.smear_gate_enabled: + sl = self.smear_lambda.to(dtype=x.dtype) + gate_in = x[:, 1:, : self.smear_window].contiguous() + g = sl * torch.sigmoid(self.smear_gate(gate_in)) + bos_mask = (input_ids[:, 1:] == 1).unsqueeze(-1) + g = g.masked_fill(bos_mask, 0.0) + x = torch.cat([x[:, :1], x[:, 1:] + g * x[:, :-1]], dim=1) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips = [] + enc_iter = ( + self.encoder_indices + if self.looping_active + else range(self.num_encoder_layers) + ) + dec_iter = ( + self.decoder_indices + if self.looping_active + else range( + self.num_encoder_layers, + self.num_encoder_layers + self.num_decoder_layers, + ) + ) + for i in enc_iter: + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + x = self.blocks[i](x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + skips.append(x) + psl = self.parallel_start_layer + lane0 = None + lane1 = None + for skip_idx, i in enumerate(dec_iter): + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + if i >= psl and psl > 0: + if lane0 is None: + lane0 = x + lane1 = x + if skip_idx < self.num_skip_weights and skips: + skip = skips.pop() + w = self.skip_weights[skip_idx].to(dtype=lane0.dtype)[None, None, :] + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=lane0.dtype))[None, None, :] + lane0 = torch.lerp(w * skip, lane0, g) + else: + lane0 = lane0 + w * skip + lane0, lane1 = self._parallel_block( + i, lane0, lane1, x0, q_w, k_w, v_w, out_w, up_w, down_w, + cu_seqlens=cu_seqlens, max_seqlen=max_seqlen, + ) + else: + if skip_idx < self.num_skip_weights and skips: + scaled_skip = ( + self.skip_weights[skip_idx].to(dtype=x.dtype)[None, None, :] + * skips.pop() + ) + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=x.dtype))[None, None, :] + x = torch.lerp(scaled_skip, x, g) + else: + x = x + scaled_skip + x = self.blocks[i](x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + if lane0 is not None: + x = self._final_parallel_hidden(lane0, lane1) + x = self.final_norm(x) + return x + + def _project_logits(self, hidden): + if self.tie_embeddings: + return F.linear(hidden, self.tok_emb.weight) + return self.lm_head(hidden) + + def forward_logits(self, input_ids, cu_seqlens=None, max_seqlen=0): + hidden = self._forward_hidden(input_ids, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + logits_proj = self._project_logits(hidden) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + def forward(self, input_ids, target_ids, cu_seqlens=None, max_seqlen=0): + hidden = self._forward_hidden(input_ids, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + logits_proj = self._project_logits(hidden) + flat_targets = target_ids.reshape(-1) + # Fused softcapped-CE kernel (training path only). Applies softcap inside the + # Triton kernel; takes pre-softcap logits_proj. Non-fused path matches stock + # PR-1736 numerics exactly (softcap in fp32, then F.cross_entropy on fp32). + if self.fused_ce_enabled: + return softcapped_cross_entropy( + logits_proj.reshape(-1, logits_proj.size(-1)), + flat_targets, + self.logit_softcap, + reduction="mean", + ) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + flat_targets, + reduction="mean", + ) + + def forward_ttt(self, input_ids, target_ids, lora): + x = self.tok_emb(input_ids) + # SmearGate on the TTT path — same inline compute as forward_logits. + if self.smear_gate_enabled: + sl = self.smear_lambda.to(dtype=x.dtype) + gate_in = x[:, 1:, : self.smear_window].contiguous() + g = sl * torch.sigmoid(self.smear_gate(gate_in)) + bos_mask = (input_ids[:, 1:] == 1).unsqueeze(-1) + g = g.masked_fill(bos_mask, 0.0) + x = torch.cat([x[:, :1], x[:, 1:] + g * x[:, :-1]], dim=1) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips = [] + enc_iter = ( + self.encoder_indices + if self.looping_active + else list(range(self.num_encoder_layers)) + ) + dec_iter = ( + self.decoder_indices + if self.looping_active + else list( + range( + self.num_encoder_layers, + self.num_encoder_layers + self.num_decoder_layers, + ) + ) + ) + slot = 0 + for i in enc_iter: + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + x = self._block_with_lora(self.blocks[i], x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w) + slot += 1 + skips.append(x) + psl = self.parallel_start_layer + lane0 = None + lane1 = None + for skip_idx, i in enumerate(dec_iter): + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + if i >= psl and psl > 0: + if lane0 is None: + lane0 = x + lane1 = x + if skip_idx < self.num_skip_weights and skips: + skip = skips.pop() + w = self.skip_weights[skip_idx].to(dtype=lane0.dtype)[None, None, :] + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=lane0.dtype))[None, None, :] + lane0 = torch.lerp(w * skip, lane0, g) + else: + lane0 = lane0 + w * skip + lane0, lane1 = self._parallel_block_with_lora( + i, lane0, lane1, x0, lora, slot, + q_w, k_w, v_w, out_w, up_w, down_w, + ) + else: + if skip_idx < self.num_skip_weights and skips: + scaled_skip = ( + self.skip_weights[skip_idx].to(dtype=x.dtype)[None, None, :] + * skips.pop() + ) + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=x.dtype))[None, None, :] + x = torch.lerp(scaled_skip, x, g) + else: + x = x + scaled_skip + x = self._block_with_lora(self.blocks[i], x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w) + slot += 1 + if lane0 is not None: + x = self._final_parallel_hidden(lane0, lane1) + x = self.final_norm(x) + if self.tie_embeddings: + logits = F.linear(x, self.tok_emb.weight) + else: + logits = self.lm_head(x) + logits = logits + lora.lm_head_lora(x) + logits = self.logit_softcap * torch.tanh(logits / self.logit_softcap) + bsz, sl, V = logits.shape + return F.cross_entropy( + logits.float().reshape(-1, V), target_ids.reshape(-1), reduction="none" + ).reshape(bsz, sl) + + def _block_with_lora(self, block, x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w): + mix = block.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = block.attn_norm(x_in) * block.ln_scale_factor + attn = block.attn + bsz, seqlen, dim = n.shape + # Keep raw Q for AttnOutGate src='q' (matches forward path semantics). + q_raw = F.linear(n, q_w.to(n.dtype)) + lora.q_loras[slot](n) + q = q_raw.reshape(bsz, seqlen, attn.num_heads, attn.head_dim) + k = F.linear(n, k_w.to(n.dtype)) + if lora.k_loras is not None: + k = k + lora.k_loras[slot](n) + k = k.reshape(bsz, seqlen, attn.num_kv_heads, attn.head_dim) + v = (F.linear(n, v_w.to(n.dtype)) + lora.v_loras[slot](n)).reshape( + bsz, seqlen, attn.num_kv_heads, attn.head_dim + ) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = attn.rotary(seqlen, n.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, attn.rope_dims) + k = apply_rotary_emb(k, cos, sin, attn.rope_dims) + q = q * attn.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if attn.use_xsa: + y = attn._xsa_efficient(y, v) + # AttnOutGate (TTT path) — inline + .contiguous() barrier, same as the eval path. + if attn.attn_out_gate: + gate_src = q_raw if attn.attn_out_gate_src == "q" else n + gate_in = gate_src[..., : attn.gate_window].contiguous() + g = 2.0 * torch.sigmoid(attn.attn_gate_proj(gate_in)) + y = y * g[..., None] + # Gated Attention (TTT path). Gate input is n (post-norm block input), same + # as eval path. .to(n.dtype) on fp32 param before bf16 broadcast. + if attn.gated_attn: + n_c = n.contiguous() + g = torch.sigmoid(F.linear(n_c, attn.attn_gate_w.to(n.dtype))) + y = y * g[..., None] + # Sparse attention head-output gate (TTT path) — must match the eval path in + # forward() exactly, else training (which applied the gate) and TTT eval (which + # skipped it) produce mismatched representations and catastrophic BPB regression. + if attn.sparse_attn_gate: + gate_in = n[..., : attn.gate_window].contiguous() + g = torch.sigmoid( + attn.sparse_attn_gate_scale + * F.linear(gate_in, attn.attn_gate_w.to(n.dtype)) + ) + y = y * g[..., None] + y = y.reshape(bsz, seqlen, dim) + attn_out = F.linear(y, out_w.to(n.dtype)) + if lora.o_loras is not None: + attn_out = attn_out + lora.o_loras[slot](n) + x_out = x_in + block.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + mlp_n = block.mlp_norm(x_out) * block.ln_scale_factor + mlp_out = block.mlp(mlp_n, up_w, down_w) + if lora.mlp_loras is not None: + mlp_out = mlp_out + lora.mlp_loras[slot](mlp_n) + x_out = x_out + block.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * mlp_out + return x_out + + def _parallel_block_with_lora( + self, block_idx, lane0, lane1, x0, lora, slot, + q_w, k_w, v_w, out_w, up_w, down_w, + ): + block = self.blocks[block_idx] + mix = block.resid_mix.to(dtype=lane0.dtype) + attn_read = mix[0][None, None, :] * lane0 + mix[1][None, None, :] * x0 + n = block.attn_norm(attn_read) * block.ln_scale_factor + attn = block.attn + bsz, seqlen, dim = n.shape + q_raw = F.linear(n, q_w.to(n.dtype)) + lora.q_loras[slot](n) + q = q_raw.reshape(bsz, seqlen, attn.num_heads, attn.head_dim) + k = F.linear(n, k_w.to(n.dtype)) + if lora.k_loras is not None: + k = k + lora.k_loras[slot](n) + k = k.reshape(bsz, seqlen, attn.num_kv_heads, attn.head_dim) + v = (F.linear(n, v_w.to(n.dtype)) + lora.v_loras[slot](n)).reshape( + bsz, seqlen, attn.num_kv_heads, attn.head_dim + ) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = attn.rotary(seqlen, n.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, attn.rope_dims) + k = apply_rotary_emb(k, cos, sin, attn.rope_dims) + q = q * attn.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if attn.use_xsa: + y = attn._xsa_efficient(y, v) + # AttnOutGate (TTT parallel path) — inline + .contiguous() barrier. + if attn.attn_out_gate: + gate_src = q_raw if attn.attn_out_gate_src == "q" else n + gate_in = gate_src[..., : attn.gate_window].contiguous() + g = 2.0 * torch.sigmoid(attn.attn_gate_proj(gate_in)) + y = y * g[..., None] + # Gated Attention (TTT parallel path). Gate input is n (post-norm block input). + if attn.gated_attn: + n_c = n.contiguous() + g = torch.sigmoid(F.linear(n_c, attn.attn_gate_w.to(n.dtype))) + y = y * g[..., None] + # Sparse attention head-output gate (TTT parallel path) — must match the + # eval path in forward() to keep train/eval semantics in sync. + if attn.sparse_attn_gate: + gate_in = n[..., : attn.gate_window].contiguous() + g = torch.sigmoid( + attn.sparse_attn_gate_scale + * F.linear(gate_in, attn.attn_gate_w.to(n.dtype)) + ) + y = y * g[..., None] + y = y.reshape(bsz, seqlen, dim) + attn_out = F.linear(y, out_w.to(n.dtype)) + if lora.o_loras is not None: + attn_out = attn_out + lora.o_loras[slot](n) + attn_out = block.attn_scale.to(dtype=attn_out.dtype)[None, None, :] * attn_out + mlp_read = lane1 + mlp_n = block.mlp_norm(mlp_read) * block.ln_scale_factor + mlp_out = block.mlp(mlp_n, up_w, down_w) + if lora.mlp_loras is not None: + mlp_out = mlp_out + lora.mlp_loras[slot](mlp_n) + mlp_out = block.mlp_scale.to(dtype=lane1.dtype)[None, None, :] * mlp_out + attn_resid = self.parallel_resid_lambdas[block_idx, 0].to(dtype=lane0.dtype) + attn_post = self.parallel_post_lambdas[block_idx, 0].to(dtype=lane0.dtype) + mlp_resid = self.parallel_resid_lambdas[block_idx, 1].to(dtype=lane0.dtype) + mlp_post = self.parallel_post_lambdas[block_idx, 1].to(dtype=lane0.dtype) + lane0 = attn_resid * lane0 + attn_post[0] * attn_out + mlp_post[0] * mlp_out + lane1 = mlp_resid * lane1 + attn_post[1] * attn_out + mlp_post[1] * mlp_out + return lane0, lane1 + + +class BatchedLinearLoRA(nn.Module): + # PR-1767: rank-scaled output (alpha/rank), like standard LoRA. Decouples + # effective magnitude from rank so changing rank does not change LR scale. + _ALPHA = float(os.environ.get("TTT_LORA_ALPHA", "144")) + # PR-1767: optionally keep A warm across per-doc resets (only B is zeroed). + # Accumulates useful feature directions across documents within a TTT phase. + _WARM_START_A = bool(int(os.environ.get("TTT_WARM_START_A", "1"))) + + def __init__(self, bsz, in_features, out_features, rank): + super().__init__() + self._bound = 1.0 / math.sqrt(in_features) + self._scale = self._ALPHA / rank + self.A = nn.Parameter( + torch.empty(bsz, rank, in_features).uniform_(-self._bound, self._bound) + ) + self.B = nn.Parameter(torch.zeros(bsz, out_features, rank)) + + def reset(self): + with torch.no_grad(): + if not self._WARM_START_A: + self.A.uniform_(-self._bound, self._bound) + self.B.zero_() + + def forward(self, x): + return ((x @ self.A.transpose(1, 2)) @ self.B.transpose(1, 2)) * self._scale + + +class BatchedTTTLoRA(nn.Module): + def __init__(self, bsz, model, rank, k_lora=True, mlp_lora=True, o_lora=True): + super().__init__() + self.bsz = bsz + dim = model.qo_bank.shape[-1] + vocab = model.tok_emb.num_embeddings + if getattr(model, "looping_active", False): + num_slots = len(model.encoder_indices) + len(model.decoder_indices) + else: + num_slots = len(model.blocks) + kv_dim = model.blocks[0].attn.num_kv_heads * ( + dim // model.blocks[0].attn.num_heads + ) + embed_dim = model.tok_emb.embedding_dim + self.lm_head_lora = BatchedLinearLoRA(bsz, embed_dim, vocab, rank) + self.q_loras = nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + self.v_loras = nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, kv_dim, rank) for _ in range(num_slots)] + ) + self.k_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, kv_dim, rank) for _ in range(num_slots)] + ) + if k_lora + else None + ) + self.mlp_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + if mlp_lora + else None + ) + self.o_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + if o_lora + else None + ) + + def reset(self): + with torch.no_grad(): + self.lm_head_lora.reset() + for loras in [self.q_loras, self.v_loras, self.k_loras, + self.mlp_loras, self.o_loras]: + if loras is not None: + for lora in loras: + lora.reset() + + +# Polar Express per-iteration minimax Newton-Schulz coefficients (PR #1344). +# Replaces the fixed (3.4445, -4.775, 2.0315) coefficients of stock Muon. +# Applied at backend_steps=5 — taking more than 5 iterations from this list +# falls back to the final (converged) tuple via the slice guard below. +_PE_COEFFS = ( + (8.156554524902461, -22.48329292557795, 15.878769915207462), + (4.042929935166739, -2.808917465908714, 0.5000178451051316), + (3.8916678022926607, -2.772484153217685, 0.5060648178503393), + (3.285753657755655, -2.3681294933425376, 0.46449024233003106), + (2.3465413258596377, -1.7097828382687081, 0.42323551169305323), +) + + +@torch.compile +def zeropower_via_newtonschulz5(G, steps=10, eps=1e-07): + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + coeffs = _PE_COEFFS[:steps] if steps <= len(_PE_COEFFS) else _PE_COEFFS + for a, b, c in coeffs: + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + + +class Muon(torch.optim.Optimizer): + def __init__( + self, + params, + lr, + momentum, + backend_steps, + nesterov=True, + weight_decay=0.0, + row_normalize=False, + ): + super().__init__( + params, + dict( + lr=lr, + momentum=momentum, + backend_steps=backend_steps, + nesterov=nesterov, + weight_decay=weight_decay, + row_normalize=row_normalize, + ), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + "p": p, + "B": B, + "padded_grad": torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + "shard": torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + "shard_mom": torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + "full_update": torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + "scale": max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + self._bank_meta.sort(key=lambda m: -m["p"].numel()) + self._built = True + + def launch_reduce_scatters(self): + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m["p"] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m["padded_grad"] + pg[: m["B"]].copy_(p.grad.bfloat16()) + if pg.shape[0] > m["B"]: + pg[m["B"] :].zero_() + fut = dist.reduce_scatter_tensor( + m["shard"], pg, op=dist.ReduceOp.AVG, async_op=True + ) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + if not self._built: + self._build() + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + row_normalize = group.get("row_normalize", False) + prev_ag_handle = None + prev_m = None + sharded = self._distributed and hasattr(self, "_rs_futures") + for idx, m in enumerate(self._bank_meta): + p = m["p"] + if p.grad is None: + continue + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m["p"] + upd = prev_m["full_update"][: prev_m["B"]] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m["scale"]) + if sharded and self._rs_futures[idx] is not None: + self._rs_futures[idx].wait() + g = m["shard"] + buf = m["shard_mom"] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + if row_normalize: + rn = update.float().norm(dim=-1, keepdim=True).clamp_min(1e-07) + update = update / rn.to(update.dtype) + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m["full_update"], update, async_op=True + ) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m["scale"]) + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m["p"] + upd = prev_m["full_update"][: prev_m["B"]] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m["scale"]) + if hasattr(self, "_rs_futures"): + del self._rs_futures + return loss + + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,skip_gates,parallel_post_lambdas,parallel_resid_lambdas,attn_gate_proj,attn_gate_w,smear_gate,smear_lambda", + ).split(",") + if pattern +) + + +PACKED_REPLICATED_GRAD_MAX_NUMEL = 1 << 15 + + +class Optimizers: + def __init__(self, h, base_model): + matrix_params = [ + base_model.qo_bank, + base_model.kv_bank, + base_model.mlp_up_bank, + base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for (name, p) in block_named_params + if p.ndim < 2 + or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + if base_model.skip_gates is not None and base_model.skip_gates.numel() > 0: + scalar_params.append(base_model.skip_gates) + if base_model.parallel_post_lambdas is not None: + scalar_params.append(base_model.parallel_post_lambdas) + if base_model.parallel_resid_lambdas is not None: + scalar_params.append(base_model.parallel_resid_lambdas) + # SmearGate params live on GPT root (not in .blocks), so add them by hand. + # Both are tiny (gate_window scalars + 1 lambda). Optimized via scalar Adam. + if getattr(base_model, "smear_gate_enabled", False): + scalar_params.append(base_model.smear_gate.weight) + scalar_params.append(base_model.smear_lambda) + token_lr = h.tied_embed_lr if h.tie_embeddings else h.embed_lr + tok_params = [ + {"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr} + ] + self.optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(h.beta1, h.beta2), + eps=h.adam_eps, + weight_decay=h.embed_wd, + fused=True, + ) + self.optimizer_muon = Muon( + matrix_params, + lr=h.matrix_lr, + momentum=h.muon_momentum, + backend_steps=h.muon_backend_steps, + weight_decay=h.muon_wd, + row_normalize=h.muon_row_normalize, + ) + for group in self.optimizer_muon.param_groups: + group["base_lr"] = h.matrix_lr + self.optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": h.scalar_lr, "base_lr": h.scalar_lr}], + betas=(h.beta1, h.beta2), + eps=h.adam_eps, + weight_decay=h.adam_wd, + fused=True, + ) + self.optimizers = [ + self.optimizer_tok, + self.optimizer_muon, + self.optimizer_scalar, + ] + self.replicated_params = list(tok_params[0]["params"]) + self.replicated_params.extend(scalar_params) + self.replicated_large_params = [] + self.replicated_packed_params = [] + for p in self.replicated_params: + if p.numel() <= PACKED_REPLICATED_GRAD_MAX_NUMEL: + self.replicated_packed_params.append(p) + else: + self.replicated_large_params.append(p) + + def __iter__(self): + return iter(self.optimizers) + + def zero_grad_all(self): + for opt in self.optimizers: + opt.zero_grad(set_to_none=True) + + def _all_reduce_packed_grads(self): + grads_by_key = collections.defaultdict(list) + for p in self.replicated_packed_params: + if p.grad is not None: + grads_by_key[(p.grad.device, p.grad.dtype)].append(p.grad) + for grads in grads_by_key.values(): + flat = torch.empty( + sum(g.numel() for g in grads), + device=grads[0].device, + dtype=grads[0].dtype, + ) + offset = 0 + for g in grads: + n = g.numel() + flat[offset : offset + n].copy_(g.contiguous().view(-1)) + offset += n + dist.all_reduce(flat, op=dist.ReduceOp.AVG) + offset = 0 + for g in grads: + n = g.numel() + g.copy_(flat[offset : offset + n].view_as(g)) + offset += n + + def step(self, distributed=False): + self.optimizer_muon.launch_reduce_scatters() + if distributed: + reduce_handles = [ + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG, async_op=True) + for p in self.replicated_large_params + if p.grad is not None + ] + self._all_reduce_packed_grads() + for handle in reduce_handles: + handle.wait() + self.optimizer_tok.step() + self.optimizer_scalar.step() + self.optimizer_muon.step() + self.zero_grad_all() + + +def restore_fp32_params(model): + for module in model.modules(): + if isinstance(module, CastedLinear): + module.float() + for name, param in model.named_parameters(): + if ( + param.ndim < 2 + or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ) and param.dtype != torch.float32: + param.data = param.data.float() + if hasattr(model, "qo_bank") and model.qo_bank is not None: + model.qo_bank.data = model.qo_bank.data.float() + model.kv_bank.data = model.kv_bank.data.float() + model.mlp_up_bank.data = model.mlp_up_bank.data.float() + model.mlp_down_bank.data = model.mlp_down_bank.data.float() + + +def collect_hessians(model, train_loader, h, device, n_calibration_batches=64): + hessians = {} + hooks = [] + for i, block in enumerate(model.blocks): + block.attn._calib = True + block.mlp._calib = True + block.mlp.use_fused = False + + def make_attn_hook(layer_idx): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + for suffix in ["c_q", "c_k", "c_v"]: + name = f"blocks.{layer_idx}.attn.{suffix}.weight" + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + y = module._last_proj_input + if y is not None: + y = y.float() + if y.ndim == 3: + y = y.reshape(-1, y.shape[-1]) + name = f"blocks.{layer_idx}.attn.proj.weight" + if name not in hessians: + hessians[name] = torch.zeros( + y.shape[1], y.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(y.T, y) + return hook_fn + + def make_mlp_hook(layer_idx): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + name = f"blocks.{layer_idx}.mlp.fc.weight" + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + h_act = module._last_down_input + if h_act is not None: + h_act = h_act.float() + if h_act.ndim == 3: + h_act = h_act.reshape(-1, h_act.shape[-1]) + name = f"blocks.{layer_idx}.mlp.proj.weight" + if name not in hessians: + hessians[name] = torch.zeros( + h_act.shape[1], h_act.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(h_act.T, h_act) + return hook_fn + + for i, block in enumerate(model.blocks): + hooks.append(block.attn.register_forward_hook(make_attn_hook(i))) + hooks.append(block.mlp.register_forward_hook(make_mlp_hook(i))) + + # Hessian hooks for embedding factorization projection layers + def make_linear_input_hook(weight_name): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if weight_name not in hessians: + hessians[weight_name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[weight_name].addmm_(x.T, x) + return hook_fn + + if model.tie_embeddings: + hook_module = model.final_norm + + def make_output_hook(name): + def hook_fn(module, inp, out): + x = out.detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + return hook_fn + + hooks.append( + hook_module.register_forward_hook(make_output_hook("tok_emb.weight")) + ) + model.eval() + with torch.no_grad(): + for _ in range(n_calibration_batches): + x, _ = train_loader.next_batch(h.train_batch_tokens, h.grad_accum_steps) + model.forward_logits(x) + for hook in hooks: + hook.remove() + for i, block in enumerate(model.blocks): + block.attn._calib = False + block.mlp._calib = False + block.mlp.use_fused = True + for name in hessians: + hessians[name] = hessians[name].cpu() / n_calibration_batches + return hessians + + +def gptq_quantize_weight(w, H, clip_sigmas=3.0, clip_range=63, block_size=128): + W_orig = w.float().clone() + rows, cols = W_orig.shape + H = H.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = 0.01 * H.diag().mean() + H.diagonal().add_(damp) + perm = torch.argsort(H.diag(), descending=True) + invperm = torch.argsort(perm) + W_perm = W_orig[:, perm].clone() + W_perm[:, dead[perm]] = 0 + H = H[perm][:, perm] + Hinv = torch.cholesky_inverse(torch.linalg.cholesky(H)) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + row_std = W_orig.std(dim=1) + s = (clip_sigmas * row_std / clip_range).clamp_min(1e-10).to(torch.float16) + sf = s.float() + Q = torch.zeros(rows, cols, dtype=torch.int8) + W_work = W_perm.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W_work[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros(rows, i2 - i1) + for j in range(i2 - i1): + w_col = W_block[:, j] + d = Hinv_block[j, j] + q_col = torch.clamp(torch.round(w_col / sf), -clip_range, clip_range) + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - q_col.float() * sf) / d + Err[:, j] = err + W_block[:, j:] -= err.unsqueeze(1) * Hinv_block[j, j:].unsqueeze(0) + if i2 < cols: + W_work[:, i2:] -= Err @ Hinv[i1:i2, i2:] + return Q[:, invperm], s + + +def _quantize_gate_int8_row(w): + # Symmetric int8-per-row quantization for small gate tensors. w shape + # (R, C) -> (R,) scales in fp16, int8 values in [-127, 127]. Single scale + # per row keeps accuracy high while halving storage vs fp16. + W = w.float().contiguous() + row_max = W.abs().amax(dim=1).clamp_min(1e-10) + s = (row_max / 127.0).to(torch.float16) + sf = s.float().view(-1, 1) + q = torch.clamp(torch.round(W / sf), -127, 127).to(torch.int8) + return q, s + + +def _lqer_pack(A, B, bits): + rng = 2 ** (bits - 1) - 1 + sA = (A.abs().amax(dim=1).clamp_min(1e-10) / rng).to(torch.float16) + sB = (B.abs().amax(dim=1).clamp_min(1e-10) / rng).to(torch.float16) + qA = torch.clamp(torch.round(A / sA.float().view(-1, 1)), -rng, rng).to(torch.int8) + qB = torch.clamp(torch.round(B / sB.float().view(-1, 1)), -rng, rng).to(torch.int8) + return qA, sA, qB, sB + + +def _lqer_pack_asym(A, B, g=64): + # A: INT2 per-matrix scalar (signed [-2,1], scale = |A|max/1.5). + sA = (A.abs().amax().clamp_min(1e-10) / 1.5).to(torch.float16) + qA = torch.clamp(torch.round(A / sA.float()), -2, 1).to(torch.int8) + # B: INT4 groupwise g over flattened B (signed [-8,7], per-group scale). + Bf = B.reshape(-1, g) + Bmax = Bf.abs().amax(dim=-1, keepdim=True).clamp_min(1e-10) + sB = (Bmax / 7.5).to(torch.float16).reshape(-1) + qB = torch.clamp(torch.round(Bf / sB.float().reshape(-1, 1)), -8, 7).to( + torch.int8 + ).reshape(B.shape) + return qA, sA, qB, sB + + +def gptq_mixed_quantize(state_dict, hessians, h): + result = {} + meta = {} + quant_gate = bool(getattr(h, "gated_attn_quant_gate", False)) + lqer_on = bool(getattr(h, "lqer_enabled", False)) + lqer_cands = {} + for (name, tensor) in state_dict.items(): + t = tensor.detach().cpu().contiguous() + # Dedicated int8-per-row path for attn_gate_w (bypasses both GPTQ and + # fp16 passthrough). Applied BEFORE the numel<=65536 passthrough check + # so the gate tensor is routed here instead of to fp16. + if ( + quant_gate + and t.is_floating_point() + and t.ndim == 2 + and name.endswith(".attn_gate_w") + # Dense GatedAttn: (num_heads, dim) = (8, 512) = 4096. + # Sparse gate: (num_heads, gate_window) = (8, 12) = 96. + # Both need int8-per-row routing; the 1024 lower bound in stock + # PR-1736 presumed dense-only. Widen to catch both. + and 32 <= t.numel() <= 8192 + ): + gq, gs = _quantize_gate_int8_row(t) + result[name + ".gq"] = gq + result[name + ".gs"] = gs + meta[name] = "gate_int8_row" + continue + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough (float16)" + continue + if "tok_emb" in name: + cs = h.embed_clip_sigmas + elif ".mlp." in name: + cs = h.mlp_clip_sigmas + elif ".attn." in name: + cs = h.attn_clip_sigmas + else: + cs = h.matrix_clip_sigmas + bits = h.embed_bits if "tok_emb" in name else h.matrix_bits + clip_range = 2 ** (bits - 1) - 1 + ret = gptq_quantize_weight( + t, hessians[name], clip_sigmas=cs, clip_range=clip_range + ) + q, s = ret + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = f"gptq (int{bits})" + if lqer_on: + W_q = q.float() * s.float().view(-1, 1) + E = t.float() - W_q + lqer_cands[name] = (E, float(E.norm())) + if lqer_on and lqer_cands: + top = sorted(lqer_cands.items(), key=lambda kv: -kv[1][1])[: h.lqer_top_k] + asym_on = bool(getattr(h, "lqer_asym_enabled", False)) + asym_g = int(getattr(h, "lqer_asym_group", 64)) + for (name, (E, _)) in top: + U, S, Vh = torch.linalg.svd(E, full_matrices=False) + r = min(h.lqer_rank, S.numel()) + A = (U[:, :r] * S[:r]).contiguous() + B = Vh[:r, :].contiguous() + if asym_on and B.numel() % asym_g == 0: + qA, sA, qB, sB = _lqer_pack_asym(A, B, asym_g) + result[name + ".lqA_a"] = qA + result[name + ".lqAs_a"] = sA + result[name + ".lqB_a"] = qB + result[name + ".lqBs_a"] = sB + meta[name] = meta[name] + "+lqer_asym" + else: + qA, sA, qB, sB = _lqer_pack(A, B, h.lqer_factor_bits) + result[name + ".lqA"] = qA + result[name + ".lqAs"] = sA + result[name + ".lqB"] = qB + result[name + ".lqBs"] = sB + meta[name] = meta[name] + "+lqer" + categories = collections.defaultdict(set) + for (name, cat) in meta.items(): + short = re.sub("\\.\\d+$", "", re.sub("blocks\\.\\d+", "blocks", name)) + categories[cat].add(short) + log("Quantized weights:") + for cat in sorted(categories): + log(f" {cat}: {', '.join(sorted(categories[cat]))}") + return result, meta + +def dequantize_mixed(result, meta, template_sd): + out = {} + for (name, orig) in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if "passthrough" in info: + t = result[name] + if t.dtype == torch.float16 and orig_dtype in ( + torch.float32, + torch.bfloat16, + ): + t = t.to(orig_dtype) + out[name] = t + continue + if info == "gate_int8_row": + gq = result[name + ".gq"] + gs = result[name + ".gs"] + out[name] = (gq.float() * gs.float().view(-1, 1)).to(orig_dtype) + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + W = q.float() * s.float().view(q.shape[0], *[1] * (q.ndim - 1)) + else: + W = q.float() * float(s.item()) + if "lqer_asym" in info: + qA_t = result[name + ".lqA_a"] + sA_t = result[name + ".lqAs_a"] + qB_t = result[name + ".lqB_a"] + sB_t = result[name + ".lqBs_a"] + qA = qA_t.float() * float(sA_t) + g_sz = qB_t.numel() // sB_t.numel() + qB = (qB_t.reshape(-1, g_sz).float() * sB_t.float().view(-1, 1)).reshape( + qB_t.shape + ) + W = W + qA @ qB + elif "lqer" in info: + qA = result[name + ".lqA"].float() * result[name + ".lqAs"].float().view(-1, 1) + qB = result[name + ".lqB"].float() * result[name + ".lqBs"].float().view(-1, 1) + W = W + qA @ qB + out[name] = W.to(orig_dtype) + return out + + +_BSHF_MAGIC = b"BSHF" + + +def _byte_shuffle(data, stride=2): + if stride <= 1 or len(data) < stride: + return data + src = np.frombuffer(data, dtype=np.uint8) + n = len(src) + out = np.empty(n, dtype=np.uint8) + dest_off = 0 + for pos in range(stride): + chunk = src[pos::stride] + out[dest_off : dest_off + len(chunk)] = chunk + dest_off += len(chunk) + return _BSHF_MAGIC + bytes([stride]) + out.tobytes() + + +def _byte_unshuffle(data): + if len(data) < 5 or data[:4] != _BSHF_MAGIC: + return data + stride = data[4] + if stride < 2: + return data[5:] + payload = np.frombuffer(data, dtype=np.uint8, offset=5) + n = len(payload) + out = np.empty(n, dtype=np.uint8) + src_off = 0 + for pos in range(stride): + chunk_len = n // stride + (1 if pos < n % stride else 0) + out[pos::stride][:chunk_len] = payload[src_off : src_off + chunk_len] + src_off += chunk_len + return out.tobytes() + + +def _compress(data, compressor): + data = _byte_shuffle(data) + if compressor == "lzma": + return lzma.compress(data, preset=6) + elif compressor == "brotli": + import brotli + + return brotli.compress(data, quality=11) + raise ValueError(f"Unknown compressor: {compressor!r}") + + +def _decompress(data, compressor): + if compressor == "lzma": + raw = lzma.decompress(data) + elif compressor == "brotli": + import brotli + + raw = brotli.decompress(data) + else: + raise ValueError(f"Unknown compressor: {compressor!r}") + raw = _byte_unshuffle(raw) + return raw + + +def _unbank_state_dict(state_dict, num_layers): + sd = {} + n = num_layers + for k, v in state_dict.items(): + t = v.detach().cpu() if v is not None else None + if k == "qo_bank": + for i in range(n): + sd[f"blocks.{i}.attn.c_q.weight"] = t[i] + sd[f"blocks.{i}.attn.proj.weight"] = t[n + i] + elif k == "kv_bank": + for i in range(n): + sd[f"blocks.{i}.attn.c_k.weight"] = t[i] + sd[f"blocks.{i}.attn.c_v.weight"] = t[n + i] + elif k == "mlp_up_bank": + for i in range(n): + sd[f"blocks.{i}.mlp.fc.weight"] = t[i] + elif k == "mlp_down_bank": + for i in range(n): + sd[f"blocks.{i}.mlp.proj.weight"] = t[i] + else: + if t is not None: + sd[k] = t + return sd + + +def _rebank_state_dict(flat_sd, num_layers, model_dim, kv_dim, hidden_dim): + sd = {} + n = num_layers + sd["qo_bank"] = torch.zeros(2 * n, model_dim, model_dim) + sd["kv_bank"] = torch.zeros(2 * n, kv_dim, model_dim) + for i in range(n): + sd["qo_bank"][i] = flat_sd[f"blocks.{i}.attn.c_q.weight"] + sd["qo_bank"][n + i] = flat_sd[f"blocks.{i}.attn.proj.weight"] + sd["kv_bank"][i] = flat_sd[f"blocks.{i}.attn.c_k.weight"] + sd["kv_bank"][n + i] = flat_sd[f"blocks.{i}.attn.c_v.weight"] + sd["mlp_up_bank"] = torch.zeros(n, hidden_dim, model_dim) + sd["mlp_down_bank"] = torch.zeros(n, model_dim, hidden_dim) + for i in range(n): + sd["mlp_up_bank"][i] = flat_sd[f"blocks.{i}.mlp.fc.weight"] + sd["mlp_down_bank"][i] = flat_sd[f"blocks.{i}.mlp.proj.weight"] + for k, v in flat_sd.items(): + if not ( + k.startswith("blocks.") + and any( + p in k + for p in [ + ".attn.c_q.", ".attn.c_k.", ".attn.c_v.", + ".attn.proj.", ".mlp.fc.", ".mlp.proj.", + ] + ) + ): + sd[k] = v + return sd + + + +def _compressed_code_size(code): + code_raw = code.encode("utf-8") + minified = subprocess.run( + ["pyminify", "--no-rename-locals", "--no-hoist-literals", "--remove-literal-statements", "-"], + input=code_raw, capture_output=True, check=True, + ).stdout + compressed = lzma.compress(minified) + encoded = base64.b85encode(compressed) + wrapper = b'import lzma as L,base64 as B\nexec(L.decompress(B.b85decode("' + encoded + b'")))\n' + return len(code_raw), len(wrapper) + + +def serialize(h, base_model, code): + code_bytes_uncompressed, code_bytes = _compressed_code_size(code) + if h.is_main_process: + torch.save(base_model.state_dict(), h.model_path) + model_bytes = os.path.getsize(h.model_path) + log(f"Serialized model: {model_bytes} bytes") + log(f"Code size (uncompressed): {code_bytes_uncompressed} bytes") + log(f"Code size (compressed): {code_bytes} bytes") + sd_cpu = _unbank_state_dict(base_model.state_dict(), h.num_layers) + device = torch.device("cuda", h.local_rank) + t0 = time.perf_counter() + calib_loader = ShuffledSequenceLoader(h, device) + log("GPTQ:collecting Hessians from calibration data...") + hessians = collect_hessians( + base_model, + calib_loader, + h, + device, + n_calibration_batches=h.gptq_calibration_batches, + ) + log(f"GPTQ:collected {len(hessians)} Hessians in {time.perf_counter()-t0:.1f}s") + quant_result, quant_meta = gptq_mixed_quantize(sd_cpu, hessians, h) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = _compress(quant_raw, h.compressor) + quant_file_bytes = len(quant_blob) + bytes_total = quant_file_bytes + code_bytes + if h.is_main_process: + with open(h.quantized_model_path, "wb") as f: + f.write(quant_blob) + log(f"Serialized model quantized+{h.compressor}: {quant_file_bytes} bytes") + log(f"Total submission size quantized+{h.compressor}: {bytes_total} bytes") + return bytes_total, quant_file_bytes + + +def deserialize(h, device): + eval_model = GPT(h).to(device).bfloat16() + restore_fp32_params(eval_model) + flat_template = _unbank_state_dict(eval_model.state_dict(), h.num_layers) + with open(h.quantized_model_path, "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(_decompress(quant_blob_disk, h.compressor)), map_location="cpu" + ) + deq_flat = dequantize_mixed(quant_state["w"], quant_state["m"], flat_template) + head_dim = h.model_dim // h.num_heads + kv_dim = h.num_kv_heads * head_dim + hidden_dim = int(h.mlp_mult * h.model_dim) + deq_state = _rebank_state_dict(deq_flat, h.num_layers, h.model_dim, kv_dim, hidden_dim) + eval_model.load_state_dict(deq_state, strict=True) + return eval_model + + +def _loss_bpb(loss_sum, token_count, byte_count): + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + return val_loss, val_bpb + + +def eval_val(h, device, val_data, model, forward_logits_fn=None): + seq_len = h.eval_seq_len + local_batch_tokens = h.val_batch_tokens // (h.world_size * h.grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + f"VAL_BATCH_SIZE must provide at least one sequence per rank; got VAL_BATCH_SIZE={h.val_batch_tokens}, WORLD_SIZE={h.world_size}, GRAD_ACCUM_STEPS={h.grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_data.val_tokens.numel() - 1) // seq_len + seq_start = total_seqs * h.rank // h.world_size + seq_end = total_seqs * (h.rank + 1) // h.world_size + + # TODO: Don't truncate this. + seq_end = seq_start + ((seq_end - seq_start) // local_batch_seqs) * local_batch_seqs + + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + run_forward_logits = ( + (model.module.forward_logits if hasattr(model, "module") else model.forward_logits) + if forward_logits_fn is None + else forward_logits_fn + ) + model.eval() + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + with torch.no_grad(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_data.val_tokens[raw_start:raw_end].to( + device=device, dtype=torch.int64, non_blocking=True + ) + x = local[:-1] + y = local[1:] + bos_pos = (x == BOS_ID).nonzero(as_tuple=True)[0].tolist() + cu_seqlens, max_seqlen = _build_cu_seqlens( + bos_pos, x.numel(), x.device, h.eval_seq_len, 64 + ) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = run_forward_logits( + x[None], cu_seqlens=cu_seqlens, max_seqlen=max_seqlen + ).detach() + per_token_loss = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y.reshape(-1), + reduction="none", + ) + val_loss_sum += per_token_loss.to(torch.float64).sum() + val_token_count += float(y.numel()) + prev_ids = x + tgt_ids = y + if val_data.caseops_enabled and val_data.val_bytes is not None: + # CaseOps: read per-token byte budget from sidecar at the same + # global positions as the target tokens y. raw_start/raw_end + # span [raw_start, raw_end), x = local[:-1], y = local[1:], + # so y is at sidecar positions [raw_start + 1, raw_end). + sidecar_slice = val_data.val_bytes[raw_start + 1 : raw_end].to( + device=device, dtype=torch.int32, non_blocking=True + ) + val_byte_count += sidecar_slice.to(torch.float64).sum() + else: + token_bytes = val_data.base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += ( + val_data.has_leading_space_lut[tgt_ids] + & ~val_data.is_boundary_token_lut[prev_ids] + ).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + model.train() + return _loss_bpb(val_loss_sum, val_token_count, val_byte_count) + + +def _find_docs(all_tokens): + bos_positions = (all_tokens == BOS_ID).nonzero(as_tuple=True)[0].numpy() + docs = [] + for i in range(len(bos_positions)): + start = int(bos_positions[i]) + end = ( + int(bos_positions[i + 1]) + if i + 1 < len(bos_positions) + else all_tokens.numel() + ) + if i + 1 < len(bos_positions): + end += 1 + assert end - start >= 2 + docs.append((start, end - start)) + return docs + + +def _build_ttt_global_batches(doc_entries, h, ascending=False): + batch_size = h.ttt_batch_size + global_doc_entries = sorted(doc_entries, key=lambda x: x[1][1]) + global_batches = [ + global_doc_entries[i : i + batch_size] + for i in range(0, len(global_doc_entries), batch_size) + ] + indexed = list(enumerate(global_batches)) + if not ascending: + indexed.sort(key=lambda ib: -max(dl for _, (_, dl) in ib[1])) + return indexed + + +def _init_batch_counter(path): + with open(path, "wb") as f: + f.write((0).to_bytes(4, "little")) + + +def _claim_next_batch(counter_path, queue_len): + try: + with open(counter_path, "r+b") as f: + fcntl.flock(f, fcntl.LOCK_EX) + idx = int.from_bytes(f.read(4), "little") + f.seek(0) + f.write((idx + 1).to_bytes(4, "little")) + f.flush() + except FileNotFoundError: + return queue_len + return idx + + +def _compute_chunk_window(ci, pred_len, num_chunks, chunk_size, eval_seq_len): + chunk_end = pred_len if ci == num_chunks - 1 else (ci + 1) * chunk_size + win_start = max(0, chunk_end - eval_seq_len) + win_len = chunk_end - win_start + chunk_start = ci * chunk_size + chunk_offset = chunk_start - win_start + chunk_len = chunk_end - chunk_start + return win_start, win_len, chunk_offset, chunk_len + + +def _accumulate_bpb( + ptl, + x, + y, + chunk_offsets, + chunk_lens, + pos_idx, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + loss_sum, + byte_sum, + token_count, + y_bytes=None, +): + pos = pos_idx[: x.size(1)].unsqueeze(0) + mask = ( + (chunk_lens.unsqueeze(1) > 0) + & (pos >= chunk_offsets.unsqueeze(1)) + & (pos < (chunk_offsets + chunk_lens).unsqueeze(1)) + ) + mask_f64 = mask.to(torch.float64) + if y_bytes is not None: + tok_bytes = y_bytes.to(torch.float64) + else: + tok_bytes = base_bytes_lut[y].to(torch.float64) + tok_bytes += (has_leading_space_lut[y] & ~is_boundary_token_lut[x]).to( + torch.float64 + ) + loss_sum += (ptl.to(torch.float64) * mask_f64).sum() + byte_sum += (tok_bytes * mask_f64).sum() + token_count += chunk_lens.to(torch.float64).sum() + + +def _loss_bpb_from_sums(loss_sum, token_count, byte_sum): + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_sum.item()) + return val_loss, val_bpb + + +def _add_to_counter(path, delta): + try: + with open(path, "r+b") as f: + fcntl.flock(f, fcntl.LOCK_EX) + cur = int.from_bytes(f.read(8), "little", signed=True) + cur += int(delta) + f.seek(0) + f.write(int(cur).to_bytes(8, "little", signed=True)) + f.flush() + return cur + except FileNotFoundError: + return int(delta) + + +def _init_int64_counter(path): + with open(path, "wb") as f: + f.write((0).to_bytes(8, "little", signed=True)) + + +def _select_ttt_doc_entries(docs, h): + doc_entries = list(enumerate(docs)) + if h.val_doc_fraction < 1.0: + sample_n = max(1, int(round(len(docs) * h.val_doc_fraction))) + sampled_indices = sorted( + random.Random(h.seed).sample(range(len(docs)), sample_n) + ) + return [(i, docs[i]) for i in sampled_indices] + return doc_entries + + +def train_val_ttt_global_sgd_distributed(h, device, val_data, base_model, val_tokens, batch_seqs=None): + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + base_model.eval() + seq_len = h.eval_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = h.global_ttt_chunk_tokens + batch_seqs = h.global_ttt_batch_seqs if batch_seqs is None else batch_seqs + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + ttt_params = [p for p in base_model.parameters()] + for p in ttt_params: + p.requires_grad_(True) + optimizer = torch.optim.SGD( + ttt_params, lr=h.global_ttt_lr, momentum=h.global_ttt_momentum + ) + t_start = time.perf_counter() + for ci in range(num_chunks): + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + is_last_chunk = ci == num_chunks - 1 + if is_last_chunk or h.global_ttt_epochs <= 0: + continue + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs <= 0: + continue + warmup_chunks = max(0, min(h.global_ttt_warmup_chunks, num_chunks - 1)) + if warmup_chunks > 0 and ci < warmup_chunks: + warmup_denom = max(warmup_chunks - 1, 1) + warmup_t = ci / warmup_denom + lr_now = ( + h.global_ttt_warmup_start_lr + + (h.global_ttt_lr - h.global_ttt_warmup_start_lr) * warmup_t + ) + else: + decay_steps = max(num_chunks - 1 - warmup_chunks, 1) + decay_ci = max(ci - warmup_chunks, 0) + lr_now = h.global_ttt_lr * 0.5 * ( + 1.0 + math.cos(math.pi * decay_ci / decay_steps) + ) + for pg in optimizer.param_groups: + pg["lr"] = lr_now + my_seq_s = chunk_seqs * h.rank // h.world_size + my_seq_e = chunk_seqs * (h.rank + 1) // h.world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ in range(h.global_ttt_epochs): + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x_flat = local[:-1] + y_flat = local[1:] + optimizer.zero_grad(set_to_none=True) + with torch.enable_grad(): + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + if h.global_ttt_respect_doc_boundaries: + bos_pos = (x_flat == BOS_ID).nonzero(as_tuple=True)[0].tolist() + cu_seqlens, max_seqlen = _build_cu_seqlens( + bos_pos, x_flat.numel(), x_flat.device, h.eval_seq_len, 64 + ) + loss = base_model( + x_flat[None], + y_flat[None], + cu_seqlens=cu_seqlens, + max_seqlen=max_seqlen, + ) + else: + x = x_flat.reshape(-1, seq_len) + y = y_flat.reshape(-1, seq_len) + loss = base_model(x, y) + loss.backward() + if dist.is_available() and dist.is_initialized(): + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.SUM) + p.grad.mul_(1.0 / h.world_size) + if h.global_ttt_grad_clip > 0: + torch.nn.utils.clip_grad_norm_(ttt_params, h.global_ttt_grad_clip) + optimizer.step() + base_model.eval() + if h.rank == 0: + elapsed = time.perf_counter() - t_start + log( + f"tttg: c{ci+1}/{num_chunks} lr:{lr_now:.6f} t:{elapsed:.1f}s" + ) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + +def eval_val_ttt_phased(h, base_model, device, val_data, forward_ttt_train): + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(False) + all_tokens = val_data.val_tokens + all_tokens_idx = all_tokens.to(torch.int32) + docs = _find_docs(all_tokens) + doc_entries = _select_ttt_doc_entries(docs, h) + prefix_doc_limit = max(0, min(len(doc_entries), int(h.phased_ttt_prefix_docs))) + num_phases = max(1, int(h.phased_ttt_num_phases)) + phase_boundaries = [] + for pi in range(num_phases): + boundary = prefix_doc_limit * (pi + 1) // num_phases + phase_boundaries.append(boundary) + current_phase = 0 + current_phase_boundary = phase_boundaries[0] + log( + "ttt_phased:" + f" total_docs:{len(doc_entries)} prefix_docs:{prefix_doc_limit} " + f"suffix_docs:{len(doc_entries) - prefix_doc_limit}" + f" num_phases:{num_phases} boundaries:{phase_boundaries}" + ) + chunk_size, eval_seq_len = h.ttt_chunk_size, h.ttt_eval_seq_len + eval_batch_set = None + if h.ttt_eval_batches: + eval_batch_set = set(int(x) for x in h.ttt_eval_batches.split(",") if x.strip()) + use_ascending = eval_batch_set is not None + global_batches_sorted = _build_ttt_global_batches( + doc_entries, h, ascending=use_ascending + ) + queue_len = len(global_batches_sorted) + counter_path = f"/tmp/ttt_counter_{h.run_id}" + prefix_counter_path = f"/tmp/ttt_prefix_counter_{h.run_id}" + pause_flag_path = f"/tmp/ttt_pause_flag_{h.run_id}" + if h.rank == 0: + _init_batch_counter(counter_path) + _init_int64_counter(prefix_counter_path) + try: + os.remove(pause_flag_path) + except FileNotFoundError: + pass + if dist.is_available() and dist.is_initialized(): + path_list = [counter_path, prefix_counter_path, pause_flag_path] + dist.broadcast_object_list(path_list, src=0) + counter_path, prefix_counter_path, pause_flag_path = path_list + dist.barrier() + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + byte_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + t_start = time.perf_counter() + reusable_lora = BatchedTTTLoRA( + h.ttt_batch_size, base_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + + def _build_opt(lora): + if h.ttt_optimizer == "sgd": + return torch.optim.SGD( + lora.parameters(), lr=h.ttt_lora_lr, + momentum=h.ttt_beta1, weight_decay=h.ttt_weight_decay, + ) + return torch.optim.AdamW( + lora.parameters(), lr=h.ttt_lora_lr, + betas=(h.ttt_beta1, h.ttt_beta2), + eps=1e-10, weight_decay=h.ttt_weight_decay, fused=True, + ) + + reusable_opt = _build_opt(reusable_lora) + local_scored_docs = [] + global_ttt_done = prefix_doc_limit == 0 + try: + while True: + queue_idx = _claim_next_batch(counter_path, queue_len) + if queue_idx >= queue_len: + break + orig_batch_idx, batch_entries = global_batches_sorted[queue_idx] + batch = [doc for _, doc in batch_entries] + bsz = len(batch) + prev_loss = loss_sum.item() + prev_bytes = byte_sum.item() + prev_tokens = token_count.item() + if bsz == reusable_lora.bsz: + reusable_lora.reset() + for s in reusable_opt.state.values(): + for k, v in s.items(): + if isinstance(v, torch.Tensor): + v.zero_() + elif k == "step": + s[k] = 0 + cur_lora = reusable_lora + cur_opt = reusable_opt + else: + cur_lora = BatchedTTTLoRA( + bsz, base_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + cur_opt = _build_opt(cur_lora) + pred_lens = [doc_len - 1 for _, doc_len in batch] + num_chunks = [(pl + chunk_size - 1) // chunk_size for pl in pred_lens] + max_nc = max(num_chunks) + num_chunks_t = torch.tensor(num_chunks, dtype=torch.int64, device=device) + for ci in range(max_nc): + active = [ci < nc for nc in num_chunks] + needs_train = any(ci < nc - 1 for nc in num_chunks) + tok_starts = torch.zeros(bsz, dtype=torch.int64) + tok_wls = torch.zeros(bsz, dtype=torch.int64) + chunk_offsets_cpu = torch.zeros(bsz, dtype=torch.int64) + chunk_lens_cpu = torch.zeros(bsz, dtype=torch.int64) + for b in range(bsz): + if not active[b]: + continue + doc_start, doc_len = batch[b] + win_start, win_len, chunk_offset, chunk_len = _compute_chunk_window( + ci, pred_lens[b], num_chunks[b], chunk_size, eval_seq_len + ) + tok_starts[b] = doc_start + win_start + tok_wls[b] = win_len + chunk_offsets_cpu[b] = chunk_offset + chunk_lens_cpu[b] = chunk_len + _, context_size, chunk_offset, _ = _compute_chunk_window( + ci, (ci + 1) * chunk_size, ci + 1, chunk_size, eval_seq_len + ) + col_idx = torch.arange(context_size + 1) + idx = tok_starts.unsqueeze(1) + col_idx.unsqueeze(0) + idx.clamp_(max=all_tokens.numel() - 1) + gathered_gpu = all_tokens_idx[idx].to( + device=device, dtype=torch.int64, non_blocking=True + ) + valid = (col_idx[:context_size].unsqueeze(0) < tok_wls.unsqueeze(1)).to( + device, non_blocking=True + ) + chunk_offsets = chunk_offsets_cpu.to(device, non_blocking=True) + chunk_lens = chunk_lens_cpu.to(device, non_blocking=True) + x = torch.where(valid, gathered_gpu[:, :context_size], 0) + y = torch.where(valid, gathered_gpu[:, 1 : context_size + 1], 0) + ctx_pos = torch.arange(context_size, device=device, dtype=torch.int64) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + per_tok_loss = forward_ttt_train(x, y, lora=cur_lora) + # CaseOps sidecar-driven byte budget. Mirror the index pattern + # used to build y from all_tokens: y[b, j] corresponds to the + # token at global position tok_starts[b] + 1 + j (when valid). + y_bytes_arg = None + if val_data.caseops_enabled and val_data.val_bytes is not None: + y_idx = ( + tok_starts.unsqueeze(1) + + 1 + + col_idx[:context_size].unsqueeze(0) + ) + y_idx = y_idx.clamp_(max=val_data.val_bytes.numel() - 1) + y_bytes_arg = val_data.val_bytes[y_idx].to( + device=device, dtype=torch.int32, non_blocking=True + ) + # Mirror the `valid` masking used for y so out-of-range tokens + # contribute zero bytes (matches y=0 substitution above). + y_bytes_arg = torch.where( + valid, y_bytes_arg, torch.zeros_like(y_bytes_arg) + ) + with torch.no_grad(): + _accumulate_bpb( + per_tok_loss, + x, + y, + chunk_offsets, + chunk_lens, + ctx_pos, + val_data.base_bytes_lut, + val_data.has_leading_space_lut, + val_data.is_boundary_token_lut, + loss_sum, + byte_sum, + token_count, + y_bytes=y_bytes_arg, + ) + if needs_train: + activate_chunk_mask = (num_chunks_t - 1 > ci).float() + for gi in range(h.ttt_grad_steps): + if gi > 0: + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + per_tok_loss = forward_ttt_train(x, y, lora=cur_lora) + per_doc = per_tok_loss[ + :, chunk_offset : chunk_offset + chunk_size + ].mean(dim=-1) + cur_opt.zero_grad(set_to_none=True) + (per_doc * activate_chunk_mask).sum().backward() + cur_opt.step() + else: + del per_tok_loss + batch_num = orig_batch_idx + 1 + doc_lens = [dl for _, dl in batch] + should_report = batch_num in eval_batch_set if eval_batch_set is not None else True + if should_report: + cur_tokens = token_count.item() + cur_loss_val = loss_sum.item() + cur_bytes_val = byte_sum.item() + dt = cur_tokens - prev_tokens + db = cur_bytes_val - prev_bytes + if dt > 0 and db > 0: + b_loss = (cur_loss_val - prev_loss) / dt + b_bpb = b_loss / math.log(2.0) * (dt / db) + else: + b_loss = b_bpb = 0.0 + r_loss = cur_loss_val / max(cur_tokens, 1) + r_bpb = r_loss / math.log(2.0) * (cur_tokens / max(cur_bytes_val, 1)) + elapsed = time.perf_counter() - t_start + log( + f"ttp: b{batch_num}/{queue_len} bl:{b_loss:.4f} bb:{b_bpb:.4f} " + f"rl:{r_loss:.4f} rb:{r_bpb:.4f} dl:{min(doc_lens)}-{max(doc_lens)} " + f"gd:{int(global_ttt_done)}" + ) + if not global_ttt_done: + local_scored_docs.extend( + (orig_batch_idx, pos, doc_start, doc_len) + for pos, (doc_start, doc_len) in enumerate(batch) + ) + prefix_done = _add_to_counter(prefix_counter_path, len(batch_entries)) + if prefix_done >= current_phase_boundary: + try: + with open(pause_flag_path, "x"): + pass + except FileExistsError: + pass + should_pause = os.path.exists(pause_flag_path) + if should_pause: + if dist.is_available() and dist.is_initialized(): + dist.barrier() + gathered_scored_docs = [None] * h.world_size + if dist.is_available() and dist.is_initialized(): + dist.all_gather_object(gathered_scored_docs, local_scored_docs) + else: + gathered_scored_docs = [local_scored_docs] + scored_docs_for_global = [] + for rank_docs in gathered_scored_docs: + if rank_docs: + scored_docs_for_global.extend(rank_docs) + scored_docs_for_global.sort(key=lambda x: (x[0], x[1])) + scored_docs_for_global = scored_docs_for_global[:current_phase_boundary] + scored_token_chunks = [ + val_data.val_tokens[doc_start : doc_start + doc_len] + for _, _, doc_start, doc_len in scored_docs_for_global + ] + if scored_token_chunks: + global_ttt_tokens = torch.cat(scored_token_chunks) + else: + global_ttt_tokens = val_data.val_tokens[:0] + if h.rank == 0: + prefix_done = 0 + try: + with open(prefix_counter_path, "rb") as f: + prefix_done = int.from_bytes( + f.read(8), "little", signed=True + ) + except FileNotFoundError: + pass + log( + f"ttpp: phase:{current_phase + 1}/{num_phases} pd:{prefix_done} " + f"gd:{len(scored_docs_for_global)} " + f"t:{time.perf_counter() - t_start:.1f}s" + ) + train_val_ttt_global_sgd_distributed( + h, device, val_data, base_model, global_ttt_tokens + ) + for p in base_model.parameters(): + p.requires_grad_(False) + reusable_lora = BatchedTTTLoRA( + h.ttt_batch_size, base_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + reusable_opt = _build_opt(reusable_lora) + current_phase += 1 + if current_phase >= num_phases: + global_ttt_done = True + else: + current_phase_boundary = phase_boundaries[current_phase] + if h.rank == 0: + try: + os.remove(pause_flag_path) + except FileNotFoundError: + pass + if dist.is_available() and dist.is_initialized(): + dist.barrier() + if h.rank == 0: + log(f"ttpr: phase:{current_phase}/{num_phases} t:{time.perf_counter() - t_start:.1f}s") + del cur_lora, cur_opt + finally: + pass + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.train() + return _loss_bpb_from_sums(loss_sum, token_count, byte_sum) + + +def timed_eval(label, fn, *args, **kwargs): + torch.cuda.synchronize() + t0 = time.perf_counter() + val_loss, val_bpb = fn(*args, **kwargs) + torch.cuda.synchronize() + elapsed_ms = 1e3 * (time.perf_counter() - t0) + log( + f"{label} val_loss:{val_loss:.8f} val_bpb:{val_bpb:.8f} eval_time:{elapsed_ms:.0f}ms" + ) + return val_loss, val_bpb + + +def train_model(h, device, val_data): + base_model = GPT(h).to(device).bfloat16() + restore_fp32_params(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + compiled_forward_logits = torch.compile( + base_model.forward_logits, dynamic=False, fullgraph=True + ) + model = compiled_model + log(f"model_params:{sum(p.numel()for p in base_model.parameters())}") + optimizers = Optimizers(h, base_model) + train_loader = DocumentPackingLoader(h, device) + max_wallclock_ms = ( + 1e3 * h.max_wallclock_seconds if h.max_wallclock_seconds > 0 else None + ) + if max_wallclock_ms is not None: + max_wallclock_ms -= h.gptq_reserve_seconds * 1e3 + log( + f"gptq:reserving {h.gptq_reserve_seconds:.0f}s, effective={max_wallclock_ms:.0f}ms" + ) + + def training_frac(step, elapsed_ms): + if max_wallclock_ms is None: + return step / max(h.iterations, 1) + return elapsed_ms / max(max_wallclock_ms, 1e-09) + + def lr_mul(frac): + if h.warmdown_frac <= 0: + return 1.0 + if frac >= 1.0 - h.warmdown_frac: + return max((1.0 - frac) / h.warmdown_frac, h.min_lr) + return 1.0 + + def step_fn(step, lr_scale): + optimizers.zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(h.grad_accum_steps): + x, y, cu_seqlens, _max_seqlen = train_loader.next_batch( + h.train_batch_tokens, h.grad_accum_steps + ) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y, cu_seqlens=cu_seqlens, max_seqlen=h.train_seq_len) + train_loss += loss.detach() + (loss / h.grad_accum_steps).backward() + train_loss /= h.grad_accum_steps + frac = ( + min(step / h.muon_momentum_warmup_steps, 1.0) + if h.muon_momentum_warmup_steps > 0 + else 1.0 + ) + muon_momentum = ( + 1 - frac + ) * h.muon_momentum_warmup_start + frac * h.muon_momentum + for group in optimizers.optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * lr_scale + if h.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), h.grad_clip_norm) + optimizers.step(distributed=h.distributed) + return train_loss + + if h.warmup_steps > 0: + initial_model_state = { + name: tensor.detach().cpu().clone() + for (name, tensor) in base_model.state_dict().items() + } + initial_optimizer_states = [ + copy.deepcopy(opt.state_dict()) for opt in optimizers + ] + model.train() + num_tokens_local = h.train_batch_tokens // h.world_size + for blk in base_model.blocks: + blk.attn.rotary(num_tokens_local, device, torch.bfloat16) + cu_bucket_size = train_loader.cu_bucket_size + warmup_cu_buckets = tuple(cu_bucket_size * i for i in range(1, 5)) + warmup_cu_iters = 3 + x, y, cu_seqlens, _ = train_loader.next_batch( + h.train_batch_tokens, h.grad_accum_steps + ) + log(f"warmup_cu_buckets:{','.join(str(b) for b in warmup_cu_buckets)} iters_each:{warmup_cu_iters}") + def _run_cu_bucket_warmup(): + for bucket_len in warmup_cu_buckets: + boundaries = list(range(0, x.size(1), max(h.train_seq_len, 1))) + if boundaries[-1] != x.size(1): + boundaries.append(x.size(1)) + cu = torch.full((bucket_len,), x.size(1), dtype=torch.int32, device=device) + cu[: len(boundaries)] = torch.tensor(boundaries, dtype=torch.int32, device=device) + for _ in range(warmup_cu_iters): + optimizers.zero_grad_all() + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + wloss = model(x, y, cu_seqlens=cu, max_seqlen=h.train_seq_len) + (wloss / h.grad_accum_steps).backward() + optimizers.zero_grad_all() + _run_cu_bucket_warmup() + if h.num_loops > 0: + base_model.looping_active = True + _run_cu_bucket_warmup() + base_model.looping_active = False + for warmup_step in range(h.warmup_steps): + step_fn(warmup_step, 1.0) + if ( + warmup_step <= 5 + or (warmup_step + 1) % 10 == 0 + or warmup_step + 1 == h.warmup_steps + ): + log(f"warmup_step: {warmup_step+1}/{h.warmup_steps}") + if h.num_loops > 0: + base_model.looping_active = True + log( + f"loop_warmup:enabled encoder:{base_model.encoder_indices} decoder:{base_model.decoder_indices}" + ) + for warmup_step in range(h.warmup_steps): + step_fn(warmup_step, 1.0) + if ( + warmup_step <= 5 + or (warmup_step + 1) % 10 == 0 + or warmup_step + 1 == h.warmup_steps + ): + log(f"loop_warmup_step: {warmup_step+1}/{h.warmup_steps}") + base_model.looping_active = False + base_model.load_state_dict(initial_model_state, strict=True) + for (opt, state) in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + optimizers.zero_grad_all() + train_loader = DocumentPackingLoader(h, device) + ema_state = { + name: t.detach().float().clone() + for (name, t) in base_model.state_dict().items() + } + ema_decay = h.ema_decay + training_time_ms = 0.0 + stop_after_step = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = ( + step == h.iterations + or stop_after_step is not None + and step >= stop_after_step + ) + should_validate = ( + last_step or h.val_loss_every > 0 and step % h.val_loss_every == 0 + ) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1e3 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + h, device, val_data, model, compiled_forward_logits + ) + log( + f"{step}/{h.iterations} val_loss: {val_loss:.4f} val_bpb: {val_bpb:.4f}" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < h.iterations: + log( + f"stopping_early: wallclock_cap train_time: {training_time_ms:.0f}ms step: {step}/{h.iterations}" + ) + break + elapsed_ms = training_time_ms + 1e3 * (time.perf_counter() - t0) + frac = training_frac(step, elapsed_ms) + scale = lr_mul(frac) + if ( + h.num_loops > 0 + and not base_model.looping_active + and frac >= h.enable_looping_at + ): + base_model.looping_active = True + log( + f"layer_loop:enabled step:{step} frac:{frac:.3f} encoder:{base_model.encoder_indices} decoder:{base_model.decoder_indices}" + ) + train_loss = step_fn(step, scale) + with torch.no_grad(): + for (name, t) in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_( + t.detach().float(), alpha=1.0 - ema_decay + ) + step += 1 + approx_training_time_ms = training_time_ms + 1e3 * (time.perf_counter() - t0) + should_log_train = h.train_log_every > 0 and ( + step <= 5 or step % h.train_log_every == 0 or stop_after_step is not None + ) + if should_log_train: + tok_per_sec = step * h.train_batch_tokens / (approx_training_time_ms / 1e3) + log( + f"{step}/{h.iterations} train_loss: {train_loss.item():.4f} train_time: {approx_training_time_ms/60000:.1f}m tok/s: {tok_per_sec:.0f}" + ) + reached_cap = ( + max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + ) + if h.distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log( + f"peak memory allocated: {torch.cuda.max_memory_allocated()//1024//1024} MiB reserved: {torch.cuda.max_memory_reserved()//1024//1024} MiB" + ) + log("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = { + name: t.to(dtype=current_state[name].dtype) for (name, t) in ema_state.items() + } + base_model.load_state_dict(avg_state, strict=True) + return base_model, compiled_model, compiled_forward_logits + + +def train_and_eval(h, device): + random.seed(h.seed) + np.random.seed(h.seed) + torch.manual_seed(h.seed) + torch.cuda.manual_seed_all(h.seed) + if h.artifact_dir and h.is_main_process: + os.makedirs(h.artifact_dir, exist_ok=True) + val_data = ValidationData(h, device) + log( + f"train_shards: {len(list(Path(h.datasets_dir).resolve().glob('fineweb_train_*.bin')))}" + ) + log(f"val_tokens: {val_data.val_tokens.numel()-1}") + # TTT_EVAL_ONLY: skip training + GPTQ, jump straight to TTT eval on a + # pre-existing quantized artifact. Used to test TTT-only improvements + # (e.g., PR-1767's alpha/warm-start/WD) without retraining. + ttt_eval_only = os.environ.get("TTT_EVAL_ONLY", "0") == "1" + if ttt_eval_only: + log("TTT_EVAL_ONLY=1 — skipping training + GPTQ, loading saved artifact for TTT eval") + log(f"ttt_lora_alpha: {BatchedLinearLoRA._ALPHA}") + log(f"ttt_warm_start_a: {BatchedLinearLoRA._WARM_START_A}") + log(f"ttt_weight_decay: {h.ttt_weight_decay}") + else: + base_model, compiled_model, compiled_forward_logits = train_model( + h, device, val_data + ) + torch._dynamo.reset() + timed_eval( + "diagnostic pre-quantization post-ema", + eval_val, + h, + device, + val_data, + compiled_model, + compiled_forward_logits, + ) + if os.environ.get("PREQUANT_ONLY", "0") == "1": + log("PREQUANT_ONLY=1 — skipping serialize/GPTQ/post-quant eval/TTT") + return + serialize(h, base_model, Path(__file__).read_text(encoding="utf-8")) + if h.distributed: + dist.barrier() + eval_model = deserialize(h, device) + if h.num_loops > 0: + eval_model.looping_active = True + if not ttt_eval_only: + compiled_model = torch.compile(eval_model, dynamic=False, fullgraph=True) + compiled_forward_logits = torch.compile( + eval_model.forward_logits, dynamic=False, fullgraph=True + ) + timed_eval( + "diagnostic quantized", + eval_val, + h, + device, + val_data, + compiled_model, + compiled_forward_logits, + ) + del eval_model + if h.ttt_enabled: + if not ttt_eval_only: + del compiled_model + if ttt_eval_only: + del eval_model + torch._dynamo.reset() + torch.cuda.empty_cache() + ttt_model = deserialize(h, device) + if h.num_loops > 0: + ttt_model.looping_active = True + for p in ttt_model.parameters(): + p.requires_grad_(False) + + if h.rope_yarn: + _yarn_seqlen = h.train_batch_tokens // h.grad_accum_steps + for block in ttt_model.blocks: + block.attn.rotary(_yarn_seqlen, device, torch.bfloat16) + else: + for block in ttt_model.blocks: + block.attn.rotary._cos_cached = None + block.attn.rotary._sin_cached = None + block.attn.rotary._seq_len_cached = 0 + block.attn.rotary(h.ttt_eval_seq_len, device, torch.bfloat16) + + def _fwd_ttt_inner(input_ids, target_ids, lora): + return ttt_model.forward_ttt(input_ids, target_ids, lora=lora) + + _fwd_ttt_compiled_inner = None + + def _fwd_ttt(input_ids, target_ids, lora): + nonlocal _fwd_ttt_compiled_inner + if _fwd_ttt_compiled_inner is None: + _fwd_ttt_compiled_inner = torch.compile(_fwd_ttt_inner, dynamic=True) + return _fwd_ttt_compiled_inner(input_ids, target_ids, lora=lora) + + fwd_ttt_compiled = _fwd_ttt + log(f"ttt_lora:warming up compile (random tokens, no val data)") + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + t_warmup = time.perf_counter() + warmup_bszes = [h.ttt_batch_size] + for bsz in warmup_bszes: + wl = BatchedTTTLoRA( + bsz, ttt_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + wo = torch.optim.AdamW( + wl.parameters(), + lr=h.ttt_lora_lr, + betas=(h.ttt_beta1, h.ttt_beta2), + eps=1e-10, + weight_decay=h.ttt_weight_decay, + fused=True, + ) + for ctx_len in (h.ttt_chunk_size, h.ttt_eval_seq_len): + xw = torch.randint(0, h.vocab_size, (bsz, ctx_len), device=device, dtype=torch.int64) + yw = torch.randint(0, h.vocab_size, (bsz, ctx_len), device=device, dtype=torch.int64) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + ptl = fwd_ttt_compiled(xw, yw, lora=wl) + ptl[:, : min(h.ttt_chunk_size, ctx_len)].mean(dim=-1).sum().backward() + wo.step() + wo.zero_grad(set_to_none=True) + del wl, wo + torch.cuda.empty_cache() + compile_elapsed = time.perf_counter() - t_warmup + log(f"ttt_lora:compile warmup done ({compile_elapsed:.1f}s)") + log("\nbeginning TTT eval timer") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_ttt_phased( + h, ttt_model, device, val_data, forward_ttt_train=fwd_ttt_compiled + ) + torch.cuda.synchronize() + ttt_eval_elapsed = time.perf_counter() - t_ttt + log( + "quantized_ttt_phased " + f"val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f} " + f"eval_time:{1e3*ttt_eval_elapsed:.0f}ms" + ) + log(f"total_eval_time:{ttt_eval_elapsed:.1f}s") + del ttt_model + + +def main(): + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError( + f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral" + ) + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + torch.set_float32_matmul_precision("high") + from torch.backends.cuda import ( + enable_cudnn_sdp, + enable_flash_sdp, + enable_math_sdp, + enable_mem_efficient_sdp, + ) + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + torch._dynamo.config.optimize_ddp = False + torch._dynamo.config.cache_size_limit = 16 + h = Hyperparameters() + set_logging_hparams(h) + if h.is_main_process: + os.makedirs(h.artifact_dir if h.artifact_dir else "logs", exist_ok=True) + log(100 * "=", console=False) + log("Hyperparameters:", console=True) + for (k, v) in sorted(vars(type(h)).items()): + if not k.startswith("_"): + log(f" {k}: {v}", console=True) + log("=" * 100, console=False) + log("Source code:", console=False) + log("=" * 100, console=False) + with open(__file__, "r", encoding="utf-8") as _src: + log(_src.read(), console=False) + log("=" * 100, console=False) + log(f"Running Python {sys.version}", console=False) + log(f"Running PyTorch {torch.__version__}", console=False) + log("=" * 100, console=False) + train_and_eval(h, device) + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/train_seed42.log b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/train_seed42.log new file mode 100644 index 0000000000..e4e34a2534 --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_SmearGateBOSFix_PR1787Base_LQERAsym_PhasedTTT/train_seed42.log @@ -0,0 +1,833 @@ +nohup: ignoring input +W0427 04:43:16.977000 185990 torch/distributed/run.py:803] +W0427 04:43:16.977000 185990 torch/distributed/run.py:803] ***************************************** +W0427 04:43:16.977000 185990 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0427 04:43:16.977000 185990 torch/distributed/run.py:803] ***************************************** +Hyperparameters: + adam_eps: 1e-08 + adam_wd: 0.02 + artifact_dir: + attn_clip_sigmas: 13.0 + attn_out_gate_enabled: False + attn_out_gate_src: proj + beta1: 0.9 + beta2: 0.95 + caseops_enabled: True + compressor: brotli + data_dir: ./data/ + datasets_dir: /dev/shm/pgolf_data + distributed: True + ema_decay: 0.9965 + embed_bits: 7 + embed_clip_sigmas: 15.0 + embed_lr: 0.6 + embed_wd: 0.085 + enable_looping_at: 0.35 + eval_seq_len: 2048 + eval_stride: 64 + fused_ce_enabled: True + gate_window: 12 + gated_attn_enabled: False + gated_attn_init_std: 0.01 + gated_attn_quant_gate: False + global_ttt_batch_seqs: 32 + global_ttt_chunk_tokens: 32768 + global_ttt_epochs: 1 + global_ttt_grad_clip: 1.0 + global_ttt_lr: 0.001 + global_ttt_momentum: 0.9 + global_ttt_respect_doc_boundaries: True + global_ttt_warmup_chunks: 0 + global_ttt_warmup_start_lr: 0.0 + gptq_calibration_batches: 16 + gptq_reserve_seconds: 0.5 + grad_accum_steps: 1 + grad_clip_norm: 0.3 + is_main_process: True + iterations: 20000 + ln_scale: True + local_rank: 0 + logfile: logs/340ad5c9-cc29-4e8b-918a-9786cbc866be.txt + logit_softcap: 30.0 + loop_end: 5 + loop_start: 3 + lqer_asym_enabled: True + lqer_asym_group: 64 + lqer_enabled: True + lqer_factor_bits: 4 + lqer_rank: 4 + lqer_top_k: 3 + matrix_bits: 6 + matrix_clip_sigmas: 12.85 + matrix_lr: 0.026 + max_wallclock_seconds: 600.0 + min_lr: 0.1 + mlp_clip_sigmas: 12.0 + mlp_mult: 4.0 + model_dim: 512 + model_path: final_model.pt + muon_backend_steps: 5 + muon_momentum: 0.97 + muon_momentum_warmup_start: 0.92 + muon_momentum_warmup_steps: 1500 + muon_row_normalize: True + muon_wd: 0.095 + num_heads: 8 + num_kv_heads: 4 + num_layers: 11 + num_loops: 2 + parallel_final_lane: mean + parallel_start_layer: 8 + phased_ttt_num_phases: 3 + phased_ttt_prefix_docs: 2000 + qk_gain_init: 5.0 + quantized_model_path: final_model.int6.ptz + rank: 0 + rope_base: 10000.0 + rope_dims: 16 + rope_train_seq_len: 2048 + rope_yarn: False + run_id: 340ad5c9-cc29-4e8b-918a-9786cbc866be + scalar_lr: 0.02 + seed: 42 + skip_gates_enabled: True + smear_gate_enabled: True + sparse_attn_gate_enabled: True + sparse_attn_gate_init_std: 0.0 + sparse_attn_gate_scale: 1.0 + tie_embeddings: True + tied_embed_init_std: 0.005 + tied_embed_lr: 0.03 + tokenizer_path: /dev/shm/pgolf_tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model + train_batch_tokens: 786432 + train_files: /dev/shm/pgolf_data/fineweb_train_*.bin + train_log_every: 500 + train_seq_len: 2048 + ttt_batch_size: 64 + ttt_beta1: 0.0 + ttt_beta2: 0.999 + ttt_chunk_size: 48 + ttt_enabled: True + ttt_eval_batches: + ttt_eval_seq_len: 2048 + ttt_grad_steps: 1 + ttt_k_lora: True + ttt_lora_lr: 0.0001 + ttt_lora_rank: 96 + ttt_mlp_lora: True + ttt_o_lora: True + ttt_optimizer: adam + ttt_weight_decay: 1.0 + val_batch_tokens: 524288 + val_bytes_files: /dev/shm/pgolf_data/fineweb_val_bytes_*.bin + val_doc_fraction: 1.0 + val_files: /dev/shm/pgolf_data/fineweb_val_*.bin + val_loss_every: 4000 + vocab_size: 8192 + warmdown_frac: 0.75 + warmup_steps: 20 + world_size: 8 + xsa_last_n: 11 +train_shards: 39 +val_tokens: 47851520 +model_params:35945671 +gptq:reserving 0s, effective=599500ms +warmup_cu_buckets:64,128,192,256 iters_each:3 +warmup_step: 1/20 +warmup_step: 2/20 +warmup_step: 3/20 +warmup_step: 4/20 +warmup_step: 5/20 +warmup_step: 6/20 +warmup_step: 10/20 +warmup_step: 20/20 +loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +loop_warmup_step: 1/20 +loop_warmup_step: 2/20 +loop_warmup_step: 3/20 +loop_warmup_step: 4/20 +loop_warmup_step: 5/20 +loop_warmup_step: 6/20 +loop_warmup_step: 10/20 +loop_warmup_step: 20/20 +0/20000 val_loss: 9.0076 val_bpb: 4.1159 +1/20000 train_loss: 9.0087 train_time: 0.0m tok/s: 12535509 +2/20000 train_loss: 12.8485 train_time: 0.0m tok/s: 10830895 +3/20000 train_loss: 10.2683 train_time: 0.0m tok/s: 9514405 +4/20000 train_loss: 8.7315 train_time: 0.0m tok/s: 8959951 +5/20000 train_loss: 7.9453 train_time: 0.0m tok/s: 8662131 +500/20000 train_loss: 2.5698 train_time: 0.8m tok/s: 8216992 +1000/20000 train_loss: 2.8006 train_time: 1.6m tok/s: 8178350 +1500/20000 train_loss: 2.6344 train_time: 2.4m tok/s: 8163805 +2000/20000 train_loss: 2.6634 train_time: 3.2m tok/s: 8157405 +layer_loop:enabled step:2173 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +2500/20000 train_loss: 2.5534 train_time: 4.3m tok/s: 7653758 +3000/20000 train_loss: 2.5637 train_time: 5.5m tok/s: 7164309 +3500/20000 train_loss: 2.5657 train_time: 6.7m tok/s: 6864774 +4000/20000 train_loss: 2.4112 train_time: 7.9m tok/s: 6658543 +4000/20000 val_loss: 2.4294 val_bpb: 1.1101 +4500/20000 train_loss: 2.2778 train_time: 9.1m tok/s: 6516065 +4900/20000 val_loss: 2.3541 val_bpb: 1.0757 +stopping_early: wallclock_cap train_time: 599623ms step: 4900/20000 +peak memory allocated: 41709 MiB reserved: 47026 MiB +ema:applying EMA weights +diagnostic pre-quantization post-ema val_loss:2.33055251 val_bpb:1.06490240 eval_time:7779ms +Serialized model: 135417533 bytes +Code size (uncompressed): 151766 bytes +Code size (compressed): 31250 bytes +GPTQ:collecting Hessians from calibration data... +GPTQ:collected 67 Hessians in 3.5s +Quantized weights: + gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight + gptq (int6)+lqer_asym: blocks.mlp.fc.weight + gptq (int7)+lqer_asym: tok_emb.weight + passthrough (float16): blocks.attn.attn_gate_w, blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, parallel_post_lambdas, parallel_resid_lambdas, skip_gates, skip_weights, smear_gate.weight, smear_lambda +Serialized model quantized+brotli: 15920836 bytes +Total submission size quantized+brotli: 15952086 bytes +diagnostic quantized val_loss:2.35058660 val_bpb:1.07405660 eval_time:63234ms +ttt_lora:warming up compile (random tokens, no val data) +ttt_lora:compile warmup done (156.2s) + +beginning TTT eval timer +ttt_phased: total_docs:50000 prefix_docs:2000 suffix_docs:48000 num_phases:3 boundaries:[666, 1333, 2000] +ttp: b775/782 bl:2.2782 bb:1.0591 rl:2.2782 rb:1.0591 dl:6892-7524 gd:0 +ttp: b771/782 bl:2.3090 bb:1.0606 rl:2.2918 rb:1.0598 dl:5523-5749 gd:0 +ttp: b765/782 bl:2.3185 bb:1.0843 rl:2.2987 rb:1.0660 dl:4393-4510 gd:0 +ttpp: phase:1/3 pd:1104 gd:666 t:218.0s +tttg: c1/111 lr:0.001000 t:2.4s +tttg: c2/111 lr:0.001000 t:2.5s +tttg: c3/111 lr:0.000999 t:2.6s +tttg: c4/111 lr:0.000998 t:2.7s +tttg: c5/111 lr:0.000997 t:2.8s +tttg: c6/111 lr:0.000995 t:3.0s +tttg: c7/111 lr:0.000993 t:3.1s +tttg: c8/111 lr:0.000990 t:3.2s +tttg: c9/111 lr:0.000987 t:3.3s +tttg: c10/111 lr:0.000984 t:3.5s +tttg: c11/111 lr:0.000980 t:3.6s +tttg: c12/111 lr:0.000976 t:3.7s +tttg: c13/111 lr:0.000971 t:3.9s +tttg: c14/111 lr:0.000966 t:4.0s +tttg: c15/111 lr:0.000961 t:4.1s +tttg: c16/111 lr:0.000955 t:4.3s +tttg: c17/111 lr:0.000949 t:4.4s +tttg: c18/111 lr:0.000942 t:4.5s +tttg: c19/111 lr:0.000935 t:4.6s +tttg: c20/111 lr:0.000928 t:4.8s +tttg: c21/111 lr:0.000921 t:4.9s +tttg: c22/111 lr:0.000913 t:5.0s +tttg: c23/111 lr:0.000905 t:5.1s +tttg: c24/111 lr:0.000896 t:5.3s +tttg: c25/111 lr:0.000887 t:5.4s +tttg: c26/111 lr:0.000878 t:5.5s +tttg: c27/111 lr:0.000868 t:5.8s +tttg: c28/111 lr:0.000859 t:5.9s +tttg: c29/111 lr:0.000848 t:6.1s +tttg: c30/111 lr:0.000838 t:6.2s +tttg: c31/111 lr:0.000827 t:6.3s +tttg: c32/111 lr:0.000817 t:6.4s +tttg: c33/111 lr:0.000805 t:6.6s +tttg: c34/111 lr:0.000794 t:7.0s +tttg: c35/111 lr:0.000782 t:7.2s +tttg: c36/111 lr:0.000770 t:7.3s +tttg: c37/111 lr:0.000758 t:7.4s +tttg: c38/111 lr:0.000746 t:7.5s +tttg: c39/111 lr:0.000733 t:7.7s +tttg: c40/111 lr:0.000721 t:7.8s +tttg: c41/111 lr:0.000708 t:7.9s +tttg: c42/111 lr:0.000695 t:8.0s +tttg: c43/111 lr:0.000681 t:8.2s +tttg: c44/111 lr:0.000668 t:8.3s +tttg: c45/111 lr:0.000655 t:8.4s +tttg: c46/111 lr:0.000641 t:8.5s +tttg: c47/111 lr:0.000627 t:8.6s +tttg: c48/111 lr:0.000613 t:8.8s +tttg: c49/111 lr:0.000599 t:8.9s +tttg: c50/111 lr:0.000585 t:9.0s +tttg: c51/111 lr:0.000571 t:9.2s +tttg: c52/111 lr:0.000557 t:9.3s +tttg: c53/111 lr:0.000543 t:9.4s +tttg: c54/111 lr:0.000529 t:9.6s +tttg: c55/111 lr:0.000514 t:9.7s +tttg: c56/111 lr:0.000500 t:9.8s +tttg: c57/111 lr:0.000486 t:9.9s +tttg: c58/111 lr:0.000471 t:10.1s +tttg: c59/111 lr:0.000457 t:10.2s +tttg: c60/111 lr:0.000443 t:10.3s +tttg: c61/111 lr:0.000429 t:10.4s +tttg: c62/111 lr:0.000415 t:10.5s +tttg: c63/111 lr:0.000401 t:10.7s +tttg: c64/111 lr:0.000387 t:10.8s +tttg: c65/111 lr:0.000373 t:10.9s +tttg: c66/111 lr:0.000359 t:11.0s +tttg: c67/111 lr:0.000345 t:11.1s +tttg: c68/111 lr:0.000332 t:11.2s +tttg: c69/111 lr:0.000319 t:11.3s +tttg: c70/111 lr:0.000305 t:11.4s +tttg: c71/111 lr:0.000292 t:11.6s +tttg: c72/111 lr:0.000279 t:11.7s +tttg: c73/111 lr:0.000267 t:11.8s +tttg: c74/111 lr:0.000254 t:11.9s +tttg: c75/111 lr:0.000242 t:12.0s +tttg: c76/111 lr:0.000230 t:12.1s +tttg: c77/111 lr:0.000218 t:12.2s +tttg: c78/111 lr:0.000206 t:12.3s +tttg: c79/111 lr:0.000195 t:12.4s +tttg: c80/111 lr:0.000183 t:12.5s +tttg: c81/111 lr:0.000173 t:12.6s +tttg: c82/111 lr:0.000162 t:12.7s +tttg: c83/111 lr:0.000152 t:12.8s +tttg: c84/111 lr:0.000141 t:12.9s +tttg: c85/111 lr:0.000132 t:13.0s +tttg: c86/111 lr:0.000122 t:13.1s +tttg: c87/111 lr:0.000113 t:13.2s +tttg: c88/111 lr:0.000104 t:13.3s +tttg: c89/111 lr:0.000095 t:13.4s +tttg: c90/111 lr:0.000087 t:13.5s +tttg: c91/111 lr:0.000079 t:13.6s +tttg: c92/111 lr:0.000072 t:13.7s +tttg: c93/111 lr:0.000065 t:13.9s +tttg: c94/111 lr:0.000058 t:14.0s +tttg: c95/111 lr:0.000051 t:14.1s +tttg: c96/111 lr:0.000045 t:14.2s +tttg: c97/111 lr:0.000039 t:14.3s +tttg: c98/111 lr:0.000034 t:14.4s +tttg: c99/111 lr:0.000029 t:14.5s +tttg: c100/111 lr:0.000024 t:14.6s +tttg: c101/111 lr:0.000020 t:14.7s +tttg: c102/111 lr:0.000016 t:14.8s +tttg: c103/111 lr:0.000013 t:14.9s +tttg: c104/111 lr:0.000010 t:15.0s +tttg: c105/111 lr:0.000007 t:15.1s +tttg: c106/111 lr:0.000005 t:15.2s +tttg: c107/111 lr:0.000003 t:15.3s +tttg: c108/111 lr:0.000002 t:15.4s +tttg: c109/111 lr:0.000001 t:15.5s +tttg: c110/111 lr:0.000000 t:15.6s +ttpr: phase:1/3 t:236.2s +ttp: b757/782 bl:2.2825 bb:1.0625 rl:2.2959 rb:1.0654 dl:3550-3633 gd:0 +ttpp: phase:2/3 pd:1808 gd:1333 t:340.9s +tttg: c1/185 lr:0.001000 t:0.1s +tttg: c2/185 lr:0.001000 t:0.1s +tttg: c3/185 lr:0.001000 t:0.2s +tttg: c4/185 lr:0.000999 t:0.4s +tttg: c5/185 lr:0.000999 t:0.6s +tttg: c6/185 lr:0.000998 t:0.7s +tttg: c7/185 lr:0.000997 t:0.8s +tttg: c8/185 lr:0.000996 t:0.9s +tttg: c9/185 lr:0.000995 t:1.6s +tttg: c10/185 lr:0.000994 t:1.7s +tttg: c11/185 lr:0.000993 t:1.8s +tttg: c12/185 lr:0.000991 t:2.0s +tttg: c13/185 lr:0.000990 t:2.1s +tttg: c14/185 lr:0.000988 t:2.2s +tttg: c15/185 lr:0.000986 t:2.3s +tttg: c16/185 lr:0.000984 t:2.4s +tttg: c17/185 lr:0.000981 t:2.5s +tttg: c18/185 lr:0.000979 t:2.6s +tttg: c19/185 lr:0.000977 t:2.7s +tttg: c20/185 lr:0.000974 t:2.8s +tttg: c21/185 lr:0.000971 t:2.9s +tttg: c22/185 lr:0.000968 t:3.0s +tttg: c23/185 lr:0.000965 t:3.2s +tttg: c24/185 lr:0.000962 t:3.3s +tttg: c25/185 lr:0.000959 t:3.4s +tttg: c26/185 lr:0.000955 t:3.5s +tttg: c27/185 lr:0.000952 t:3.6s +tttg: c28/185 lr:0.000948 t:4.6s +tttg: c29/185 lr:0.000944 t:4.7s +tttg: c30/185 lr:0.000940 t:4.8s +tttg: c31/185 lr:0.000936 t:4.9s +tttg: c32/185 lr:0.000932 t:5.0s +tttg: c33/185 lr:0.000927 t:5.1s +tttg: c34/185 lr:0.000923 t:5.2s +tttg: c35/185 lr:0.000918 t:5.3s +tttg: c36/185 lr:0.000913 t:5.4s +tttg: c37/185 lr:0.000908 t:5.5s +tttg: c38/185 lr:0.000904 t:5.6s +tttg: c39/185 lr:0.000898 t:5.7s +tttg: c40/185 lr:0.000893 t:5.8s +tttg: c41/185 lr:0.000888 t:5.9s +tttg: c42/185 lr:0.000882 t:6.1s +tttg: c43/185 lr:0.000877 t:6.2s +tttg: c44/185 lr:0.000871 t:6.3s +tttg: c45/185 lr:0.000865 t:6.4s +tttg: c46/185 lr:0.000860 t:6.5s +tttg: c47/185 lr:0.000854 t:6.6s +tttg: c48/185 lr:0.000847 t:6.7s +tttg: c49/185 lr:0.000841 t:6.8s +tttg: c50/185 lr:0.000835 t:6.9s +tttg: c51/185 lr:0.000829 t:7.0s +tttg: c52/185 lr:0.000822 t:7.1s +tttg: c53/185 lr:0.000816 t:7.2s +tttg: c54/185 lr:0.000809 t:7.3s +tttg: c55/185 lr:0.000802 t:7.4s +tttg: c56/185 lr:0.000795 t:7.5s +tttg: c57/185 lr:0.000788 t:7.7s +tttg: c58/185 lr:0.000781 t:7.8s +tttg: c59/185 lr:0.000774 t:7.9s +tttg: c60/185 lr:0.000767 t:8.0s +tttg: c61/185 lr:0.000760 t:8.1s +tttg: c62/185 lr:0.000752 t:8.2s +tttg: c63/185 lr:0.000745 t:8.3s +tttg: c64/185 lr:0.000738 t:8.4s +tttg: c65/185 lr:0.000730 t:8.5s +tttg: c66/185 lr:0.000722 t:8.6s +tttg: c67/185 lr:0.000715 t:8.9s +tttg: c68/185 lr:0.000707 t:9.0s +tttg: c69/185 lr:0.000699 t:9.1s +tttg: c70/185 lr:0.000691 t:9.2s +tttg: c71/185 lr:0.000683 t:9.3s +tttg: c72/185 lr:0.000675 t:9.4s +tttg: c73/185 lr:0.000667 t:9.5s +tttg: c74/185 lr:0.000659 t:9.6s +tttg: c75/185 lr:0.000651 t:9.7s +tttg: c76/185 lr:0.000643 t:9.8s +tttg: c77/185 lr:0.000635 t:9.9s +tttg: c78/185 lr:0.000627 t:10.0s +tttg: c79/185 lr:0.000618 t:10.1s +tttg: c80/185 lr:0.000610 t:10.2s +tttg: c81/185 lr:0.000602 t:10.3s +tttg: c82/185 lr:0.000593 t:10.5s +tttg: c83/185 lr:0.000585 t:10.6s +tttg: c84/185 lr:0.000577 t:10.7s +tttg: c85/185 lr:0.000568 t:10.8s +tttg: c86/185 lr:0.000560 t:10.9s +tttg: c87/185 lr:0.000551 t:11.0s +tttg: c88/185 lr:0.000543 t:11.1s +tttg: c89/185 lr:0.000534 t:11.2s +tttg: c90/185 lr:0.000526 t:11.3s +tttg: c91/185 lr:0.000517 t:11.4s +tttg: c92/185 lr:0.000509 t:11.5s +tttg: c93/185 lr:0.000500 t:11.6s +tttg: c94/185 lr:0.000491 t:11.7s +tttg: c95/185 lr:0.000483 t:11.8s +tttg: c96/185 lr:0.000474 t:11.9s +tttg: c97/185 lr:0.000466 t:12.0s +tttg: c98/185 lr:0.000457 t:12.1s +tttg: c99/185 lr:0.000449 t:12.2s +tttg: c100/185 lr:0.000440 t:12.3s +tttg: c101/185 lr:0.000432 t:12.4s +tttg: c102/185 lr:0.000423 t:12.6s +tttg: c103/185 lr:0.000415 t:12.7s +tttg: c104/185 lr:0.000407 t:12.8s +tttg: c105/185 lr:0.000398 t:12.9s +tttg: c106/185 lr:0.000390 t:13.0s +tttg: c107/185 lr:0.000382 t:13.1s +tttg: c108/185 lr:0.000373 t:13.2s +tttg: c109/185 lr:0.000365 t:13.3s +tttg: c110/185 lr:0.000357 t:13.4s +tttg: c111/185 lr:0.000349 t:13.5s +tttg: c112/185 lr:0.000341 t:13.6s +tttg: c113/185 lr:0.000333 t:15.5s +tttg: c114/185 lr:0.000325 t:15.6s +tttg: c115/185 lr:0.000317 t:15.7s +tttg: c116/185 lr:0.000309 t:15.8s +tttg: c117/185 lr:0.000301 t:15.9s +tttg: c118/185 lr:0.000293 t:16.0s +tttg: c119/185 lr:0.000285 t:16.1s +tttg: c120/185 lr:0.000278 t:16.2s +tttg: c121/185 lr:0.000270 t:16.3s +tttg: c122/185 lr:0.000262 t:16.4s +tttg: c123/185 lr:0.000255 t:16.5s +tttg: c124/185 lr:0.000248 t:16.6s +tttg: c125/185 lr:0.000240 t:16.7s +tttg: c126/185 lr:0.000233 t:16.8s +tttg: c127/185 lr:0.000226 t:16.9s +tttg: c128/185 lr:0.000219 t:17.1s +tttg: c129/185 lr:0.000212 t:17.2s +tttg: c130/185 lr:0.000205 t:17.3s +tttg: c131/185 lr:0.000198 t:17.4s +tttg: c132/185 lr:0.000191 t:17.5s +tttg: c133/185 lr:0.000184 t:18.3s +tttg: c134/185 lr:0.000178 t:18.4s +tttg: c135/185 lr:0.000171 t:18.5s +tttg: c136/185 lr:0.000165 t:18.7s +tttg: c137/185 lr:0.000159 t:18.8s +tttg: c138/185 lr:0.000153 t:18.9s +tttg: c139/185 lr:0.000146 t:19.0s +tttg: c140/185 lr:0.000140 t:19.1s +tttg: c141/185 lr:0.000135 t:19.2s +tttg: c142/185 lr:0.000129 t:19.3s +tttg: c143/185 lr:0.000123 t:19.4s +tttg: c144/185 lr:0.000118 t:19.5s +tttg: c145/185 lr:0.000112 t:19.6s +tttg: c146/185 lr:0.000107 t:19.7s +tttg: c147/185 lr:0.000102 t:19.8s +tttg: c148/185 lr:0.000096 t:19.9s +tttg: c149/185 lr:0.000092 t:20.0s +tttg: c150/185 lr:0.000087 t:20.1s +tttg: c151/185 lr:0.000082 t:20.2s +tttg: c152/185 lr:0.000077 t:20.3s +tttg: c153/185 lr:0.000073 t:20.4s +tttg: c154/185 lr:0.000068 t:20.5s +tttg: c155/185 lr:0.000064 t:20.6s +tttg: c156/185 lr:0.000060 t:20.7s +tttg: c157/185 lr:0.000056 t:20.9s +tttg: c158/185 lr:0.000052 t:21.0s +tttg: c159/185 lr:0.000048 t:21.1s +tttg: c160/185 lr:0.000045 t:21.2s +tttg: c161/185 lr:0.000041 t:21.3s +tttg: c162/185 lr:0.000038 t:21.4s +tttg: c163/185 lr:0.000035 t:21.5s +tttg: c164/185 lr:0.000032 t:21.6s +tttg: c165/185 lr:0.000029 t:21.7s +tttg: c166/185 lr:0.000026 t:21.8s +tttg: c167/185 lr:0.000023 t:21.9s +tttg: c168/185 lr:0.000021 t:22.0s +tttg: c169/185 lr:0.000019 t:22.1s +tttg: c170/185 lr:0.000016 t:22.2s +tttg: c171/185 lr:0.000014 t:22.3s +tttg: c172/185 lr:0.000012 t:22.4s +tttg: c173/185 lr:0.000010 t:22.5s +tttg: c174/185 lr:0.000009 t:22.6s +tttg: c175/185 lr:0.000007 t:22.7s +tttg: c176/185 lr:0.000006 t:22.8s +tttg: c177/185 lr:0.000005 t:22.9s +tttg: c178/185 lr:0.000004 t:23.0s +tttg: c179/185 lr:0.000003 t:23.1s +tttg: c180/185 lr:0.000002 t:23.3s +tttg: c181/185 lr:0.000001 t:23.4s +tttg: c182/185 lr:0.000001 t:23.5s +tttg: c183/185 lr:0.000000 t:23.6s +tttg: c184/185 lr:0.000000 t:23.7s +ttpr: phase:2/3 t:367.2s +ttp: b746/782 bl:2.4164 bb:1.0648 rl:2.3107 rb:1.0653 dl:2884-2943 gd:0 +ttp: b745/782 bl:2.2362 bb:1.0238 rl:2.3027 rb:1.0609 dl:2842-2883 gd:0 +ttpp: phase:3/3 pd:2448 gd:2000 t:384.3s +tttg: c1/250 lr:0.001000 t:0.1s +tttg: c2/250 lr:0.001000 t:0.1s +tttg: c3/250 lr:0.001000 t:0.2s +tttg: c4/250 lr:0.001000 t:0.3s +tttg: c5/250 lr:0.000999 t:0.4s +tttg: c6/250 lr:0.000999 t:0.5s +tttg: c7/250 lr:0.000999 t:0.8s +tttg: c8/250 lr:0.000998 t:0.9s +tttg: c9/250 lr:0.000997 t:1.1s +tttg: c10/250 lr:0.000997 t:1.2s +tttg: c11/250 lr:0.000996 t:1.3s +tttg: c12/250 lr:0.000995 t:1.4s +tttg: c13/250 lr:0.000994 t:1.5s +tttg: c14/250 lr:0.000993 t:1.6s +tttg: c15/250 lr:0.000992 t:1.7s +tttg: c16/250 lr:0.000991 t:1.8s +tttg: c17/250 lr:0.000990 t:1.9s +tttg: c18/250 lr:0.000989 t:2.0s +tttg: c19/250 lr:0.000987 t:2.1s +tttg: c20/250 lr:0.000986 t:2.2s +tttg: c21/250 lr:0.000984 t:2.3s +tttg: c22/250 lr:0.000983 t:2.4s +tttg: c23/250 lr:0.000981 t:2.5s +tttg: c24/250 lr:0.000979 t:2.6s +tttg: c25/250 lr:0.000977 t:2.7s +tttg: c26/250 lr:0.000975 t:2.8s +tttg: c27/250 lr:0.000973 t:2.9s +tttg: c28/250 lr:0.000971 t:3.1s +tttg: c29/250 lr:0.000969 t:3.2s +tttg: c30/250 lr:0.000967 t:3.3s +tttg: c31/250 lr:0.000965 t:3.4s +tttg: c32/250 lr:0.000962 t:3.5s +tttg: c33/250 lr:0.000960 t:3.6s +tttg: c34/250 lr:0.000957 t:3.7s +tttg: c35/250 lr:0.000955 t:3.8s +tttg: c36/250 lr:0.000952 t:3.9s +tttg: c37/250 lr:0.000949 t:4.0s +tttg: c38/250 lr:0.000947 t:4.1s +tttg: c39/250 lr:0.000944 t:4.2s +tttg: c40/250 lr:0.000941 t:4.3s +tttg: c41/250 lr:0.000938 t:4.4s +tttg: c42/250 lr:0.000935 t:4.6s +tttg: c43/250 lr:0.000931 t:4.7s +tttg: c44/250 lr:0.000928 t:4.8s +tttg: c45/250 lr:0.000925 t:4.9s +tttg: c46/250 lr:0.000922 t:5.0s +tttg: c47/250 lr:0.000918 t:5.1s +tttg: c48/250 lr:0.000915 t:5.2s +tttg: c49/250 lr:0.000911 t:5.3s +tttg: c50/250 lr:0.000907 t:5.4s +tttg: c51/250 lr:0.000904 t:5.5s +tttg: c52/250 lr:0.000900 t:5.6s +tttg: c53/250 lr:0.000896 t:5.7s +tttg: c54/250 lr:0.000892 t:5.8s +tttg: c55/250 lr:0.000888 t:5.9s +tttg: c56/250 lr:0.000884 t:6.0s +tttg: c57/250 lr:0.000880 t:6.1s +tttg: c58/250 lr:0.000876 t:6.2s +tttg: c59/250 lr:0.000872 t:6.3s +tttg: c60/250 lr:0.000868 t:8.1s +tttg: c61/250 lr:0.000863 t:8.2s +tttg: c62/250 lr:0.000859 t:8.3s +tttg: c63/250 lr:0.000855 t:8.4s +tttg: c64/250 lr:0.000850 t:8.5s +tttg: c65/250 lr:0.000846 t:8.7s +tttg: c66/250 lr:0.000841 t:8.8s +tttg: c67/250 lr:0.000836 t:8.9s +tttg: c68/250 lr:0.000832 t:9.0s +tttg: c69/250 lr:0.000827 t:9.1s +tttg: c70/250 lr:0.000822 t:9.2s +tttg: c71/250 lr:0.000817 t:9.3s +tttg: c72/250 lr:0.000812 t:9.4s +tttg: c73/250 lr:0.000807 t:9.5s +tttg: c74/250 lr:0.000803 t:9.6s +tttg: c75/250 lr:0.000797 t:9.7s +tttg: c76/250 lr:0.000792 t:9.8s +tttg: c77/250 lr:0.000787 t:9.9s +tttg: c78/250 lr:0.000782 t:10.0s +tttg: c79/250 lr:0.000777 t:10.1s +tttg: c80/250 lr:0.000772 t:10.2s +tttg: c81/250 lr:0.000766 t:10.4s +tttg: c82/250 lr:0.000761 t:11.3s +tttg: c83/250 lr:0.000755 t:11.4s +tttg: c84/250 lr:0.000750 t:11.5s +tttg: c85/250 lr:0.000745 t:11.6s +tttg: c86/250 lr:0.000739 t:11.8s +tttg: c87/250 lr:0.000733 t:11.9s +tttg: c88/250 lr:0.000728 t:12.0s +tttg: c89/250 lr:0.000722 t:12.1s +tttg: c90/250 lr:0.000717 t:12.2s +tttg: c91/250 lr:0.000711 t:12.3s +tttg: c92/250 lr:0.000705 t:12.4s +tttg: c93/250 lr:0.000699 t:12.5s +tttg: c94/250 lr:0.000694 t:12.6s +tttg: c95/250 lr:0.000688 t:12.7s +tttg: c96/250 lr:0.000682 t:12.8s +tttg: c97/250 lr:0.000676 t:12.9s +tttg: c98/250 lr:0.000670 t:13.0s +tttg: c99/250 lr:0.000664 t:13.1s +tttg: c100/250 lr:0.000658 t:13.2s +tttg: c101/250 lr:0.000652 t:13.3s +tttg: c102/250 lr:0.000646 t:13.4s +tttg: c103/250 lr:0.000640 t:13.5s +tttg: c104/250 lr:0.000634 t:13.6s +tttg: c105/250 lr:0.000628 t:13.8s +tttg: c106/250 lr:0.000622 t:13.9s +tttg: c107/250 lr:0.000616 t:14.0s +tttg: c108/250 lr:0.000610 t:14.1s +tttg: c109/250 lr:0.000603 t:14.2s +tttg: c110/250 lr:0.000597 t:14.9s +tttg: c111/250 lr:0.000591 t:15.0s +tttg: c112/250 lr:0.000585 t:15.1s +tttg: c113/250 lr:0.000579 t:15.2s +tttg: c114/250 lr:0.000572 t:15.3s +tttg: c115/250 lr:0.000566 t:15.5s +tttg: c116/250 lr:0.000560 t:15.6s +tttg: c117/250 lr:0.000554 t:15.7s +tttg: c118/250 lr:0.000547 t:15.8s +tttg: c119/250 lr:0.000541 t:15.9s +tttg: c120/250 lr:0.000535 t:16.0s +tttg: c121/250 lr:0.000528 t:16.1s +tttg: c122/250 lr:0.000522 t:16.2s +tttg: c123/250 lr:0.000516 t:16.3s +tttg: c124/250 lr:0.000509 t:16.4s +tttg: c125/250 lr:0.000503 t:16.5s +tttg: c126/250 lr:0.000497 t:16.6s +tttg: c127/250 lr:0.000491 t:16.7s +tttg: c128/250 lr:0.000484 t:16.8s +tttg: c129/250 lr:0.000478 t:16.9s +tttg: c130/250 lr:0.000472 t:17.0s +tttg: c131/250 lr:0.000465 t:17.1s +tttg: c132/250 lr:0.000459 t:17.3s +tttg: c133/250 lr:0.000453 t:17.4s +tttg: c134/250 lr:0.000446 t:17.5s +tttg: c135/250 lr:0.000440 t:17.6s +tttg: c136/250 lr:0.000434 t:17.7s +tttg: c137/250 lr:0.000428 t:17.8s +tttg: c138/250 lr:0.000421 t:17.9s +tttg: c139/250 lr:0.000415 t:18.0s +tttg: c140/250 lr:0.000409 t:18.1s +tttg: c141/250 lr:0.000403 t:18.2s +tttg: c142/250 lr:0.000397 t:18.3s +tttg: c143/250 lr:0.000390 t:18.4s +tttg: c144/250 lr:0.000384 t:18.5s +tttg: c145/250 lr:0.000378 t:18.6s +tttg: c146/250 lr:0.000372 t:18.7s +tttg: c147/250 lr:0.000366 t:18.8s +tttg: c148/250 lr:0.000360 t:18.9s +tttg: c149/250 lr:0.000354 t:19.0s +tttg: c150/250 lr:0.000348 t:19.1s +tttg: c151/250 lr:0.000342 t:19.2s +tttg: c152/250 lr:0.000336 t:19.3s +tttg: c153/250 lr:0.000330 t:19.5s +tttg: c154/250 lr:0.000324 t:19.6s +tttg: c155/250 lr:0.000318 t:19.7s +tttg: c156/250 lr:0.000312 t:19.8s +tttg: c157/250 lr:0.000306 t:19.9s +tttg: c158/250 lr:0.000301 t:20.0s +tttg: c159/250 lr:0.000295 t:20.1s +tttg: c160/250 lr:0.000289 t:20.2s +tttg: c161/250 lr:0.000283 t:20.3s +tttg: c162/250 lr:0.000278 t:20.4s +tttg: c163/250 lr:0.000272 t:20.5s +tttg: c164/250 lr:0.000267 t:20.6s +tttg: c165/250 lr:0.000261 t:20.7s +tttg: c166/250 lr:0.000255 t:20.8s +tttg: c167/250 lr:0.000250 t:20.9s +tttg: c168/250 lr:0.000245 t:21.0s +tttg: c169/250 lr:0.000239 t:21.2s +tttg: c170/250 lr:0.000234 t:21.3s +tttg: c171/250 lr:0.000228 t:21.4s +tttg: c172/250 lr:0.000223 t:21.5s +tttg: c173/250 lr:0.000218 t:21.6s +tttg: c174/250 lr:0.000213 t:21.7s +tttg: c175/250 lr:0.000208 t:21.8s +tttg: c176/250 lr:0.000203 t:21.9s +tttg: c177/250 lr:0.000197 t:22.0s +tttg: c178/250 lr:0.000193 t:22.1s +tttg: c179/250 lr:0.000188 t:22.2s +tttg: c180/250 lr:0.000183 t:22.3s +tttg: c181/250 lr:0.000178 t:22.4s +tttg: c182/250 lr:0.000173 t:22.5s +tttg: c183/250 lr:0.000168 t:22.6s +tttg: c184/250 lr:0.000164 t:22.7s +tttg: c185/250 lr:0.000159 t:22.8s +tttg: c186/250 lr:0.000154 t:23.0s +tttg: c187/250 lr:0.000150 t:23.1s +tttg: c188/250 lr:0.000145 t:23.2s +tttg: c189/250 lr:0.000141 t:23.3s +tttg: c190/250 lr:0.000137 t:23.4s +tttg: c191/250 lr:0.000132 t:23.5s +tttg: c192/250 lr:0.000128 t:23.6s +tttg: c193/250 lr:0.000124 t:23.7s +tttg: c194/250 lr:0.000120 t:23.8s +tttg: c195/250 lr:0.000116 t:23.9s +tttg: c196/250 lr:0.000112 t:24.0s +tttg: c197/250 lr:0.000108 t:24.1s +tttg: c198/250 lr:0.000104 t:24.2s +tttg: c199/250 lr:0.000100 t:24.3s +tttg: c200/250 lr:0.000096 t:24.4s +tttg: c201/250 lr:0.000093 t:24.5s +tttg: c202/250 lr:0.000089 t:24.6s +tttg: c203/250 lr:0.000085 t:24.7s +tttg: c204/250 lr:0.000082 t:24.8s +tttg: c205/250 lr:0.000078 t:24.9s +tttg: c206/250 lr:0.000075 t:25.0s +tttg: c207/250 lr:0.000072 t:25.1s +tttg: c208/250 lr:0.000069 t:25.2s +tttg: c209/250 lr:0.000065 t:25.3s +tttg: c210/250 lr:0.000062 t:25.4s +tttg: c211/250 lr:0.000059 t:25.5s +tttg: c212/250 lr:0.000056 t:25.6s +tttg: c213/250 lr:0.000053 t:25.7s +tttg: c214/250 lr:0.000051 t:25.8s +tttg: c215/250 lr:0.000048 t:25.9s +tttg: c216/250 lr:0.000045 t:26.0s +tttg: c217/250 lr:0.000043 t:26.1s +tttg: c218/250 lr:0.000040 t:26.3s +tttg: c219/250 lr:0.000038 t:26.4s +tttg: c220/250 lr:0.000035 t:26.5s +tttg: c221/250 lr:0.000033 t:26.6s +tttg: c222/250 lr:0.000031 t:26.7s +tttg: c223/250 lr:0.000029 t:26.8s +tttg: c224/250 lr:0.000027 t:26.9s +tttg: c225/250 lr:0.000025 t:27.0s +tttg: c226/250 lr:0.000023 t:27.1s +tttg: c227/250 lr:0.000021 t:27.2s +tttg: c228/250 lr:0.000019 t:27.3s +tttg: c229/250 lr:0.000017 t:27.4s +tttg: c230/250 lr:0.000016 t:27.5s +tttg: c231/250 lr:0.000014 t:27.6s +tttg: c232/250 lr:0.000013 t:27.7s +tttg: c233/250 lr:0.000011 t:27.8s +tttg: c234/250 lr:0.000010 t:27.9s +tttg: c235/250 lr:0.000009 t:28.0s +tttg: c236/250 lr:0.000008 t:28.1s +tttg: c237/250 lr:0.000007 t:28.2s +tttg: c238/250 lr:0.000006 t:28.3s +tttg: c239/250 lr:0.000005 t:28.4s +tttg: c240/250 lr:0.000004 t:28.5s +tttg: c241/250 lr:0.000003 t:28.6s +tttg: c242/250 lr:0.000003 t:28.7s +tttg: c243/250 lr:0.000002 t:28.9s +tttg: c244/250 lr:0.000001 t:29.0s +tttg: c245/250 lr:0.000001 t:29.1s +tttg: c246/250 lr:0.000001 t:29.2s +tttg: c247/250 lr:0.000000 t:29.3s +tttg: c248/250 lr:0.000000 t:29.4s +tttg: c249/250 lr:0.000000 t:29.5s +ttpr: phase:3/3 t:416.3s +ttp: b736/782 bl:2.2480 bb:1.0591 rl:2.2979 rb:1.0607 dl:2526-2550 gd:1 +ttp: b735/782 bl:2.3872 bb:1.0983 rl:2.3050 rb:1.0637 dl:2495-2526 gd:1 +ttp: b721/782 bl:2.3122 bb:1.0268 rl:2.3054 rb:1.0613 dl:2144-2163 gd:1 +ttp: b717/782 bl:2.2557 bb:1.0328 rl:2.3026 rb:1.0596 dl:2070-2088 gd:1 +ttp: b708/782 bl:2.3067 bb:1.0318 rl:2.3028 rb:1.0581 dl:1924-1937 gd:1 +ttp: b698/782 bl:2.2529 bb:1.0312 rl:2.3005 rb:1.0569 dl:1803-1814 gd:1 +ttp: b691/782 bl:2.4523 bb:1.0676 rl:2.3068 rb:1.0574 dl:1725-1737 gd:1 +ttp: b682/782 bl:2.3446 bb:1.0580 rl:2.3083 rb:1.0574 dl:1638-1646 gd:1 +ttp: b675/782 bl:2.3667 bb:1.0584 rl:2.3104 rb:1.0574 dl:1578-1586 gd:1 +ttp: b665/782 bl:2.3308 bb:1.0471 rl:2.3110 rb:1.0571 dl:1500-1507 gd:1 +ttp: b658/782 bl:2.2574 bb:1.0219 rl:2.3094 rb:1.0560 dl:1452-1459 gd:1 +ttp: b650/782 bl:2.3121 bb:1.0500 rl:2.3095 rb:1.0558 dl:1398-1406 gd:1 +ttp: b642/782 bl:2.3215 bb:1.0394 rl:2.3098 rb:1.0554 dl:1349-1356 gd:1 +ttp: b633/782 bl:2.2780 bb:1.0235 rl:2.3090 rb:1.0546 dl:1297-1302 gd:1 +ttp: b624/782 bl:2.3592 bb:1.0679 rl:2.3102 rb:1.0549 dl:1249-1255 gd:1 +ttp: b616/782 bl:2.4002 bb:1.0411 rl:2.3122 rb:1.0546 dl:1205-1211 gd:1 +ttp: b608/782 bl:2.3460 bb:1.0779 rl:2.3129 rb:1.0551 dl:1168-1172 gd:1 +ttp: b600/782 bl:2.2620 bb:1.0135 rl:2.3119 rb:1.0542 dl:1133-1137 gd:1 +ttp: b595/782 bl:2.3530 bb:1.0621 rl:2.3127 rb:1.0544 dl:1110-1115 gd:1 +ttp: b586/782 bl:2.2570 bb:1.0320 rl:2.3116 rb:1.0540 dl:1073-1076 gd:1 +ttp: b577/782 bl:2.2922 bb:1.0317 rl:2.3113 rb:1.0536 dl:1037-1041 gd:1 +ttp: b569/782 bl:2.3087 bb:1.0439 rl:2.3113 rb:1.0534 dl:1007-1010 gd:1 +ttp: b561/782 bl:2.2442 bb:1.0123 rl:2.3102 rb:1.0527 dl:979-983 gd:1 +ttp: b553/782 bl:2.2848 bb:1.0301 rl:2.3098 rb:1.0524 dl:952-955 gd:1 +ttp: b545/782 bl:2.3329 bb:1.0316 rl:2.3101 rb:1.0521 dl:927-930 gd:1 +ttp: b537/782 bl:2.3809 bb:1.0739 rl:2.3111 rb:1.0524 dl:902-905 gd:1 +ttp: b529/782 bl:2.3119 bb:1.0156 rl:2.3111 rb:1.0519 dl:878-882 gd:1 +ttp: b521/782 bl:2.3568 bb:1.0682 rl:2.3117 rb:1.0521 dl:854-858 gd:1 +ttp: b513/782 bl:2.3634 bb:1.0375 rl:2.3124 rb:1.0519 dl:832-835 gd:1 +ttp: b500/782 bl:2.3225 bb:1.0629 rl:2.3125 rb:1.0520 dl:796-799 gd:1 +ttp: b490/782 bl:2.3885 bb:1.0548 rl:2.3134 rb:1.0521 dl:771-773 gd:1 +ttp: b482/782 bl:2.3332 bb:1.0489 rl:2.3136 rb:1.0520 dl:752-754 gd:1 +ttp: b474/782 bl:2.3407 bb:1.0717 rl:2.3139 rb:1.0522 dl:733-735 gd:1 +ttp: b466/782 bl:2.3903 bb:1.0305 rl:2.3146 rb:1.0520 dl:714-717 gd:1 +ttp: b458/782 bl:2.2071 bb:1.0236 rl:2.3136 rb:1.0517 dl:697-700 gd:1 +ttp: b450/782 bl:2.3669 bb:1.0375 rl:2.3141 rb:1.0516 dl:680-682 gd:1 +ttp: b442/782 bl:2.2549 bb:1.0290 rl:2.3135 rb:1.0514 dl:664-666 gd:1 +ttp: b434/782 bl:2.3689 bb:1.0517 rl:2.3140 rb:1.0514 dl:647-648 gd:1 +ttp: b426/782 bl:2.2544 bb:1.0432 rl:2.3135 rb:1.0513 dl:632-634 gd:1 +ttp: b418/782 bl:2.2852 bb:1.0376 rl:2.3133 rb:1.0512 dl:617-618 gd:1 +ttp: b410/782 bl:2.3228 bb:1.0199 rl:2.3134 rb:1.0509 dl:601-603 gd:1 +ttp: b402/782 bl:2.2406 bb:0.9971 rl:2.3128 rb:1.0505 dl:586-588 gd:1 +ttp: b394/782 bl:2.2488 bb:0.9901 rl:2.3123 rb:1.0501 dl:571-573 gd:1 +ttp: b386/782 bl:2.3407 bb:1.0992 rl:2.3125 rb:1.0504 dl:557-559 gd:1 +ttp: b377/782 bl:2.2269 bb:1.0202 rl:2.3119 rb:1.0502 dl:542-544 gd:1 +ttp: b368/782 bl:2.3656 bb:1.1017 rl:2.3123 rb:1.0505 dl:527-528 gd:1 +ttp: b360/782 bl:2.3111 bb:1.0812 rl:2.3123 rb:1.0507 dl:513-515 gd:1 +ttp: b352/782 bl:2.4255 bb:1.0976 rl:2.3130 rb:1.0510 dl:499-501 gd:1 +ttp: b344/782 bl:2.3746 bb:1.0583 rl:2.3134 rb:1.0511 dl:488-489 gd:1 +ttp: b313/782 bl:2.2866 bb:1.0774 rl:2.3132 rb:1.0512 dl:440-442 gd:1 +ttp: b305/782 bl:2.3406 bb:1.0879 rl:2.3134 rb:1.0514 dl:429-430 gd:1 +ttp: b296/782 bl:2.3843 bb:1.0978 rl:2.3137 rb:1.0516 dl:415-417 gd:1 +ttp: b286/782 bl:2.3777 bb:1.1091 rl:2.3141 rb:1.0519 dl:400-402 gd:1 +ttp: b279/782 bl:2.3160 bb:1.0943 rl:2.3141 rb:1.0521 dl:391-392 gd:1 +ttp: b272/782 bl:2.3667 bb:1.0932 rl:2.3143 rb:1.0523 dl:382-383 gd:1 +ttp: b265/782 bl:2.3699 bb:1.1027 rl:2.3146 rb:1.0525 dl:372-374 gd:1 +ttp: b243/782 bl:2.3602 bb:1.0830 rl:2.3147 rb:1.0526 dl:345-346 gd:1 +ttp: b234/782 bl:2.4219 bb:1.1476 rl:2.3152 rb:1.0530 dl:334-335 gd:1 +ttp: b226/782 bl:2.3706 bb:1.0995 rl:2.3154 rb:1.0532 dl:324-325 gd:1 +ttp: b215/782 bl:2.3956 bb:1.0981 rl:2.3157 rb:1.0533 dl:312-313 gd:1 +ttp: b205/782 bl:2.3235 bb:1.1125 rl:2.3157 rb:1.0536 dl:301-302 gd:1 +ttp: b197/782 bl:2.3620 bb:1.1165 rl:2.3159 rb:1.0538 dl:292-294 gd:1 +ttp: b188/782 bl:2.3497 bb:1.1034 rl:2.3160 rb:1.0539 dl:282-283 gd:1 +ttp: b179/782 bl:2.3625 bb:1.1263 rl:2.3161 rb:1.0541 dl:273-274 gd:1 +ttp: b171/782 bl:2.4757 bb:1.1416 rl:2.3166 rb:1.0544 dl:266-266 gd:1 +ttp: b160/782 bl:2.3845 bb:1.1135 rl:2.3168 rb:1.0546 dl:255-255 gd:1 +ttp: b150/782 bl:2.3287 bb:1.1057 rl:2.3169 rb:1.0547 dl:245-246 gd:1 +ttp: b142/782 bl:2.3897 bb:1.1123 rl:2.3171 rb:1.0549 dl:237-238 gd:1 +ttp: b134/782 bl:2.4242 bb:1.1368 rl:2.3174 rb:1.0551 dl:230-231 gd:1 +ttp: b126/782 bl:2.3995 bb:1.1431 rl:2.3176 rb:1.0553 dl:222-223 gd:1 +ttp: b118/782 bl:2.4664 bb:1.1258 rl:2.3179 rb:1.0555 dl:215-216 gd:1 +ttp: b109/782 bl:2.4918 bb:1.1875 rl:2.3184 rb:1.0558 dl:207-208 gd:1 +ttp: b103/782 bl:2.4594 bb:1.1838 rl:2.3187 rb:1.0561 dl:202-202 gd:1 +ttp: b93/782 bl:2.4711 bb:1.1852 rl:2.3190 rb:1.0563 dl:192-193 gd:1 +ttp: b85/782 bl:2.4965 bb:1.1956 rl:2.3194 rb:1.0566 dl:185-186 gd:1 +ttp: b77/782 bl:2.5144 bb:1.2350 rl:2.3198 rb:1.0569 dl:178-179 gd:1 +ttp: b68/782 bl:2.5108 bb:1.1718 rl:2.3202 rb:1.0572 dl:170-171 gd:1 +ttp: b57/782 bl:2.4643 bb:1.1604 rl:2.3204 rb:1.0573 dl:160-161 gd:1 +ttp: b49/782 bl:2.4542 bb:1.1670 rl:2.3207 rb:1.0575 dl:152-153 gd:1 +ttp: b39/782 bl:2.4482 bb:1.1851 rl:2.3209 rb:1.0577 dl:142-143 gd:1 +ttp: b30/782 bl:2.5960 bb:1.2657 rl:2.3213 rb:1.0580 dl:133-134 gd:1 +ttp: b20/782 bl:2.5784 bb:1.2347 rl:2.3216 rb:1.0582 dl:122-123 gd:1 +ttp: b10/782 bl:2.6238 bb:1.1755 rl:2.3220 rb:1.0584 dl:107-109 gd:1 +ttp: b1/782 bl:2.8421 bb:1.1831 rl:2.3224 rb:1.0585 dl:27-83 gd:1 +quantized_ttt_phased val_loss:2.32247803 val_bpb:1.06128183 eval_time:519546ms +total_eval_time:519.5s