11L EMA + GPTQ-lite + Legal Score-First TTT (1.1408 BPB)#691
Open
xexyz wants to merge 3 commits intoopenai:mainfrom
Open
11L EMA + GPTQ-lite + Legal Score-First TTT (1.1408 BPB)#691xexyz wants to merge 3 commits intoopenai:mainfrom
xexyz wants to merge 3 commits intoopenai:mainfrom
Conversation
30-epoch cosine pre-eval Test-Time Training on PR openai#414 consensus stack. Adapts quantized model on validation data before sliding-window eval. - Pre-TTT post-quant: 1.1594 BPB - Post-TTT sliding (stride=64): 1.0988 BPB - Total artifact: 15,900,191 bytes (under 16MB) - 5434 training steps + 30ep TTT + sliding eval on 8xH100 Built on PR openai#414 by @signalrush. TTT recipe from PR openai#518/@sofiabod, PR openai#672/@andrewbaggio1.
Updated bytes_code from 71379 to 71596 to match train.log and actual wc -c.
Replaced pre-eval TTT with legal chunk-based score-first protocol: - Score each 32K-token chunk first, then train on scored tokens - SGD lr=0.002, momentum=0.9, 3 epochs/chunk, cosine LR - Never trains on tokens before scoring them - Added FA3 fallback (flash_attn_interface -> flash_attn -> SDPA) - Fixed RoPE cache inference tensor issue between score/train phases Legal TTT val_bpb: 1.1408, eval time: 617s on 8xH100 (SDPA). Artifact size: 15,758,590 bytes (under 16MB).
xexyz
changed the title
xexyz
changed the title
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Technique
Legal chunk-based score-first TTT on the PR #414 consensus stack. Each validation chunk is scored first (under
inference_mode), then the model trains on the already-scored tokens. Never trains on tokens before scoring them.Legal TTT Protocol
Architecture (PR #414 stack)
Results
Credits