Record: Fused softcap CE + WD=2.0 (warm-start stability fix) — val_bpb 1.06957 (3-seed mean)#1886
Open
renqianluo wants to merge 1 commit intoopenai:mainfrom
Open
Record: Fused softcap CE + WD=2.0 (warm-start stability fix) — val_bpb 1.06957 (3-seed mean)#1886renqianluo wants to merge 1 commit intoopenai:mainfrom
renqianluo wants to merge 1 commit intoopenai:mainfrom
Conversation
…d mean) Stacks Fused CE Triton kernel (from @nprime06 PR openai#1787) onto our PR openai#1768. Discovered that fused CE introduces fp32-accumulation differences that, combined with warm-start LoRA A, cause TTT divergence on seeds 314/1337 at WD=1.0 (both collapse to ~1.121). Novel fix: raise TTT_WEIGHT_DECAY 1.0 -> 2.0. This regularizes the cross-batch A drift enough that all 3 seeds stay stable AND keep the warm-start gain. 3-seed mean 1.06957 BPB (−0.00124 vs PR openai#1768's 1.07081). Each seed individually beats its PR openai#1768 counterpart. Trained on RunPod 8xH100 instead of Zoom cluster. Identical SP8192 tokenizer and FineWeb selection as upstream willdepueoai/parameter-golf.
This was referenced May 1, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Stacks @nprime06's Fused softcap CE Triton kernel (PR #1787) on top of our PR #1768. Discovered a divergence interaction — Fused CE's fp32-accumulation differences combined with warm-start A push seeds 314/1337 into a TTT attractor that collapses to ~1.121 BPB at WD=1.0, while seed 42 trains cleanly.
Novel: WD=2.0 unlocks the combination
Raising
TTT_WEIGHT_DECAYfrom 1.0 → 2.0 regularizes the cross-batch A drift enough that all 3 seeds stay stable while keeping the warm-start gain.Results
Compliance: train ≤599.7s, eval 393–478s, artifact 15.98MB. Issue #1017 conditions 1–4 verified.
Hardware
Trained on RunPod 8xH100 80GB SXM (PyTorch 2.9.1+cu128, FA3, Triton 3.5.1). Identical SP8192 tokenizer and FineWeb selection as upstream
willdepueoai/parameter-golf.Attribution
@nprime06 (PR #1787 — Fused CE), PR #1344 (Polar Express NS), @dexhunter (PR #1736 — GatedAttn), @samacqua (PR #1530 — VarLen+FusedMLP+doc-LoRA), @bigbag (PR #1493), @renqianluo (PR #1767, PR #1768).