Skip to content

record: val_bpb=1.1622, NorMuon + int6 STE + SWA + sliding window#89

Open
vmfunc wants to merge 1 commit intoopenai:mainfrom
vmfunc:submission/normuon-int6ste-swa-slidingwindow
Open

record: val_bpb=1.1622, NorMuon + int6 STE + SWA + sliding window#89
vmfunc wants to merge 1 commit intoopenai:mainfrom
vmfunc:submission/normuon-int6ste-swa-slidingwindow

Conversation

@vmfunc
Copy link
Copy Markdown

@vmfunc vmfunc commented Mar 19, 2026

mean val_bpb=1.1622 across 3 seeds on 8xH100 (1.1624, 1.1623, 1.1618). stacks six orthogonal improvements:

  • int6 STE, fake per-row int6 quantization during training w/ straight-through estimator. model learns to handle post-training quant. gap is only +0.002 bpb
  • fp16 embedding passthrough tied embed/logit head kept in fp16 instead of quantized, most quant-sensitive tensor, no STE protection
  • MLP 3x (1536 hidden) int6 compression frees enough artifact bytes to fit the wider model
  • NorMuon row-normalized Newton-Schulz updates (from modded-nanogpt) second-moment normalization on top of Muon
  • SWA over 7 checkpoints during warmdown
  • sliding window eval stride=64 (every scored token gets 960 tokens of context) ~0.033 bpb improvement
run seed steps post-quant bpb sliding window bpb
1 1337 11917 1.1956 1.1624
2 42 11925 1.1955 1.1623
3 2025 11917 1.1951 1.1618

artifact: 15.5MB (code 54KB + int6+zstd model 15.4MB). ~50ms/step, 600s wall clock

mean val_bpb=1.1622 across 3 seeds (1.1624, 1.1623, 1.1618).
int6 fake quant w/ STE, fp16 embed passthrough, MLP 3x, NorMuon,
stochastic weight averaging during warmdown, sliding window stride=64.
15.5MB artifact, 8xH100, 600s, ~12k steps.
@vmfunc vmfunc force-pushed the submission/normuon-int6ste-swa-slidingwindow branch from f6a92be to c887ef4 Compare March 19, 2026 15:18
NotADevIAmaMeatPopsicle added a commit to NotADevIAmaMeatPopsicle/parameter-golf that referenced this pull request Mar 19, 2026
NorMuon adds per-row second-moment tracking after Newton-Schulz
orthogonalization, then normalizes and rescales to preserve total
norm. Based on arXiv:2510.05491 and PR openai#89. Expected -0.005 to
-0.010 BPB improvement. Drop-in replacement (same class name).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants