[Non-Record] QAT Dead-Code Analysis + 7 Novel Technique Sweep (1xH100)

records/track_non_record_16mb/2026-03-28_QAT_DeadCode_Analysis_NovelTechniques_1xH100/README.md

@@ -0,0 +1,150 @@
+# QAT Dead-Code Bug + 7 Untried Techniques: What I Found
+
+**Non-Record Submission (Research Contribution)**
+**Author:** [@wfproc](https://github.com/wfproc)
+**Base:** PR #549 by @abaybektursun (1.1194 bpb, current SOTA)
+**Hardware:** 1xH100 80GB SXM, 600s wallclock (~940 steps per run)
+**Best pre-quant:** 1.3646 val_bpb | **Best post-quant (int6 sliding):** 2.3411 val_bpb
+
+Not competitive with the 8xH100 leaderboard (7x fewer steps), but the findings transfer to any hardware.
+
+---
+
+## The Short Version
+
+I spent two days running 20+ experiments on the SOTA #549 stack, testing techniques nobody else had submitted results for. The main finding: Late QAT is dead code in #315-derived submissions due to a torch.compile bug. Fixing it actually made the int6 gap *worse*, not better. I also tested 7 novel techniques from recent papers (Muon-VS, Deep Delta Learning, stable depth recurrence, anti-layer removal, NS step reduction, spectral SVD compression, wider models). All negative. The throughput tax at this scale kills everything that adds overhead.
+
+---
+
+## The QAT Dead-Code Bug
+
+PR #315's README flagged this already, but I wanted to confirm it and try to fix it.
+
+The issue: `CastedLinear._qat_enabled` is a class attribute set to `False`. When `torch.compile(fullgraph=True)` traces the model, it sees `_qat_enabled=False` and eliminates the STE branch from the compiled graph. Setting it to `True` later (when Late QAT is supposed to activate) does nothing. The compiled code is frozen.
+
+This is present in the SOTA (#549). I verified by checking step times: with the original code, step time stays flat at ~625ms throughout training. If QAT was actually running, you'd see a jump when it activates (STE adds overhead).
+
+I tried three ways to fix it:
+
+| Fix | What happened |
+|---|---|
+| Mutable list `[False]` | `FailOnRecompileLimitHit` error, `fullgraph=True` won't recompile |
+| Instance attribute per module | Same recompilation error |
+| Tensor multiplier (0.0 or 1.0) | Works. Graph shape stays the same, no recompile needed |
+
+The tensor approach: `w = w + qat_scale * (w_q - w).detach()`. When scale is 0.0, it's a no-op. When 1.0, full STE. The compiled graph handles both without recompilation.
+
+With the fix, step time jumps from ~625ms to ~670ms when QAT activates. So QAT is actually running now. But the results got worse:
+
+| Metric | Dead QAT (original) | Working QAT (tensor fix) |
+|---|---|---|
+| val_bpb pre-quant | 1.3631 | 1.3646 |
+| val_bpb int6 sliding | 2.3107 | 2.3411 |
+
+Why? A few theories:
+
+1. At only ~940 steps (1xH100), the model hasn't converged enough for STE to help. The weights are still moving too fast.
+2. WD=0.04 + EMA 0.997 already push weights toward int6-friendly distributions without needing explicit fake quantization.
+3. The tensor multiply `qat_scale * (w_q - w).detach()` always executes (even at scale=0.0), adding a tiny overhead that costs a few training steps.
+
+I can't rule out that QAT would help at 7000 steps (8xH100). But it's also possible that the dead code is genuinely the right call for this stack.
+
+---
+
+## Novel Technique Sweep
+
+All tested on the SOTA #549 stack on 1xH100. Baseline: 1.3631 val_bpb (pre-quant, slope=0.75, 947 steps).
+
+| # | Technique | Paper | val_bpb | Delta | Why it failed |
+|---|---|---|---|---|---|
+| 1 | Muon-VS (variance-adaptive) | 2601.14603 | 1.3884 | +0.019 | Variance buffer needs 100s of steps to warm up, +2.2% overhead |
+| 2 | Deep Delta Learning | 2601.00417 | 1.3720 | +0.003 | Beta params stay near zero at init, within noise |
+| 3 | Thinking Deeper recurrence | 2603.21676 | 1.4163 | +0.053 | 21% step overhead, 163 fewer steps |
+| 4 | Anti-layer removal | 2603.19348 | n/a | n/a | No anti-layers found in 11L stack |
+| 5 | Newton-Schulz steps=3 | - | 1.3798 | +0.017 | Worse orthogonalization quality, +9 steps doesn't compensate |
+| 6 | Spectral SVD (640d 12L) | Novel | 1.4646 | +0.102 | 46% slower, only 649 steps, SVD can't separate signal from noise |
+| 7 | Wider model (576d 11L) | - | 1.4504 | +0.087 | 33% slower, only 711 steps |
+
+Also confirmed LeakyReLU(0.75)^2 > LeakyReLU(0.5)^2 by 0.006 bpb, matching PR #977.
+
+### The throughput tax
+
+This keeps coming up: at ~83ms/step on H100, each millisecond of added overhead costs roughly 0.007 bpb. Techniques 1, 3, 6, and 7 all added overhead. None of them delivered enough quality improvement per step to compensate. The only technique that worked (LeakyReLU slope change) adds zero overhead.
+
+### Anti-layer results (full table)
+
+Trained 943 steps, then zeroed each layer's attn_scale + mlp_scale and re-evaluated:
+
+| Layer | Delta bpb when removed | Role |
+|---|---|---|
+| 0 | +1.127 | Critical (embedding interaction) |
+| 1 | +0.346 | Critical |
+| 2 | +0.237 | Important |
+| 3 | +0.087 | Moderate |
+| 4 | +0.081 | Moderate |
+| 5-6 | +0.036 | Low |
+| 7-8 | +0.031-0.034 | Low (layer 8 least important) |
+| 9-10 | +0.040-0.044 | Low |
+
+No layer hurts performance when present. The LN Scale (1/sqrt(L+1)) in the stack may prevent anti-layer formation by dampening deep layers.
+
+---
+
+## Other Findings
+
+**SWA during QAT:** PR #989 found SWA sabotages QAT. I added a one-line fix that stops SWA accumulation when QAT activates. Worth adopting if you use both.
+
+**Prune-then-quantize:** Implemented but only partially tested. Prunes smallest-magnitude weights before int6 quantization (arXiv:2603.18426). One partial run showed neutral results. The pruned zeros don't compress better than quantized values under lzma, so the artifact size didn't shrink. Needs full evaluation.
+
+---
+
+## Reproducing These Results
+
+From the repo root:
+```bash
+# Download data (if not already present)
+python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 10
+
+# Main run (slope=0.75, includes QAT tensor-scale fix)
+RUN_ID=main \
+LEAKY_SLOPE=0.75 \
+torchrun --standalone --nproc_per_node=1 \
+records/track_non_record_16mb/2026-03-28_QAT_DeadCode_Analysis_NovelTechniques_1xH100/train_gpt.py
+
+# With prune-then-quantize (5%)
+RUN_ID=prune5 \
+LEAKY_SLOPE=0.75 \
+PRUNE_FRACTION=0.05 \
+torchrun --standalone --nproc_per_node=1 \
+records/track_non_record_16mb/2026-03-28_QAT_DeadCode_Analysis_NovelTechniques_1xH100/train_gpt.py
+
+# With anti-layer diagnostic (runs layer ablation after training)
+RUN_ID=antilayer \
+LEAKY_SLOPE=0.75 \
+ANTILAYER_DIAGNOSTIC=1 \
+torchrun --standalone --nproc_per_node=1 \
+records/track_non_record_16mb/2026-03-28_QAT_DeadCode_Analysis_NovelTechniques_1xH100/train_gpt.py
+```
+
+Other toggles: `MUON_VS=1`, `DEEP_DELTA=1`, `RECURRENCE_LAYERS=2 RECURRENCE_STEPS=2`, `SVD_RANK=128 MODEL_DIM=640 NUM_LAYERS=12`.
+
+---
+
+## What I'd try next with more compute
+
+1. QAT fix at 7000 steps (8xH100) to see if the STE actually helps when the model is better converged
+2. Spectral SVD compression at full scale (the concept needs 7000+ steps for weight matrices to develop spectral structure)
+3. Prune-then-quantize sweep (5%, 10%, 15%)
+4. LeakyReLU slope sweep at 0.80, 0.85, 0.90
+
+---
+
+## Included Files
+
+- `train_gpt.py` - modified SOTA #549 script with all techniques as env var toggles
+- `submission.json` - metadata
+- `train.log` - QAT fix run output (slope=0.75 with tensor-scale QAT, 1xH100, 943 steps)
+
+## Acknowledgments
+
+Built on PR #549 (@abaybektursun), PR #315 (@jfprincz, first to flag the QAT torch.compile issue), PR #977 (@michaelwinczuk, LeakyReLU 0.75), PR #989 (@alexanderaperry-arch, SWA-QAT finding). Issue #140 commentary by @notapplica was really helpful for prioritizing what to try.

records/track_non_record_16mb/2026-03-28_QAT_DeadCode_Analysis_NovelTechniques_1xH100/requirements.txt

@@ -0,0 +1,5 @@
+torch>=2.6.0
+numpy
+sentencepiece
+flash-attn
+zstandard

records/track_non_record_16mb/2026-03-28_QAT_DeadCode_Analysis_NovelTechniques_1xH100/submission.json

@@ -0,0 +1,18 @@
+{
+    "author": "wfproc",
+    "github_id": "wfproc",
+    "name": "QAT Dead-Code Analysis + Novel Technique Sweep (1xH100)",
+    "blurb": "Non-record research submission. Found that torch.compile constant-folds Late QAT in #315-derived code (confirmed in SOTA #549). Tested a fix via tensor-scale STE, but working QAT actually made the int6 gap worse. Also swept 7 untried techniques from recent papers (Muon-VS, Deep Delta Learning, Thinking Deeper recurrence, anti-layer removal, NS step reduction, spectral SVD compression, wider models). All negative on 1xH100. Key finding: at 83ms/step, each 1ms of overhead costs about 0.007 bpb.",
+    "date": "2026-03-28T00:00:00Z",
+    "track": "non-record-16mb",
+    "val_loss": 2.3041,
+    "val_bpb": 1.3646,
+    "pre_quant_val_loss": null,
+    "pre_quant_val_bpb": 1.3646,
+    "step_stop": 943,
+    "wallclock_seconds": 600,
+    "bytes_total": 7303141,
+    "bytes_model_int6_lzma": 7200188,
+    "bytes_code": 102953,
+    "gpu": "1xH100-80GB-SXM"
+}

records/track_non_record_16mb/2026-03-28_QAT_DeadCode_Analysis_NovelTechniques_1xH100/train.log

@@ -0,0 +1,16 @@
+val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model
+step:0/20000 val_loss:6.9309 val_bpb:4.1049 train_time:0ms step_avg:0.02ms
+swa:start step:250
+late_qat:enabled step:416 scale:0.1499
+step:500/20000 train_loss:2.4300 train_time:318487ms step_avg:636.97ms
+step:943/20000 val_loss:2.3041 val_bpb:1.3646 train_time:600509ms step_avg:636.81ms
+stopping_early: wallclock_cap train_time:600509ms step:943/20000
+DIAGNOSTIC post_ema val_loss:2.4761 val_bpb:1.4665 eval_time:15592ms
+Serialized model: 106158518 bytes
+Serialized model int6+lzma: 7200188 bytes
+Total submission size int6+lzma: 7303141 bytes
+final_int6_roundtrip val_loss:3.9539 val_bpb:2.3417 eval_time:28383ms
+final_int6_roundtrip_exact val_loss:3.95388609 val_bpb:2.34171405
+final_int6_sliding_window val_loss:3.9528 val_bpb:2.3411 stride:64 eval_time:565334ms
+final_int6_sliding_window_exact val_loss:3.95279298 val_bpb:2.34107285
+final_int8_zlib_roundtrip_exact val_loss:3.95279298 val_bpb:2.34107285