SGLang Diffusion Benchmark and Profile

Use this skill when measuring denoise performance, finding the slow op, checking whether an existing fast path can solve it, or verifying that a hotspot is real before any kernel work in sglang.multimodal_gen.

This skill is diagnosis-first. It owns:

• checked-in denoise benchmark presets
• same-GPU quality/BCG applicability checks with repeated exact, lossless, and high rows
• perf dump collection and before/after comparison
• torch.profiler trace capture and quick hotspot ranking
• mapping hot kernels back to known fast paths and fusion families
• packaging confirmed kernel work with enough evidence for the appropriate kernel, Nsight, or framework-specific optimization workflow

This skill does not own low-level kernel authoring or standalone Nsight workflows.

Preflight

Before running any benchmark, profiler, or kernel-validation command:

• use scripts/diffusion_skill_env.py to derive the repo root from sglang.__file__
• verify the repo is writable
• export HF_TOKEN before using gated Hugging Face models such as black-forest-labs/FLUX.*
• export FLASHINFER_DISABLE_VERSION_CHECK=1
• set SGLANG_DIFFUSION_SYNC_STAGE_PROFILING=1 when comparing stage-level denoise/decode timings; the preset helper sets it by default unless the caller explicitly overrides it
• for downloaded checkpoints, use the preset helper's task-owned --model-cache-root together with --cleanup-model-cache; verify the JSONL ledger reports zero residual weight files before moving to the next model
• choose idle GPU(s) before starting perf work; for a comparison matrix, hold the same GPU set and verify it has no foreign process at every run boundary

Native Backend Gate

All diffusion benchmark and profiling results owned by this skill must come from the native SGLang diffusion backend.

Treat any of the following as a hard stop condition:

• Falling back to diffusers backend
• Using diffusers backend
• Loaded diffusers pipeline

If any benchmark, perf-dump, or torch.profiler command prints one of those signals:

• stop the workflow immediately
• do not keep the generated numbers or traces as SGLang benchmark evidence
• do not continue to hotspot classification or kernel work
• first fix model resolution, pipeline selection, overlay/materialization, or other backend-selection issues so the model runs on the native SGLang diffusion path

Main Reference

• benchmark-and-profile.md — canonical denoise benchmark, perf dump, and torch.profiler workflow; uses checked-in nightly-aligned presets plus current-source extras such as LongCat image/edit, Qwen base edit/layered, SD3.5, SANA-Video/SANA-WM, LingBot Video/World, Cosmos3 Edge/Super I2V/distilled and the explicit Super TP2 x CFG2 comparator, LTX-2.5 and its diffusion decoder, MiniMax-H3, FLUX.2 Klein, Ideogram4, ERNIE/GLM/SANA image models, FastWan2.1/2.2, the Blackwell-only Wan2.2 NVFP4 comparator, LTX-2.3, HunyuanVideo, MOVA, Helios, image edit, Hunyuan3D shape, and a separate Pi0.5 action-policy lane
• existing-fast-paths.md — map bottlenecks to existing fused kernels, MoE routing, packed QKV paths, fused QK norm + RoPE, distributed overlap patterns, and open optimization PRs before proposing new code
• scripts/diffusion_skill_env.py — preflight helper: repo root discovery from the skill's owning checkout before falling back to sglang.__file__, write-access probe, benchmark/profile output directories, idle GPU selection
• scripts/bench_diffusion_denoise.py — end-to-end denoise benchmark preset runner via sglang generate; defaults to eager/lossless, supports explicit quality and BCG comparators plus a same-GPU applicability matrix, rejects invalid BCG capture/fallback logs and late high-quality DiT fusion mounts, forces H3 to its eager consistency mode, enables synchronized stage attribution, validates nightly preset drift, and can clean one isolated model cache after the full matrix in a finally block with a JSONL ledger

Opportunity Discovery Rule

Before calling a diffusion hotspot "new", first classify it with existing-fast-paths.md.

Always rule out these existing families first:

• HunyuanVideo VAE GroupNorm+SiLU
• LTX upsampler GroupNorm+SiLU
• Z-Image bf16-native Triton RMSNorm scale/tanh-residual modulation
• SANA packed self-attention Q/K/V and cross-attention K/V GEMMs
• SANA-Video's packed projections and request-scoped BF16-input linear attention at quality=lossless or quality=high; keep the second attention GEMM in FP32 and compare against quality=exact before changing its precision further
• SANA-Video reuse of SANA's bit-exact bias/activation, residual-gate, and LayerNorm-modulation fast paths before adding video-only kernels
• MiniMax-H3 indexed modulation, fused QK norm + RoPE, packed Ulysses QKV, USP relayout, and batched TP AdaLN collectives
• bit-exact diffusion adaLN modulation and fused LayerNorm + modulation for FLUX.1, GLM-Image, and SANA
• request-scoped DiT and VAE fast paths at quality=lossless or quality=high
• LingBot Video's default-on fused group-limited top-k expert selection before treating its router's topk/mask/gather chain as a new hotspot
• Wan causal-VAE cache/padding and DupUp3D data-movement fusions
• fused diffusion QK norm + RoPE
• LTX2 split RoPE
• LTX2 residual-gate add
• LTX-2.5 diffusion-decoder NATTEN selection before interpreting a FlexAttention fallback trace
• varlen USP attention pack/scatter
• NVFP4 / Nunchaku packed QKV
• Nunchaku fused GELU MLP
• Ulysses / USP attention overlap
• turbo-layer async all-to-all overlap
• torch.compile compute / communication reorder
• breakable CUDA graph capture for supported fixed-resolution pipelines
• dual-stream diffusion execution

The checked-in helper defaults to eager. Use --torch-compile only for a controlled comparator, never for the eager ground truth. The legacy --no-torch-compile spelling remains accepted but is redundant.

For kernel/BCG discovery, run --quality-bcg-matrix. It executes Eager/BCG as A-B-B-A at exact, then repeats the pair at lossless and high, on one locked GPU set and one isolated checkpoint cache. The lossless/high+BCG rows are applicability checks, not presumed-valid performance cells. A BCG row is invalid unless the log contains [Diffusion BCG] captured and contains no support-disable, capture-failure, serving-signature-miss, or late quality-fusion marker. In particular, a request-scoped DiT fusion mounted after exact warmup capture would be bypassed by replay; reject that row even when capture and signature checks pass. For video presets, the helper declares both the request resolution and --warmup-num-frames so the synthetic BCG warmup captures the requested temporal shape. Treat any remaining temporal or conditioning signature miss as Eager fallback, not as a valid BCG measurement.

A zero process exit is not sufficient evidence: every accepted row must also contain its requested perf dump and a generated image, video, audio, or 3D mesh file. The helper gives every cell a unique output name and rejects missing artifacts.

On machines with a read-only Hugging Face cache, combine --model-cache-root <task-owned-dir> with one or more --seed-model-cache-root <read-only-HF-home-or-hub> options. The helper exposes cached repos through a task-owned copy-on-write directory overlay, downloads misses only into the isolated cache, and removes links plus new downloads in its normal cleanup finally block without modifying the seed cache.

Keep prompt, negative prompt, seed, shape, steps, guidance, dtype, topology, and residency fixed. Lossless comparisons require byte-identical artifacts. For quality=lossless and quality=high, report aggregate and worst-frame SSIM/PSNR; the repository defaults are 0.95/28 dB for images and 0.92/24 dB for video unless the model's checked-in consistency metadata defines a different threshold. A performance PR needs repeated saved-request e2e improvement of at least 1.5%, a representative profile, and before/after image or video evidence.

MiniMax-H3 is always an eager consistency case on current main. Use --model minimax-h3-t2va; its preset writes the H3 request fields through a generated config and suppresses the helper's global compile default. Do not turn the model's nominal BCG support gate into a performance claim: prompt- dependent packed-sequence host boundaries can differ between warmup and the serving request. A valid H3 BCG experiment must prove that every captured segment replays, keeps the MP4 byte-identical, and does not trade latency for the extra graph memory.

For FLUX-family manual profiling runs with a quantized transformer override:

• use sglang generate directly
• pass the override as --transformer-path <dir>
• prefer --prompt-path <file> when also fixing --output-file-name
• if the base model is already cached locally and the machine has unreliable HF access, use the local cached --model-path plus HF_HUB_OFFLINE=1
• remember that --profile changes latency substantially; use the non-profile perf dump for the real before/after benchmark claim