Repository navigation
bump SGLang-diffusion to sglang-miles branch - #257
Merged
Merged
Conversation
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rockdu
added
run-ci-e2e
Rockdu
added
run-ci-e2e
Rockdu
force-pushed
the
ci/sglang-miles-h3-to-sglang-miles
branch
2 times, most recently
from
September 23, 2026 15:34
5b2b56a to
650be62
Compare
Rockdu
changed the title
Attribution: sgl-project/sglang#37616 (9cb38a3d57, "filter duplicate precision variants across custom loaders") is the only commit between sglang-miles-h3's merge base (5375babbac) and sglang-miles (5a8da8cc3f) that moves this test. A CI-runner bisect (record-e2e-standards on <commit> + the four sglang-miles-h3 picks) shows the series equal to the old standard up to 9cb38a3d57^ and equal to this standard from 9cb38a3d57 on. Why: SD3.5-medium ships both model.safetensors and model.fp16.safetensors (bit-identical) for its CLIP text encoders. Before #37616, sgl-d's own CLIP loader saw both and raised "Duplicate tensor names detected across safetensors files" (517 names for CLIP-G). The pipeline then silently fell back to transformers' CLIPTextModel, which has no projection head and drops text_projection.weight ("UNEXPECTED"). So the CLIP-G half of pooled_projections was the unprojected pooler output: cos -0.016 vs the projected embedding. CLIP-L's projection is near-identity (cos 1.0). The concatenated 2048-d vector gives cos 0.333 vs diffusers' encode_prompt, with norms 50.43 vs 44.68, exactly as measured. After #37616 the sgl-d CLIPTextModelWithProjection loads, all 149 + 389 params match the HF files, and cos is 0.99996. encoder_hidden_states (penultimate layer, no projection) were identical on both sides. Recorded on the h200/3gpu runner against sgl-project/sglang#40786 (run 35798499525). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ity patch sgl-project/sglang#36680 packs Qwen-Image's add_{q,k,v}_proj into one MergedColumnParallelLinear (to_added_qkv). The parity patch forced every LoRA layer through a 2D PEFT-ordered delta, which crashes on the merged layer's 3D stacked LoRA (size 8 vs 3072). Let sgl-d's native LoRA forwards run; the norm/RoPE/split_seqs parity patches stay. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rockdu
force-pushed
the
ci/sglang-miles-h3-to-sglang-miles
branch
from
September 23, 2026 22:50
650be62 to
9b3ba42
Compare
…used epilogue off sgl-project/sglang#33555 passes freqs_complex to apply_qk_norm_with_optional_rope, which the patched replacement rejected. The new fused QK-norm/RoPE epilogue (sm90, unquantized, on by default) also bypasses the patched norms and RoPE, so disable it when the patch group is applied. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Attribution, split by source: - Step 0 (first rollout and the first train step): only this PR's 7d83ef7, which drops the LoRA forward overrides from the qwen_image parity patch. The old sglang base (5375babbac + sglang-miles-h3 picks) with this branch already gives this standard's step-0 values bit for bit. - Everything after the first weight update: sgl-project/sglang#36680 (71cee04ebe, "Optimize Qwen-Image TP collectives and attention"), the only contributing commit in a CI-runner bisect from 5375babbac to 5a8da8cc3f with this branch held fixed. It packs add_{q,k,v}_proj into one to_added_qkv, so once LoRA is live the added-text Q/K/V run as one packed GEMM plus a stacked LoRA delta instead of three GEMMs. Effect: the train<->rollout residual grows slightly (log_prob_mean_abs_diff ~3.1e-5 -> 3.2-4.7e-5) and the reward series barely moves. Recorded on the h200/5gpu runner against sgl-project/sglang#40786 (run 35803865496). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Attribution: sgl-project/sglang#35796 (0447ade326, "fall back to a component's default attention backend") is the only commit between sglang-miles-h3's merge base (5375babbac) and sglang-miles (5a8da8cc3f) that moves this test. This comes from a CI-runner bisect of <commit> + the four sglang-miles-h3 picks (+ the MiniMax-H3 LoRA-wrap fix where #37903's code exists); 5375babbac reproduces the old standard bit for bit. Why: #35796 pins the H3 video VAE's attention to default_attention_backend=TORCH_SDPA ({FA, TORCH_SDPA} supported), where it used to inherit the global FA backend. The decoded frames move by rounding, so PickScore rewards shift in the 5th decimal (step 0 mean 0.74778 -> 0.74774), and the GRPO advantages and train metrics follow. The DiT path is unchanged and the train<->rollout residual keeps its magnitude (log_prob_mean_abs_diff 4.13e-5 vs 4.08e-5). Recorded on the h200/3gpu runner against sgl-project/sglang#40786 (run 35812567370). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Attribution: sgl-project/sglang#37616 (9cb38a3d57) only, the same cause as the SD3 NFT standard. Probes on the CI runner (<commit> + the four sglang-miles-h3 picks) give: 5375babbac == 9cb38a3d57^ == old standard; 9cb38a3d57 == 5a8da8cc3f == this standard. Why: SD3.5-medium ships both model.safetensors and model.fp16.safetensors (bit-identical) for its CLIP text encoders. Before #37616, sgl-d's own CLIP loader saw both and raised "Duplicate tensor names detected across safetensors files" (517 names for CLIP-G). The pipeline then silently fell back to transformers' CLIPTextModel, which has no projection head and drops text_projection.weight ("UNEXPECTED"). So the CLIP-G half of pooled_projections was the unprojected pooler output: cos -0.016 vs the projected embedding. CLIP-L's projection is near-identity (cos 1.0). The concatenated 2048-d vector gives cos 0.333 vs diffusers' encode_prompt, with norms 50.43 vs 44.68, exactly as measured. After #37616 the sgl-d CLIPTextModelWithProjection loads, all 149 + 389 params match the HF files, and cos is 0.99996. encoder_hidden_states (penultimate layer, no projection) were identical on both sides. The step-0 mean reward drops 0.416 -> 0.375 here, but that is sampling noise from the recipe's 8 prompts, not a regression. In a paired run of 128 test prompts x 4 seeds with this recipe's sampling (512 samples per side), the old base scores 0.4366 and sglang-miles 0.4686: +0.032, 95% CI [+0.0006, +0.064]. Resampling 8 of those prompts gives a new-old spread of [-0.087, +0.155], and a <= -0.040 draw has probability ~0.12. Recorded on the h200/3gpu runner against sgl-project/sglang#40786 (run 35927893055). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rockdu
force-pushed
the
ci/sglang-miles-h3-to-sglang-miles
branch
from
September 24, 2026 23:47
9b3ba42 to
944201c
Compare
The previous curve came from a sglang build whose SD3.5 CLIP text encoders fell back to transformers' CLIPTextModel and dropped text_projection (fixed by sgl-project/sglang#37616). The new curve is the same recipe run on sgl-project/sglang#40786. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rockdu
marked this pull request as ready for review
September 25, 2026 22:47
sgl-project/sglang#40786 landed the sglang-miles-h3 commits on sglang-miles, so CI, the Docker image and the docs now follow sglang-miles. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rockdu
added
run-ci-e2e
Rockdu
changed the title
Rockdu
added a commit
that referenced
this pull request
Sep 25, 2026
One run of scripts/run_diffusion_grpo_pickscore_5gpu_flowgrpo_aligned.py (wandb radixarkai/miles-diffusion-CIs/runs/npcsyfm8) on the code of #257: eval/pickscore_test rises from 0.859 at rollout 29 to 0.900 at rollout 389. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rockdu
added a commit
that referenced
this pull request
Sep 25, 2026
One run of scripts/run_diffusion_grpo_pickscore_5gpu_flowgrpo_aligned.py (wandb radixarkai/miles-diffusion-CIs/runs/npcsyfm8) on the code of #257: eval/pickscore_test rises from 0.859 at rollout 29 to 0.900 at rollout 389. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Merged
2 of 6 tasks
OnePunchMonk
pushed a commit
to OnePunchMonk/miles_diffusion
that referenced
this pull request
Oct 5, 2026
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.
Moves miles_diffusion from
sglang-miles-h3tosglang-miles, now that sgl-project/sglang#40786 (the four H3 picks plus two LoRA-wrap fixes) has landed there.Changes
ci: build the rollout engine fromsglang-miles(DEFAULT_SGLANG_REFin.github/workflows/_run-ci.yml,SGLANG_DIFFUSION_BRANCHindocker/Dockerfile, and the docs).qwen_imagerollout patch: drop the LoRA forward overrides, since they break on sglang's packed added-QKV layer (#36680). Also follow the newapply_qk_norm_with_optional_ropesignature and keep the fused QK-norm/RoPE epilogue off.sglang-mileschanges (details in each commit message):text_projection)Validation
e2e 13/13 green against #40786. 400-rollout Qwen-Image recipe run: https://wandb.ai/radixarkai/miles-diffusion-CIs/runs/npcsyfm8
🤖 Generated with Claude Code