feat: depth recurrence + cosine recovery TTT#697
Open
Danishlynx wants to merge 2 commits intoopenai:mainfrom
Open
feat: depth recurrence + cosine recovery TTT#697Danishlynx wants to merge 2 commits intoopenai:mainfrom
Danishlynx wants to merge 2 commits intoopenai:mainfrom
Conversation
Based on merged SOTA (PR openai#549 stack + LeakyReLU² + Legal TTT, 1.1194 bpb): 1. Depth recurrence: repeat layers 4-5 → 13 virtual layers from 11 physical - Per-repetition learnable scale parameters - U-Net skip connections adapted for virtual layer count - DEPTH_RECURRENCE=4,5 env var 2. Enhanced TTT with cosine recovery phase: - After standard score-first TTT, runs N additional cosine-LR epochs on all scored data to repair int6 quantization damage - Re-scores with standard sliding window eval - TTT_RECOVERY_EPOCHS=20, TTT_RECOVERY_LR=0.001 env vars 3. FlashAttention 3 fallback to SDPA for non-Hopper GPUs - Manual GQA head repeat for PyTorch <2.5 compatibility Smoke-tested on 1xH100 SXM 80GB. Both features validated. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- fullgraph=True → fullgraph=False for torch.compile (conditional branches in _run_layers break fullgraph) - Create fresh uncompiled model for TTT eval to avoid stale inference tensor state from compiled eval model - Clear Rotary cos/sin caches when transitioning between inference_mode (scoring) and train mode (adaptation) to prevent "Inference tensors cannot be saved for backward" errors - Manual GQA head repeat for PyTorch <2.5 SDPA compatibility Validated: TTT now runs end-to-end on 1xH100, achieving 1.3859 bpb (from 1.5158 post-quant baseline, -0.13 bpb improvement) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
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.
Based on merged SOTA (PR #549 stack + LeakyReLU² + Legal TTT, 1.1194 bpb):
Depth recurrence: repeat layers 4-5 → 13 virtual layers from 11 physical
Enhanced TTT with cosine recovery phase:
FlashAttention 3 fallback to SDPA for non-Hopper GPUs
Smoke-tested on 1xH100 SXM 80GB. Both features validated.