Non-record: MDLM Diffusion — val_var_bpb 1.1465 (first diffusion to beat AR baseline)#1106
Merged
valerio-oai merged 1 commit intoopenai:mainfrom May 3, 2026
Merged
Conversation
First discrete diffusion model to beat the AR baseline (1.22) in parameter-golf. MDLM with log-linear noise, adaLN, frozen visible-token logits, discrete ELBO eval. 27 hyperparameter experiments. Validated on 2xH100 (TensorPool), 31 min training. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
He-Wenhao
pushed a commit
to He-Wenhao/parameter-golf
that referenced
this pull request
Apr 10, 2026
PR openai#1106 found eps=0.1 >> 0.001 was the single biggest improvement. With eps=0.1, 10% of tokens remain visible at t=1, giving the model anchors for denoising. Larger terminal KL but much easier task. Also revert lr=1e-3, warmdown=1000 (v8's lr=2e-3 made artifact >16MB). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
He-Wenhao
pushed a commit
to He-Wenhao/parameter-golf
that referenced
this pull request
Apr 10, 2026
- train_mdlm_combined.py: full MDLM training script (PR openai#1053 infra + PR openai#1106 MDLM + our innovations) - sweep.sh/sweep2.sh: 12-experiment hyperparameter sweep (eps, arch, loss, seq_len) - results.tsv: updated with v10-v13 experiments, corrected descriptions Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
valerio-oai
approved these changes
May 3, 2026
Contributor
valerio-oai
left a comment
valerio-oai
left a comment
There was a problem hiding this comment.
Selected for the notable non-record submissions section.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
val_var_bpb: 1.1465 (512 eval steps) | 33M params | 2xH100 80GB HBM3 | Non-record
First discrete diffusion model to beat the AR baseline (1.22 BPB). Beats previous best diffusion (#820, 1.625; #1053, 1.360) by 0.21+ BPB.
Results
Approach
MDLM (Sahoo et al. 2024) with log-linear noise, adaLN timestep conditioning, frozen visible-token logits, antithetic sampling, discrete absorbing-mask ELBO eval. 11L 512d, 6000 steps, AdamW.
Key findings from 27 experiments
Hardware
Developed on GB10 (Project DIGITS). Validated on 2xH100 (TensorPool, 31 min). 8xH100 unavailable (#821). Extrapolated: ~8 min on 8xH100.