Record: MuonEq-R + Depth Recurrence + N61 Mixed GPTQ — val_bpb 1.0924 (3-seed mean)#1279
Open
dexhunter wants to merge 1 commit intoopenai:mainfrom
Open
Record: MuonEq-R + Depth Recurrence + N61 Mixed GPTQ — val_bpb 1.0924 (3-seed mean)#1279dexhunter wants to merge 1 commit intoopenai:mainfrom
dexhunter wants to merge 1 commit intoopenai:mainfrom
Conversation
… (3-seed mean) Improves PR openai#1260 (1.0929) by using N_INT6=61 (one more int6 layer) with a smaller mini runner (21,396 bytes) that creates enough headroom. 3-seed mean: 1.0924 BPB / 2.5133 nats (seeds 42, 0, 7) All seeds under 16MB (max: 15,996,591 bytes) No TTT, no SLOT, no eval-time adaptation. Techniques: MuonEq-R optimizer, depth recurrence (layers 4,5 shared MLP), 61 int6 + 5 int5 Hessian-ranked GPTQ, brotli-11 compression. Built on PR openai#1218 by @clarkkev.
Open
4 tasks
yuyeon
added a commit
to yuyeon/parameter-golf
that referenced
this pull request
Apr 3, 2026
New architecture: instead of N independent transformer blocks, use K shared blocks cycled to N virtual layers, with per-layer FiLM conditioning (learned scale vectors for attn/mlp/residual per virtual layer). Saves massive parameters — 3 shared blocks for 9 virtual layers uses ~6.5M vs 17.1M params, freeing artifact budget. This is genuinely novel for parameter-golf: no submission has tried feature-wise linear modulation for depth conditioning. The closest is PR openai#1279's LoRA adapters, but FiLM is much cheaper (1024 params per virtual layer vs ~8K for LoRA rank-4). Experiments running: Standard 9L vs FiLM 3→9 vs FiLM 3→18 vs FiLM 1→9. Also includes best_full_run.log: Kitchen Sink seq2048 at 600s reached 1.2698 BPB (1338 steps, 15.6MB artifact). 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.
Summary
Key Innovation: N_INT6=61
PR #1260 used N_INT6=60. By regenerating a smaller self-extracting mini runner (21,396 bytes vs 87K standalone), we freed enough artifact budget to fit one additional int6 layer. N_INT6=61 improves BPP by ~0.001 per seed with zero architecture change — purely a quantization precision upgrade.
Results (8xH100 80GB SXM, PyTorch 2.9.1+cu128)
Changes from PR #1218
Credits
Test plan