warmdown-quantization val_bpb = 1.2154 #61
Merged
0hq merged 2 commits intoopenai:mainfrom Mar 19, 2026
Merged
Conversation
Novel finding: aggressive LR decay (WARMDOWN_ITERS=20000) reduces int8 quantization penalty from 0.014 to 0.005 BPB. Combined with FP16 tied embeddings and moderate NTK-RoPE extrapolation (eval@1408). Full warmdown sweep across 10 values and detailed analysis in README.
South-33
added a commit
to South-33/parameter-golf
that referenced
this pull request
Mar 19, 2026
- add a PR-audit research log entry covering the clean takeaways from pull requests openai#36 through openai#70 - promote long-context training plus matching long-context eval as a first-class clean branch based on PR openai#61 and PR openai#63 - refine mixed-precision export notes to emphasize using int6/int8 byte savings to fund wider MLP capacity, based on PR openai#65 - update the current snapshot and research thesis so future agents do not over-focus on exporter-only ideas after the broader PR sweep
South-33
added a commit
to South-33/parameter-golf
that referenced
this pull request
Mar 19, 2026
- fix the PR-audit notes to attribute the long-context branch to PR openai#65 rather than PR openai#61 - record PR openai#61 as schedule-side evidence about long warmdown reducing quantization damage - keep the ideas backlog aligned with the actual GitHub PR content before using it for next-step decisions
Closed
50225c5 to
4a65f69
Compare
Collaborator
|
Confused, your title says 1.1574 but your header says "val_bpb = 1.2154 (baseline: 1.2244, improvement: 0.009 BPB / 0.017 nats)" |
0hq
approved these changes
Mar 19, 2026
Collaborator
0hq
left a comment
There was a problem hiding this comment.
Tentatively approving, just to keep leaderboard up to date for others.
Collaborator
|
Before I officially add to leaderboard, mind running again to verify that the 1.1574 result is within noise? 1-2 more runs that show the same result would be great. |
Collaborator
|
Ah sorry I see that this is no longer the latest PR, moving there. |
Contributor
Author
|
Hey Will sorry about the confusion on the documentation here. I saw people updating their PR's and I thought that would be a good idea so I went back to update this then changed my mind and reverted it but didn't update the title and comment since I thought nobody would be looking this far back. Missed the title updated now. |
This was referenced Mar 20, 2026
scottspace
pushed a commit
to scottspace/parameter-golf
that referenced
this pull request
Mar 21, 2026
* Warmdown-quantization co-optimization, val_bpb=1.2154 Novel finding: aggressive LR decay (WARMDOWN_ITERS=20000) reduces int8 quantization penalty from 0.014 to 0.005 BPB. Combined with FP16 tied embeddings and moderate NTK-RoPE extrapolation (eval@1408). Full warmdown sweep across 10 values and detailed analysis in README. * breakthrough: 1.1574 BPB via int6 + MLP 3x + sliding window stride=256 --------- Co-authored-by: Sam Larson <saml212@users.noreply.github.com>
leonardcser
pushed a commit
to leonardcser/parameter-golf
that referenced
this pull request
Mar 21, 2026
* Warmdown-quantization co-optimization, val_bpb=1.2154 Novel finding: aggressive LR decay (WARMDOWN_ITERS=20000) reduces int8 quantization penalty from 0.014 to 0.005 BPB. Combined with FP16 tied embeddings and moderate NTK-RoPE extrapolation (eval@1408). Full warmdown sweep across 10 values and detailed analysis in README. * breakthrough: 1.1574 BPB via int6 + MLP 3x + sliding window stride=256 --------- Co-authored-by: Sam Larson <saml212@users.noreply.github.com>
nedcut
pushed a commit
to nedcut/parameter-golf
that referenced
this pull request
Mar 26, 2026
* Warmdown-quantization co-optimization, val_bpb=1.2154 Novel finding: aggressive LR decay (WARMDOWN_ITERS=20000) reduces int8 quantization penalty from 0.014 to 0.005 BPB. Combined with FP16 tied embeddings and moderate NTK-RoPE extrapolation (eval@1408). Full warmdown sweep across 10 values and detailed analysis in README. * breakthrough: 1.1574 BPB via int6 + MLP 3x + sliding window stride=256 --------- Co-authored-by: Sam Larson <saml212@users.noreply.github.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
2 participants
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.
Score
val_bpb = 1.2154 (baseline: 1.2244, improvement: 0.009 BPB / 0.017 nats)
Key Finding
Post-training int8 quantization is the dominant BPB bottleneck on 8xH100. The quantization penalty alone (0.014 BPB at default settings) is larger than most hyperparameter improvements combined. We reduce it 3x via an always-decaying LR schedule.
Novel Contributions
1. Always-decaying LR schedule (WARMDOWN_ITERS=20000): With ~12,200 actual steps, the LR decays linearly from 61% of peak at step 0 to near-zero at the final step. Post-quant penalty drops from 0.014 to 0.005 BPB. Full curve mapped across 10 warmdown values (2400-30000).
2. FP16 tied embeddings: Keep tok_emb.weight in fp16 during int8 export. Costs ~500KB, offset by MLP_HIDDEN=992.
3. Optimal NTK-RoPE extrapolation: eval@1408 (1.375x training length) beats eval@2048 on well-trained 8xH100 models. Full curve from 1024 to 2048.
4. Optimizer-warmdown interaction: MUON_BACKEND_STEPS=5 beats 7 at high warmdown (reversal from low warmdown). When warmdown already smooths weights, more steps > better orthogonalization.
Configuration
15.91MB artifact. 8xH100 SXM (RunPod). See README.md for full warmdown sweep data.