[Diffusion] Optimize Qwen-Image TP collectives and attention by BBuf · Pull Request #36680 · sgl-project/sglang

BBuf · 2026-08-27T10:07:59Z

Summary

This PR consolidates the Qwen-Image TP kernel work previously split across
#36680, #36693, and #36694:

route CUDA diffusion TP collectives through the SRT custom-all-reduce
dispatcher and admit 24 MiB Qwen-Image row-parallel outputs to
CustomAllReduceV2
keep the three unquantized Qwen added-text Q/K/V weights packed, run
three reference GEMMs for default quality=lossless, and mount the single
column-parallel GEMM only for request-scoped quality=high; Nunchaku keeps
its existing quantized packed path
pack masked text-prefix and image Q/K/V directly with a segmented Triton
kernel instead of materializing three concatenated tensors
use the FA3 static persistent scheduler for batch-one dense packed sequences
and scatter back to the padded/BCG bucket when needed
preserve the existing fallbacks for unsupported shapes, devices,
quantization backends, and non-CUDA platforms
document the validated two-H200 recipe and update the diffusion performance
skill/kernel documentation

Request quality semantics

Only the unquantized BF16 added-text QKV GEMM packing is non-bit-exact. It is
now a request-scoped quality site:

--quality=lossless (default): applies the packed Q/K/V weight slices as the
original three independent projection GEMMs, preserving the reference
reduction order and output;
--quality=high: mounts the single packed GEMM used for the performance
numbers below;
switching back to lossless unmounts the site at the next batch boundary.

The checkpoint remains packed in memory in both modes, so this does not add a
second copy of the Q/K/V weights. The segmented pack, attention scheduler, and
collective changes keep their existing independently validated fallbacks.

Actual blast radius

This is broader than the original Qwen-only benchmark, but it does not
speed up every diffusion model or every Qwen checkpoint automatically.

Changed path	Models/configurations that can execute it
CUDA diffusion TP custom-all-reduce dispatch + 32 MiB V2 workspace	Any native CUDA diffusion run with TP > 1; checked-in affected presets include FLUX.1/2, Qwen-Image-2512, Qwen-Image-Edit-2511, Z-Image-Turbo, Ideogram 4 FP8, Cosmos3-Super T2V/I2V, MiniMax-H3, and the 4-GPU Cosmos TP variants
Generic masked-varlen FA3 dense/static fast path	Native FA attention with a 2-D mask plus varlen metadata: ERNIE-Image, Krea-2, Ideogram 4, LingBot Video MoE, Z-Image, and Qwen-Image families
Quality-gated fused added-text QKV + segmented text/image packing	The fused added-QKV GEMM is used only by `quality=high` on unquantized Qwen-Image architecture checkpoints: Qwen-Image, 2512, Edit, Edit-2509, Edit-2511, Layered, and FireRed Image Edit 1.0/1.1. Default `lossless` uses three projection GEMMs; segmented packing remains independently guarded.

LTX-2/2.3 and single-GPU FLUX.2 Klein do not execute these changed branches in
their tested presets; their measurements below are controls, not PR speedup
claims. Qwen/Ideogram NVFP4 are Blackwell-only and the quantized projection path
does not inherit the unquantized fused-QKV change.

Motivation and profiler evidence

The original Qwen-Image-2512 TP2 path left three independent gaps relative to
vLLM-Omni:

Diffusion directly instantiated the legacy custom all-reduce
implementation. The 24 MiB row-parallel outputs therefore fell back to NCCL
instead of using the CUDA V2 dispatcher.
Each attention block materialized joint text/image Q, K, and V tensors
before packing. The trace contained 1,212 CatArray launches taking
10.768 ms in the sampled window.
SGLang selected FA3 VarlenDynamic at 159.403 us/call, while vLLM selected
StaticPersistent at 147.402 us/call for this batch-one dense sequence.

The combined path changes those hot spots as follows:

24 MiB collective microbenchmark: NCCL 104.196 us -> V2 graph one-shot pull
81.754 us (21.5% faster)
materialized QKV pack -> segmented pack: 47.7293 us -> 21.0982 us at the
production batch-one shape
FA3 attention: VarlenDynamic 159.403 us -> StaticPersistent 147.063 us,
matching the vLLM kernel timing
under quality=high, fused added-text QKV removes 1,212 small split-K
GEMMs/reductions and saves about 0.43 ms per denoise step

Cross-model H200 audit

Method

Base: a7e3f590cabb676a02181fb3328026587822f934
PR: 8e056cfbfe68ddc46c1f869b7853ec9fc7e94cfe
Same physical NVIDIA H200 pair (GPUs 6/7)
ABBA order: base-a, PR-a, PR-b, base-b; values below are two-run medians
Native SGLang backend, fixed prompt/seed, checked-in model preset. These
measurements were collected before the request gate was added, when the packed
GEMM ran unconditionally; they therefore describe the current quality=high
path. The new default quality=lossless path retains the reference three-GEMM
added-QKV computation. Eager with compile/BCG off unless explicitly shown
Every checkpoint used a task-owned cache. It was deleted and verified empty
immediately after that model, including failed 403 attempts.

Full JSON summaries, output metrics, all contact sheets, and both cleanup
ledgers are in the
evidence branch.

Successfully executed affected checkpoints

Positive percentages mean lower PR latency. Bold rows cross the audit's 1.5%
signal threshold.

Preset / checkpoint	Topology	Base denoise (s)	PR denoise (s)	Change	Output result
Qwen-Image-2512	TP2 + BCG	8.5406	7.8657	+8.58%	repeats stable; Base/PR SSIM 0.9713
Qwen-Image	1 GPU + BCG	12.6528	12.3580	+2.39%	bit-exact
Qwen-Image-Edit	1 GPU eager	32.0573	32.2576	-0.62%	bit-exact
Qwen-Image-Edit-2509	1 GPU eager	22.3352	22.3464	-0.05%	bit-exact
Qwen-Image-Edit-2511	TP2 eager	14.2704	14.2651	+0.04%	bit-exact
Qwen-Image-Layered	1 GPU eager	32.9077	32.8971	+0.03%	repeats stable; min SSIM 0.9891 over 4 layers
FireRed-Image-Edit-1.0	CFG2 eager	11.4178	11.4694	-0.45%	bit-exact
FireRed-Image-Edit-1.1	CFG2 eager	11.4151	11.4170	-0.02%	bit-exact
Z-Image-Turbo	TP2 eager	0.5575	0.5307	+5.05%	bit-exact
Z-Image	1 GPU eager	9.2629	9.3055	-0.46%	bit-exact
Cosmos3-Super T2V	TP2 eager	113.2269	113.0971	+0.11%	MP4 and sampled frames bit-exact
Cosmos3-Super I2V	TP2 eager	112.9026	112.9645	-0.05%	MP4 and sampled frames bit-exact
ERNIE-Image	1 GPU eager	13.0868	12.8640	+1.73%	prompt enhancement nondeterministic on both sides
ERNIE-Image-Turbo	1 GPU eager	12.9039	13.0717	-1.28%	prompt enhancement nondeterministic on both sides
LingBot Video MoE 30B	1 GPU eager	3.9435	3.9126	+0.79%	MP4 and sampled frames bit-exact

Qwen conclusion: no, the entire Qwen family does not get a measurable
speedup. The two T2I BCG lanes are clear wins; Edit, Edit-2509, Edit-2511,
Layered, and FireRed 1.0/1.1 are neutral or slightly slower at these fixed
shapes. The Qwen-specific code is shared, but its saved launches are too small
relative to the full denoiser on several variants, and only TP runs benefit
from the custom-all-reduce change.

ERNIE's internal prompt-enhancement generation is nondeterministic across fresh
processes on both base and PR. Its pictures are included for disclosure, but
their cross-side SSIM is not evidence of a PR correctness change. Denoise timing
is measured after the independently generated prompt.

Affected but blocked on this assignment

Preset	Status
FLUX.1-dev TP2	Hugging Face 403 gated repo; isolated metadata cache cleaned
FLUX.2-dev TP2	Hugging Face 403 gated repo; isolated metadata cache cleaned
Ideogram 4 FP8 TP2	Hugging Face 403 gated repo; isolated metadata cache cleaned
Ideogram 4 Fast / Instant	public wrappers resolve to gated Ideogram weights (403); caches cleaned
Krea-2 registry root	Registered placeholder returns Hugging Face 404; metadata cache cleaned
Krea-2 Turbo / Raw	Hugging Face 403 gated repos; caches cleaned
MiniMax-H3 TP2+Ulysses2	checked-in preset requires 4 GPUs; assignment had 2
Cosmos3-Super T2V CFG2+TP2	preset requires 4 GPUs; assignment had 2
Cosmos3-Super distilled T2I TP4	preset requires 4 GPUs; assignment had 2
Qwen-Image NVFP4 / Ideogram 4 NVFP4	Blackwell-only; cannot execute on H200

Negative controls (not attributed to this PR)

Preset	Base (s)	PR (s)	Observed change	Why it is a control
FLUX.2 Klein 4B	0.2390	0.2332	+2.52%	1 GPU, no masked-varlen metadata
FLUX.2 Klein Base 4B	6.0333	5.9889	+0.74%	1 GPU, no masked-varlen metadata
LTX-2	9.0477	9.4805	-4.57%	CFG2, not TP; no varlen metadata
LTX-2.3 two-stage	13.5873	12.7691	+6.41%	CFG2, not TP; no varlen metadata

These controls show why an isolated ABBA shift is not generalized unless the
changed branch is proven to execute.

Generated output: Base vs PR

Every successfully executed affected checkpoint has a Base/PR sheet below.
Video sheets show the first and middle frames; Layered shows all four layers.

Qwen family (8 checkpoints)

Qwen-Image-2512

Qwen-Image

Qwen-Image-Edit

Qwen-Image-Edit-2509

Qwen-Image-Edit-2511

Qwen-Image-Layered

FireRed-Image-Edit-1.0

FireRed-Image-Edit-1.1

Z-Image, Cosmos3, ERNIE, and LingBot (7 checkpoints)

Z-Image-Turbo

Z-Image

Cosmos3-Super T2V

Cosmos3-Super I2V

ERNIE-Image (nondeterministic prompt enhancement on both sides)

ERNIE-Image-Turbo (nondeterministic prompt enhancement on both sides)

LingBot Video MoE 30B

The control sheets are retained in the evidence branch but intentionally not
presented as affected-model results.

Separate vLLM-Omni comparison lane

This is the original fixed no-CFG benchmark and is separate from the cross-model
ABBA sweep above. It used Qwen-Image-2512, 1024x1024, 50 steps, guidance scale
1, seed 42, TP2, BCG enabled, compile off, five measured requests, and the same
adjacent full-NVLink H200 pair.

Runtime	Median client latency	Median denoise step
SGLang, combined candidate	4.101 s	76.9 ms
SGLang, CARv2-only baseline	4.256 s	78.6 ms
fastest observed vLLM-Omni regional-compile path	4.442 s	-
vLLM-Omni forced eager	7.794 s	-

The combined SGLang candidate is 7.7% faster than the fastest observed
vLLM-Omni result under that fixed workload. The fused added-QKV part of this
candidate now requires quality=high; default lossless keeps the reference
projection order.

Correctness and cleanup

Deterministic checkpoints were repeat-stable. Eleven affected lanes were
bit-exact Base vs PR. Qwen-Image-2512 (SSIM 0.9713 / PSNR 30.15 dB) and
Qwen-Image-Layered (minimum SSIM 0.9891 across four layers) were not bit-exact,
but both were stable within each side.
The quality=high BF16 Qwen added-text projection fusion changes reduction
association. Default quality=lossless now runs the reference three GEMMs. In
the original no-CFG high-path lane, the fixed-seed image measured PSNR 40.55 dB and
pixel MAE 1.393/255 against the unfused baseline.
ERNIE base and Turbo were nondeterministic within both base and PR because
prompt enhancement is regenerated in each fresh process; the output sheets
are disclosure rather than a correctness attribution.
The two cleanup ledgers contain 29 records (successful models, 403/404 attempts,
and LTX-2.3's extra materialized checkpoint). Every record reports zero
remaining bytes/weight files. The task cache itself is now empty except for
its marker file.

Test plan

pre-commit run --from-ref origin/main --to-ref HEAD
Qwen-Image fused-QKV and BCG capture tests on H200: 12 passed, 3 subtests passed
request-scoped added-QKV quality gate on latest PR head: lossless/high/unmount switching, 14 import/Qwen tests passed on GB300
segmented pack/layout tests on H200: 16 passed
segmented pack benchmark on H200: 47.7293 us -> 21.0982 us (batch 1), 17.8346 us -> 7.7642 us (batch 2)
Qwen-Image-2512 TP2 BCG model-backed benchmark, five measured requests
same-pair vLLM-Omni comparison
affected-model H200 ABBA sweep, generated-media comparison, and per-model cleanup

The BF16 reduction-order tradeoff is now isolated behind quality=high and
documented explicitly. Supersedes #36693 and #36694.

CI States

Latest PR Test (Base): ⏳ Run #33459254730
Latest PR Test (Extra): ⏳ Run #33459254596
Latest PR Test (AMD ROCm 7.2): ⏳ Run #33459254715

Preserve the PR's Qwen-Image NVFP4 fallback mapping while adopting main's transformer override quantization detection.

BBuf · 2026-08-31T06:10:45Z

Updated the H100 consistency-data pin in 1ed919b73d to 883cf11, the merge commit of ci-data-diffusion#6.

This revision keeps the Qwen-Image fused-QKV GT from ci-data-diffusion#5, while restoring the three Sana H100 frames to the main-compatible blobs used before 88b17d4.

The Sana mismatch was unrelated to this PR's Qwen-Image runtime changes. 88b17d4 had captured output from the unmerged auto-residency experiment in #35335 / #36703, where post-warmup component placement changes were not numerically invariant for the two-stage Sana pipeline. That experiment later set SanaWMPipelineConfig.supports_auto_residency = False and moved its own ci-data pin back to the pre-88b17d4 revision, but the experimental images remained in the shared ci-data history and were picked up by this PR's repository-wide revision bump.

I also merged the latest main in b385ffe72a. The only content conflict was in transformer_load_utils.py; the resolution keeps this PR's quant_param_names_mapping for Qwen-Image NVFP4 fallback discovery and preserves main's new replacement-checkpoint quantization detection.

Validation performed locally:

the three Sana blob IDs in 883cf11 exactly match the pre-88b17d4 blobs;
the Qwen-Image fused-QKV GT remains unchanged;
git diff --check, Python compilation, and pre-commit on the touched/conflict-resolved files passed.

The updated head is b385ffe72a; CI should now compare both Qwen-Image and Sana against the correct references.

BBuf · 2026-08-31T09:09:36Z

Applied the quality-gating follow-up in e5b6cdcbfd: the unquantized BF16 added-text QKV packed GEMM is now mounted only for request-scoped quality=high. Default quality=lossless applies the three packed weight slices as three independent reference GEMMs, without duplicating weights; switching back to lossless unmounts the site. On GB300, the lossless/high/unmount path test passed, the full quality-site suite passed (23 tests), and the latest-head import/Qwen selection passed (14 tests). The historical performance/SSIM numbers in the PR body are now explicitly labeled as the high path.

…-fix

BBuf · 2026-08-31T10:49:44Z

/rerun-failed-ci

Co-authored-by: Cursor <cursoragent@cursor.com>

BBuf · 2026-09-01T01:37:53Z

https://github.com/sgl-project/sglang/actions/runs/33398581329/job/99632213429?pr=36680

…ject#36680) Co-authored-by: Mick <mickjagger19@icloud.com> Co-authored-by: Cursor <cursoragent@cursor.com>

…-wrapped layers sgl-project#36680 added _split_unquantized_merged_linear, which reads output_partition_sizes and weight straight off to_added_qkv. Once LoRA wraps the layer (MergedColumnParallelLinearWithLoRA) the attribute is missing, and splitting the base weight would drop the LoRA delta anyway. That broke miles_diffusion test_qwenimage_pickscore_grpo_5xGPU on sglang-miles. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

sgl-project/sglang#36680 packs Qwen-Image's add_{q,k,v}_proj into one MergedColumnParallelLinear (to_added_qkv). The parity patch routed every ColumnParallelLinearWithLoRA through the 2D PEFT delta, which fails on the merged layer's 3D stacked LoRA (size 8 vs 3072). Split the packed weight into three reference GEMMs, each with its own PEFT-ordered delta, which matches the separate linears PEFT trains. 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>

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>

…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>

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>

…-wrapped layers sgl-project#36680 added _split_unquantized_merged_linear, which reads output_partition_sizes and weight straight off to_added_qkv. Once LoRA wraps the layer (MergedColumnParallelLinearWithLoRA) the attribute is missing, and splitting the base weight would drop the LoRA delta anyway. That broke miles_diffusion test_qwenimage_pickscore_grpo_5xGPU on sglang-miles. 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>

github-actions Bot added documentation Improvements or additions to documentation diffusion SGLang Diffusion labels Aug 27, 2026

BBuf added 3 commits August 27, 2026 20:33

perf(diffusion): use custom all-reduce v2 for TP

cb34bdd

Fuse Qwen-Image text QKV projections

d383209

Optimize masked diffusion attention packing

2a37870

BBuf force-pushed the perf/diffusion-custom-allreduce-v2 branch from d46d461 to 2a37870 Compare August 27, 2026 12:34

github-actions Bot added the jit-kernel label Aug 27, 2026

BBuf changed the title ~~perf(diffusion): use CustomAllReduceV2 for TP collectives~~ [Diffusion] Optimize Qwen-Image TP collectives and attention Aug 27, 2026

This was referenced Aug 27, 2026

[Diffusion] Fuse Qwen-Image text QKV projections #36693

Closed

[Diffusion] Optimize segmented varlen attention packing #36694

Closed

BBuf marked this pull request as ready for review August 27, 2026 13:07

BBuf requested review from AgainstEntropy, HaiShaw, JustinTong0323, mickqian, ping1jing2, sogalin, wisclmy0611, yichiche, yingluosanqian and zijiexia as code owners August 27, 2026 13:07

Merge branch 'main' into perf/diffusion-custom-allreduce-v2

8e056cf

BBuf added run-ci CI: run the baseline test suite on this PR run-ci-extra CI: also run the extra suite (requires run-ci) labels Aug 27, 2026

BBuf marked this pull request as draft August 27, 2026 18:27

amd-bot mentioned this pull request Aug 28, 2026

[CI Monitor] Daily Report - 2026-08-28 bingxche/sglang-ci-bot#180

Open

BBuf marked this pull request as ready for review August 28, 2026 01:22

BBuf added 3 commits August 28, 2026 09:22

Merge branch 'main' into perf/diffusion-custom-allreduce-v2

e7d6121

Merge branch 'main' into perf/diffusion-custom-allreduce-v2

d1e5b4e

Merge branch 'main' into perf/diffusion-custom-allreduce-v2

d4edee1

fix(diffusion): preserve Qwen-Image NVFP4 fallback mappings

aa11013

BBuf mentioned this pull request Aug 31, 2026

diffusion-ci: restore main-compatible Sana H100 GT sgl-project/ci-data-diffusion#6

Merged

BBuf added 2 commits August 31, 2026 14:07

test(diffusion): pin corrected Sana consistency GT

1ed919b

Merge main into perf/diffusion-custom-allreduce-v2

b385ffe

Preserve the PR's Qwen-Image NVFP4 fallback mapping while adopting main's transformer override quantization detection.

mickqian and others added 2 commits August 31, 2026 15:44

Merge branch 'main' into perf/diffusion-custom-allreduce-v2

dfedfc1

Gate Qwen-Image fused added QKV by quality

e5b6cdc

Merge remote-tracking branch 'origin/main' into codex/pr36680-ci-data…

272e6d1

…-fix

BBuf and others added 4 commits August 31, 2026 19:15

Merge origin/main to resolve Qwen-Image TP collective test conflicts.

990fd1a

Co-authored-by: Cursor <cursoragent@cursor.com>

fix(diffusion): restore canonical lossless GT pin

e979cfb

Merge origin/main to resolve Qwen-Image TP test conflicts.

a2d110d

Co-authored-by: Cursor <cursoragent@cursor.com>

Merge origin/main to resolve Qwen-Image TP QKV packing conflicts.

45da391

Co-authored-by: Cursor <cursoragent@cursor.com>

BBuf merged commit 71cee04 into sgl-project:main Sep 1, 2026
89 of 181 checks passed

StevenChenSE pushed a commit to StevenChenSE/sglang that referenced this pull request Sep 6, 2026

[Diffusion] Optimize Qwen-Image TP collectives and attention (sgl-pro…

629ff43

…ject#36680) Co-authored-by: Mick <mickjagger19@icloud.com> Co-authored-by: Cursor <cursoragent@cursor.com>

This was referenced Sep 23, 2026

[Diffusion][sglang-miles] Bring sglang-miles-h3 H3 rollout commits onto sglang-miles + LoRA-wrap fixes #40786

Merged

bump SGLang-diffusion to sglang-miles branch radixark/miles_diffusion#257

Merged

Rockdu mentioned this pull request Sep 25, 2026

[diffusion] fix: LoRA-wrapped linears crash Qwen-Image and MiniMax-H3 inference #41272

Merged

1 task done

Navigation Menu

Repository navigation

[Diffusion] Optimize Qwen-Image TP collectives and attention - #36680

[Diffusion] Optimize Qwen-Image TP collectives and attention#36680
BBuf merged 19 commits into
sgl-project:mainfrom
BBuf:perf/diffusion-custom-allreduce-v2

BBuf commented Aug 27, 2026 •

edited by github-actions Bot

Loading

Uh oh!

BBuf commented Aug 31, 2026

Uh oh!

BBuf commented Aug 31, 2026

Uh oh!

BBuf commented Aug 31, 2026

Uh oh!

BBuf commented Sep 1, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

BBuf commented Aug 27, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Request quality semantics

Actual blast radius

Motivation and profiler evidence

Cross-model H200 audit

Method

Successfully executed affected checkpoints

Affected but blocked on this assignment

Negative controls (not attributed to this PR)

Generated output: Base vs PR

Separate vLLM-Omni comparison lane

Correctness and cleanup

Test plan

CI States

Uh oh!

BBuf commented Aug 31, 2026

Uh oh!

BBuf commented Aug 31, 2026

Uh oh!

BBuf commented Aug 31, 2026

Uh oh!

BBuf commented Sep 1, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

BBuf commented Aug 27, 2026 •

edited by github-actions Bot

Loading