diff --git a/README.md b/README.md index 39012623ea..c23dfb61c9 100644 --- a/README.md +++ b/README.md @@ -232,4 +232,4 @@ The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching Join the [OpenAI Discord server](https://discord.com/invite/openai) and visit the Parameter Golf channels (#parameter-golf-discussions, #parameter-golf-announcements) and ask questions. -This repository adapts code from `modded-nanogpt`, see [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) for attribution. +This repository adapts code from `modded-nanogpt`, see [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) for attribution. \ No newline at end of file diff --git a/records/track_10min_16mb/2026-04-03_RandomLinearMaps_LoRA_Adapters/README.md b/records/track_10min_16mb/2026-04-03_RandomLinearMaps_LoRA_Adapters/README.md new file mode 100644 index 0000000000..406f83af7e --- /dev/null +++ b/records/track_10min_16mb/2026-04-03_RandomLinearMaps_LoRA_Adapters/README.md @@ -0,0 +1,84 @@ +# Random Linear Maps + Learned LoRA Adapters + +## Summary + +This submission implements the **"Learning adapters on random linear maps"** idea from the challenge wishlist — a previously unclaimed approach that inverts the standard train→compress paradigm. + +**Core idea**: Instead of training all weights and then compressing to fit in 16MB, we: + +1. **Freeze most weights as pseudo-random projections** initialized from a deterministic seed (stored in code = 0 bytes in artifact). +2. **Only train small LoRA-style low-rank adapters** (rank 16) on each layer, plus embeddings, norms, and control parameters. +3. **At save time**, serialize only the trained adapter weights + seed. +4. **At load time**, regenerate the full random backbone from the seed and apply the trained adapters. + +## Why This Is Interesting + +- **Massive model for free**: The frozen random backbone (12 layers, 768 dim, 3x MLP) has ~70M+ parameters but costs **0 bytes** in the artifact since it's reproducible from a seed. +- **Only ~5-10M trainable params**: These fit easily in 16MB even at FP16, leaving headroom for wider/deeper architectures. +- **Theoretically motivated**: Random features are surprisingly powerful (Random Kitchen Sinks, Lottery Ticket Hypothesis). The frozen random projections provide a rich feature basis that the adapters learn to combine. +- **Novel for this challenge**: Nobody has tried this approach — it's fundamentally different from the quantization-focused submissions on the leaderboard. + +## Architecture + +| Component | Value | +| -------------------- | -------------------- | +| Layers | 12 | +| Model dim | 768 | +| Heads | 12 (4 KV heads, GQA) | +| MLP mult | 3x | +| LoRA rank | 16 | +| Vocab | 1024 (sp1024) | +| Backbone seed | 42 | +| Trainable params | ~5-10M | +| Frozen random params | ~70M+ | + +## Key Components + +### LoRALinear Module + +Each linear layer has: + +- A **frozen random base weight** `W` (from deterministic seed, stored as buffer) +- **Trainable low-rank adapters** `A` (rank×in) and `B` (out×rank) +- Output: `W@x + (B@A)@x * scale` +- `B` initialized to zero so initial behavior = pure random projection + +### Optimizer Split + +- **Muon**: LoRA adapter matrices (lora_A, lora_B) +- **Adam**: Token embeddings, scalar/control parameters, norms + +### Serialization + +- Only trainable parameters are saved (not frozen buffers) +- Int8 + zlib compression on the trainable subset +- At load time: regenerate random backbone from seed, apply dequantized adapters + +## Running + +```bash +cd /workspace/parameter-golf + +python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 1 + +RUN_ID=random_lora_v1 \ +DATA_PATH=./data/datasets/fineweb10B_sp1024/ \ +TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ +VOCAB_SIZE=1024 \ +torchrun --standalone --nproc_per_node=1 records/track_10min_16mb/2026-04-03_RandomLinearMaps_LoRA_Adapters/train_gpt.py +``` + +## Potential Improvements + +- **Selective unfreezing**: Unfreeze first/last layer base weights for better embedding-to-hidden and hidden-to-logit projections. +- **Larger LoRA rank** on critical layers (attention Q/K vs MLP). +- **Different random initialization** schemes (orthogonal, spectral norm matching). +- **Hybrid**: Freeze only MLP base weights (largest), train attention fully. +- **Combine with proven techniques**: BigramHash, sliding eval, EMA/SWA. + +## Theoretical Background + +- [Random Features for Large-Scale Kernel Machines](https://papers.nips.cc/paper/2007/hash/013a006f03dbc5392effeb8f18fda755-Abstract.html) (Rahimi & Recht, 2007) +- [The Lottery Ticket Hypothesis](https://arxiv.org/abs/1803.03635) (Frankle & Carlin, 2018) +- [LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/abs/2106.09685) (Hu et al., 2021) +- [Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuning](https://arxiv.org/abs/2012.13255) (Aghajanyan et al., 2020) diff --git a/records/track_10min_16mb/2026-04-03_RandomLinearMaps_LoRA_Adapters/submission.json b/records/track_10min_16mb/2026-04-03_RandomLinearMaps_LoRA_Adapters/submission.json new file mode 100644 index 0000000000..d3745d721a --- /dev/null +++ b/records/track_10min_16mb/2026-04-03_RandomLinearMaps_LoRA_Adapters/submission.json @@ -0,0 +1,9 @@ +{ + "author": "austinluk", + "github_id": "austinluk", + "date": "2026-04-03", + "val_bpb": null, + "description": "Random Linear Maps + Learned LoRA Adapters: freeze most weights as pseudo-random (from seed = 0 bytes in artifact), train only low-rank adapters. Enables a much larger backbone (12L, 768d, 3x MLP) under 16MB since frozen weights cost nothing in the artifact.", + "track": "10min_16mb", + "tags": ["random-linear-maps", "lora", "adapters", "creative"] +} diff --git a/records/track_10min_16mb/2026-04-03_RandomLinearMaps_LoRA_Adapters/train_gpt.py b/records/track_10min_16mb/2026-04-03_RandomLinearMaps_LoRA_Adapters/train_gpt.py new file mode 100644 index 0000000000..a3714eba7d --- /dev/null +++ b/records/track_10min_16mb/2026-04-03_RandomLinearMaps_LoRA_Adapters/train_gpt.py @@ -0,0 +1,955 @@ +""" +Random Linear Maps + Learned LoRA Adapters + +Creative submission for Parameter Golf: instead of training all weights and compressing, +we freeze most weights as pseudo-random (reproducible from a seed = 0 bytes in artifact) +and only train small LoRA-style low-rank adapters on top. + +This inverts the standard paradigm: instead of training → quantizing → compressing, +we start "compressed" (random weights are free) and only learn what we must. + +Key ideas: +- Base transformer weights are frozen random projections, seeded deterministically. +- Only LoRA adapters (rank-16), biases, norms, embeddings, and control params are trained. +- The trainable parameter count is ~5-10M, fitting easily in 16MB even at FP16. +- The frozen backbone can be much wider/deeper than normal since it costs 0 bytes. +- At save time, we only serialize the trained adapters + seed. +- At load time, we regenerate the random backbone from the seed and apply adapters. + +Architecture: 12 layers, 768 dim, 12 heads, 4 KV heads, 3x MLP, LoRA rank 16. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +RANDOM_SEED_BACKBONE = 42 # Deterministic seed for frozen random weights (stored in code, 0 bytes in artifact) +LORA_RANK = 16 # Low-rank adapter dimension + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Bigger model since frozen weights are free in artifact + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 12)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 768)) + num_heads = int(os.environ.get("NUM_HEADS", 12)) + mlp_mult = int(os.environ.get("MLP_MULT", 3)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + lora_rank = int(os.environ.get("LORA_RANK", str(LORA_RANK))) + + # Optimizer hyperparameters + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + adapter_lr = float(os.environ.get("ADAPTER_LR", 0.01)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + 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-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): + super().__init__(params, dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov)) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + 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: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION +# ----------------------------- + +def build_sentencepiece_luts(sp, vocab_size, device): + sp_vocab_size = int(sp.vocab_size()) + 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): + files = [Path(p) for p in sorted(glob.glob(pattern))] + 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 eval_val(args, model, rank, world_size, device, grad_accum_steps, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut): + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError("VAL_BATCH_SIZE too small") + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + 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) + model.eval() + with torch.inference_mode(): + 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 * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~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) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + + +# ----------------------------- +# QUANTIZATION (for adapter weights only) +# ----------------------------- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + p for p 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", + ).split(",") if p +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = CONTROL_TENSOR_NAME_PATTERNS +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_Q = 99.99984 / 100.0 + + +def tensor_nbytes(t): + return int(t.numel()) * int(t.element_size()) + + +def keep_float_tensor(name, t, passthrough_orig_dtypes): + if any(p in name for p in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + + +def quantize_float_tensor(t): + t32 = t.float() + if t32.ndim == 2: + clip_abs = torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) if t32.numel() else torch.empty((t32.shape[0],), dtype=torch.float32) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + + +def quantize_state_dict_int8(state_dict): + quantized, scales, dtypes, passthrough = {}, {}, {}, {} + passthrough_orig_dtypes = {} + qmeta = {} + stats = dict.fromkeys(("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), 0) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj = {"__quant_format__": "int8_clean_per_row_v1", "quantized": quantized, "scales": scales, "dtypes": dtypes, "passthrough": passthrough} + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + + +def dequantize_state_dict_int8(obj): + out = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + out[name] = (q.float() * float(s.item())).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file): + header_bytes = 256 * np.dtype(" 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + def __init__(self, pattern, rank, world_size, device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens, seq_len, grad_accum_steps): + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + + +# ----------------------------- +# LoRA ADAPTER MODULE +# ----------------------------- + +class LoRALinear(nn.Module): + """Frozen random base weight + trainable low-rank adapter. + + The base weight W is initialized from a deterministic seed and frozen. + Only the low-rank matrices A (down) and B (up) are trained. + Output = W @ x + (B @ A) @ x * scale + """ + def __init__(self, in_features: int, out_features: int, rank: int, seed: int, bias: bool = False, scale: float = 1.0): + super().__init__() + self.in_features = in_features + self.out_features = out_features + self.rank = rank + self.scale = scale + + # Frozen random base weight — initialized from seed, never trained + rng = torch.Generator() + rng.manual_seed(seed) + # Kaiming uniform initialization for the frozen base + std = 1.0 / math.sqrt(in_features) + base_weight = torch.empty(out_features, in_features).uniform_(-std, std, generator=rng) + self.register_buffer("base_weight", base_weight) # Not a parameter — won't be saved/trained + + # Trainable LoRA adapters + self.lora_A = nn.Parameter(torch.zeros(rank, in_features)) # Down projection + self.lora_B = nn.Parameter(torch.zeros(out_features, rank)) # Up projection + nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5)) + # B starts at zero so initial output = base_weight @ x (no adapter effect) + + self.bias_param = nn.Parameter(torch.zeros(out_features)) if bias else None + + def forward(self, x: Tensor) -> Tensor: + # Base (frozen) contribution + base_out = F.linear(x, self.base_weight.to(x.dtype)) + # LoRA (trained) contribution + lora_out = F.linear(F.linear(x, self.lora_A.to(x.dtype)), self.lora_B.to(x.dtype)) * self.scale + out = base_out + lora_out + if self.bias_param is not None: + out = out + self.bias_param.to(x.dtype) + return out + + +class FrozenLinear(nn.Module): + """Fully frozen random linear — no adapters, just cheap capacity.""" + def __init__(self, in_features: int, out_features: int, seed: int): + super().__init__() + rng = torch.Generator() + rng.manual_seed(seed) + std = 1.0 / math.sqrt(in_features) + weight = torch.empty(out_features, in_features).uniform_(-std, std, generator=rng) + self.register_buffer("weight", weight) + + def forward(self, x: Tensor) -> Tensor: + return F.linear(x, self.weight.to(x.dtype)) + + +# ----------------------------- +# TRANSFORMER MODULES (with LoRA) +# ----------------------------- + +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) + + +def restore_low_dim_params_to_fp32(module): + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + def __init__(self, dim, base=10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + 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._seq_len_cached != seq_len or self._cos_cached.device != device: + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x, cos, sin): + 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, lora_rank, layer_seed): + super().__init__() + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + kv_dim = self.num_kv_heads * self.head_dim + + # Q, K, V, Proj all use LoRA on frozen random bases + self.c_q = LoRALinear(dim, dim, rank=lora_rank, seed=layer_seed * 10 + 1) + self.c_k = LoRALinear(dim, kv_dim, rank=lora_rank, seed=layer_seed * 10 + 2) + self.c_v = LoRALinear(dim, kv_dim, rank=lora_rank, seed=layer_seed * 10 + 3) + self.proj = LoRALinear(dim, dim, rank=lora_rank, seed=layer_seed * 10 + 4) + + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x): + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + 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) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, k, v, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + def __init__(self, dim, mlp_mult, lora_rank, layer_seed): + super().__init__() + hidden = mlp_mult * dim + self.fc = LoRALinear(dim, hidden, rank=lora_rank, seed=layer_seed * 10 + 5) + self.proj = LoRALinear(hidden, dim, rank=lora_rank, seed=layer_seed * 10 + 6) + + def forward(self, x): + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__(self, dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init, lora_rank, layer_seed): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, lora_rank, layer_seed) + self.mlp = MLP(dim, mlp_mult, lora_rank, layer_seed) + 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()) + + def forward(self, x, x0): + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) + return x + + +class GPT(nn.Module): + def __init__(self, vocab_size, num_layers, model_dim, num_heads, num_kv_heads, + mlp_mult, tie_embeddings, tied_embed_init_std, logit_softcap, rope_base, + qk_gain_init, lora_rank): + super().__init__() + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) # Trainable + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init, + lora_rank, layer_seed=RANDOM_SEED_BACKBONE * 1000 + i) + for i in range(num_layers) + ]) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else LoRALinear(model_dim, vocab_size, rank=lora_rank, seed=RANDOM_SEED_BACKBONE * 1000 + 999) + 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) + + def trainable_state_dict(self): + """Return only trainable parameters (adapters + embeddings + norms + scalars). + Frozen random base weights (buffers) are excluded.""" + return {name: param for name, param in self.named_parameters()} + + def forward(self, input_ids, target_ids): + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main(): + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # Distributed + CUDA setup + 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")) + 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") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + 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) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg, console=True): + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0(subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, console=False) + log0("=" * 100, console=False) + + # Tokenizer + validation setup + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError(f"VOCAB_SIZE mismatch: {args.vocab_size} vs {int(sp.vocab_size())}") + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts(sp, args.vocab_size, device) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # Model setup — with LoRA adapters on frozen random backbone + base_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + lora_rank=args.lora_rank, + ).to(device).bfloat16() + + restore_low_dim_params_to_fp32(base_model) + + # Count trainable vs total params + total_params = sum(p.numel() for p in base_model.parameters()) + total_buffers = sum(b.numel() for b in base_model.buffers()) + trainable_params = sum(p.numel() for p in base_model.parameters() if p.requires_grad) + frozen_buffer_params = total_buffers + + log0(f"model_architecture: Random Linear Maps + LoRA Adapters") + log0(f"total_params (trainable): {trainable_params}") + log0(f"frozen_buffer_params (random, from seed): {frozen_buffer_params}") + log0(f"effective_model_size: {trainable_params + frozen_buffer_params}") + log0(f"lora_rank: {args.lora_rank}") + log0(f"backbone_seed: {RANDOM_SEED_BACKBONE}") + log0(f"artifact_contains: ONLY trainable params ({trainable_params}) + code + seed") + + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer — only train adapters, embeddings, norms, scalars + # Separate LoRA adapter matrices for Muon, everything else for Adam + adapter_matrix_params = [] + scalar_params = [] + seen_param_ids = set() + + def add_unique(target_list, param): + pid = id(param) + if pid in seen_param_ids: + return + seen_param_ids.add(pid) + target_list.append(param) + + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + if "lora_A" in name or "lora_B" in name: + add_unique(adapter_matrix_params, p) + elif p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS): + add_unique(scalar_params, p) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), eps=args.adam_eps, fused=True, + ) + optimizer_muon = Muon( + adapter_matrix_params, lr=args.adapter_lr, + momentum=args.muon_momentum, backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.adapter_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), eps=args.adam_eps, fused=True, + ) + optimizers = [optimizer_tok, optimizer_muon, optimizer_scalar] + + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0(f"adapter_lr:{args.adapter_lr} embed_lr:{token_lr} scalar_lr:{args.scalar_lr}") + log0(f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} iterations:{args.iterations}") + log0(f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}") + log0(f"seed:{args.seed}") + + # Data loader & warmup + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all(): + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step, elapsed_ms): + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + if args.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() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + 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) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # Main training loop + training_time_ms = 0.0 + stop_after_step = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val(args, model, rank, world_size, device, grad_accum_steps, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut) + log0(f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms") + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0(f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms step:{step}/{args.iterations}") + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = args.train_log_every > 0 and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + if should_log_train: + log0(f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms") + + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if 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 + + log0(f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB") + + # Serialization — ONLY save trainable parameters (adapters + embeddings + norms) + # Frozen random weights are regenerated from seed at load time + if master_process: + trainable_sd = {name: param.detach().cpu() for name, param in base_model.named_parameters()} + torch.save(trainable_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model (trainable only): {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + # Int8 quantize the trainable params only + trainable_sd = {name: param.detach().cpu() for name, param in base_model.named_parameters()} + quant_obj, quant_stats = quantize_state_dict_int8(trainable_sd) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0(f"Serialized model int8+zlib: {quant_file_bytes} bytes (payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)") + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + # Roundtrip validation — reload quantized adapters onto fresh random backbone + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + dequant_sd = dequantize_state_dict_int8(quant_state) + # Load only trainable params back + current_sd = base_model.state_dict() + for name in dequant_sd: + current_sd[name] = dequant_sd[name] + base_model.load_state_dict(current_sd, strict=True) + + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val(args, model, rank, world_size, device, grad_accum_steps, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut) + torch.cuda.synchronize() + log0(f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main()