Update SGLang diffusion skills (#27496)

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-add-model/SKILL.md

@@ -64,6 +64,8 @@ See existing Modular examples: `QwenImagePipeline` (uses `add_standard_t2i_stage
 | Scheduler implementations | `python/sglang/multimodal_gen/runtime/models/schedulers/` |
 | Model/VAE/DiT configs | `python/sglang/multimodal_gen/configs/models/dits/`, `vaes/`, `encoders/` |
 | Central registry | `python/sglang/multimodal_gen/registry.py` |
+| Model component registry | `python/sglang/multimodal_gen/runtime/models/registry.py` |
+| Current support list | `docs/diffusion/compatibility_matrix.md` |
 
 ---
 
@@ -92,7 +94,7 @@ Once you have the reference code, study it thoroughly:
 **Before creating any new files, check whether an existing pipeline or stage can be reused or extended.** Only create new pipelines/stages when the existing ones would require extensive modifications or when no similar implementation exists.
 
 Specifically:
-1. **Compare the new model's architecture against existing pipelines** (Flux, Wan, Qwen-Image, GLM-Image, HunyuanVideo, LTX, etc.). If the new model shares most of its structure with an existing one (e.g., same text encoders, similar latent format, compatible denoising loop), prefer:
+1. **Compare the new model's architecture against existing pipelines** before creating files. Current native families include LTX-2/2.3, HunyuanVideo/FastHunyuan, Wan/FastWan/TurboWan/LingBot World, MOVA, FLUX/FLUX.2/Klein, Z-Image, Qwen-Image/edit/layered, GLM-Image, SD3, Hunyuan3D, Helios, Cosmos3, SANA, FireRed, ERNIE-Image, JoyAI, and Ideogram4. If the new model shares most of its structure with an existing one (e.g., same text encoders, similar latent format, compatible denoising loop), prefer:
    - Adding a new config variant to the existing pipeline rather than creating a new pipeline class
    - Reusing the existing `BeforeDenoisingStage` with minor parameter differences
    - Using `add_standard_t2i_stages()` / `add_standard_ti2i_stages()` / `add_standard_ti2v_stages()` if the model fits standard patterns
@@ -524,7 +526,9 @@ register_configs(
 
 `register_configs()` does not take a `model_family` argument. It registers the
 sampling and pipeline config classes, then resolves models by exact
-`hf_model_paths` or optional detector predicates.
+`hf_model_paths` or optional detector predicates. Prefer exact `hf_model_paths`
+for public checkpoints used in docs or tests; use detector predicates only for
+families where local mirrors, renamed repos, or generated paths are common.
 
 The `EntryClass` in your pipeline file is automatically discovered by the registry's `_discover_and_register_pipelines()` function -- no additional registration needed for the pipeline class itself.
 
@@ -552,6 +556,10 @@ After implementation, **you must verify that the generated output is not noise**
 |-------|----------|---------------------|----------------|
 | GLM-Image | `runtime/pipelines/glm_image.py` | `stages/model_specific_stages/glm_image.py` | `configs/pipeline_configs/glm_image.py` |
 | Qwen-Image-Layered | `runtime/pipelines/qwen_image.py` (`QwenImageLayeredPipeline`) | `stages/model_specific_stages/qwen_image_layered.py` | `configs/pipeline_configs/qwen_image.py` (`QwenImageLayeredPipelineConfig`) |
+| Cosmos3 | `runtime/pipelines/cosmos3_pipeline.py` | `stages/model_specific_stages/cosmos3.py` | `configs/pipeline_configs/cosmos3.py` |
+| ErnieImage | `runtime/pipelines/ernie_image.py` | `runtime/pipelines/ernie_image.py` | `configs/pipeline_configs/ernie_image.py` |
+| Hunyuan3D | `runtime/pipelines/hunyuan3d_pipeline.py` | `stages/model_specific_stages/hunyuan3d.py` | `configs/pipeline_configs/hunyuan3d.py` |
+| LingBot World realtime | `runtime/pipelines/lingbot_world_causal_dmd_pipeline.py` | `stages/model_specific_stages/lingbot_world/` | `configs/pipeline_configs/lingbot_world.py` |
 
 ### Modular Style (when standard stages fit well)
 
@@ -560,6 +568,10 @@ After implementation, **you must verify that the generated output is not noise**
 | Qwen-Image (T2I) | `runtime/pipelines/qwen_image.py` | Uses `add_standard_t2i_stages()` — standard text encoding + latent prep fits this model |
 | Qwen-Image-Edit | `runtime/pipelines/qwen_image.py` | Uses `add_standard_ti2i_stages()` — standard image-to-image flow |
 | Flux | `runtime/pipelines/flux.py` | Uses `add_standard_t2i_stages()` with custom `prepare_mu` |
+| FLUX.2 / FLUX.2 Klein | `runtime/pipelines/flux_2.py`, `flux_2_klein.py` | Reuses FLUX.2 stages; Klein differences live in config and sampling params |
+| Z-Image | `runtime/pipelines/zimage_pipeline.py` | Uses standard image pipeline stages plus Z-Image-specific config/model code |
+| Ideogram4 | `runtime/pipelines/ideogram.py` | Uses dedicated text encoding and denoising stages while keeping standard latent prep |
+| SANA | `runtime/pipelines/sana.py` | Spatial image pipeline; reuse the spatial image config pattern |
 | Wan | `runtime/pipelines/wan_pipeline.py` | Uses `add_standard_ti2v_stages()` |
 
 ---

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-ako4all-kernel/SKILL.md

@@ -1,145 +0,0 @@
----
-name: sglang-diffusion-ako4all-kernel
-description: Use when optimizing an existing SGLang diffusion kernel with AKO4ALL, including AKO4ALL repo hygiene, custom microbench setup, ncu-guided iteration, and end-to-end denoise validation. Also use when a sibling AKO4ALL repo must be cloned or refreshed before starting kernel tuning work.
----
-
-# SGLang Diffusion AKO4ALL Kernel
-
-Use this skill to run the full AKO4ALL-based optimization loop for an existing SGLang diffusion kernel.
-It is the default implementation path once the benchmark/profile skill has already shown that a hotspot is real and not covered by an existing fast path. This workflow bootstraps a custom AKO harness, benchmarks and profiles the kernel, iterates with `ncu`, ports the best version back to `sglang`, then validates with targeted tests and model-level denoise runs.
-
-This skill assumes a sibling repo layout like:
-
-```text
-<base-dir>/
-├── sglang/
-└── AKO4ALL/
-```
-
-If `AKO4ALL/` is missing under the current base directory, clone it first.
-
-## Use This Skill When
-
-- tuning an existing diffusion Triton, CUDA JIT, CuTeDSL, or runtime-integrated kernel in `sglang`
-- `sglang-diffusion-benchmark-profile` has already ruled out an existing in-repo fast path or overlap family
-- creating a custom AKO4ALL harness for a real diffusion kernel instead of using the default benchmark tasks
-- validating that a kernel-level win transfers to Qwen, FLUX, Wan, Hunyuan, MOVA, or other diffusion denoise latency
-- preparing PR artifacts such as microbench tables, `ncu` before/after data, and proof image outputs
-
-Do not start here when the bottleneck has not been proven yet.
-First use [../sglang-diffusion-benchmark-profile/SKILL.md](../sglang-diffusion-benchmark-profile/SKILL.md) to:
-- measure the real denoise regression
-- collect the perf dump baseline
-- capture one representative `torch.profiler` trace
-- rule out existing mainline fast paths
-- prove the run stayed on the native SGLang diffusion backend, not a diffusers fallback
-
-Before opening AKO, also read
-[../sglang-diffusion-benchmark-profile/existing-fast-paths.md](../sglang-diffusion-benchmark-profile/existing-fast-paths.md).
-It records current mainline fusions plus the open PR watchlist for diffusion
-kernel, VAE, attention, cache, and scheduling work. If an open PR already covers
-the same shape family, use it as prior art or decide whether to rebase/extend it
-instead of starting a duplicate kernel.
-
-If a future specialized optimization skill matches the kernel family better than AKO4ALL, hand off there instead. The diagnosis contract stays the same.
-
-## Mandatory AKO4ALL Preflight
-
-Before any AKO work:
-
-1. Run `scripts/ensure_ako4all_clean.sh [base-dir]`.
-2. If `<base-dir>/AKO4ALL` does not exist, the script clones it.
-3. Do not continue unless `AKO4ALL` is:
-   - on the upstream default branch, usually `main`
-   - fully clean with no tracked or untracked local changes
-   - exactly synced to `upstream/<default-branch>`
-4. If the script reports local commits, divergence, or a dirty worktree, stop and clean or re-clone the repo before continuing.
-
-The script creates an `upstream` remote automatically when missing.
-By default it uses the existing `origin` URL, or `AKO4ALL_URL` if you need to override the clone source.
-
-## Workflow
-
-### 1. Scope the Kernel
-
-- Identify the exact kernel entry point and runtime call sites in `sglang`.
-- Record the target shapes, dtypes, model families, and whether the kernel is on a hot path.
-- Reuse existing unit tests and benchmark entry points when they already exist.
-- Record whether the hotspot overlaps an open PR from `existing-fast-paths.md`;
-  if it does, note the PR number in the AKO context and final PR artifacts.
-
-### 2. Bootstrap the AKO Harness
-
-Inside the clean `AKO4ALL` repo:
-
-- read `TASK.md` and `HINTS.md`
-- create a custom harness instead of relying on the stock benchmark tasks
-- mirror the real SGLang kernel into:
-  - `input/reference.py`
-  - `input/<kernel>.py`
-  - `solution/<kernel>.py`
-  - `bench/bench_<kernel>.py`
-- keep a short context note in `context/` when the kernel has model-specific shape assumptions or perf conclusions
-
-The custom benchmark should:
-
-- cover representative diffusion shapes
-- check correctness against the reference kernel
-- report aggregate runtime plus per-shape results when useful
-
-### 3. Establish the Baseline
-
-- run the AKO custom microbench before changing the kernel
-- capture one representative `ncu` baseline on the hottest meaningful shape
-- note whether the bottleneck looks like registers, occupancy, instruction count, launch config, or memory latency
-
-### 4. Iterate in AKO4ALL
-
-- change one idea at a time
-- rerun the microbench after every change
-- update `ITERATIONS.md` with hypothesis, result, and next step
-- prefer simple, explainable wins over clever rewrites that do not transfer
-
-After 3 consecutive no-improvement or regression iterations:
-
-- rerun `ncu`
-- re-read `ITERATIONS.md`
-- change direction instead of continuing blind sweeps
-
-### 5. Port the Best Version Back to SGLang
-
-- apply the best candidate to the real `sglang` kernel file
-- run import or syntax checks and targeted tests first
-- keep the AKO `solution/` version aligned with the main-tree version you actually want to keep
-
-### 6. Validate on Real Models
-
-- use the benchmark/profile skill for denoise perf dumps and before/after comparison
-- prefer exact local snapshot validation when testing local edits on a GPU box
-- run targeted kernel tests first
-- run model-level denoise benchmarks with perf dumps
-- compare baseline vs optimized runs with `compare_perf.py`
-- if the PR needs proof that generation still works, save one real model output image
-
-### 7. Prepare PR Artifacts
-
-At minimum, keep:
-
-- one microbench table
-- one denoise-stage table
-- one end-to-end table
-- one `ncu` before/after pair on the most representative kernel shape
-- one generated image when the kernel affects production inference
-
-See [references/ako-loop.md](references/ako-loop.md) for the checklist and common stop rules.
-
-## Operating Rules
-
-- Treat AKO4ALL repo hygiene as a gate, not a suggestion.
-- Prefer exact local snapshot validation over hand-wavy “remote tree is close enough”.
-- Do not start or justify kernel work from traces collected after
-  `Falling back to diffusers backend`, `Using diffusers backend`, or
-  `Loaded diffusers pipeline`; fix backend selection and rerun the
-  benchmark/profile workflow first.
-- Keep model-level validation honest: if microbench improves but denoise does not, do not keep the AKO-only variant in the main code path.
-- When writing conclusions, explain the win in terms of measurable causes such as lower registers per thread, higher occupancy, fewer executed instructions, or better scheduler eligibility.

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-ako4all-kernel/references/ako-loop.md

@@ -1,54 +0,0 @@
-# AKO Loop Checklist
-
-Use this checklist after `scripts/ensure_ako4all_clean.sh` succeeds.
-
-## Minimum Repo Layout
-
-Inside `AKO4ALL/`, prefer these files for a diffusion kernel task:
-
-- `input/reference.py`
-- `input/<kernel>.py`
-- `solution/<kernel>.py`
-- `bench/bench_<kernel>.py`
-- `context/<kernel>_notes.md`
-
-## Baseline Checklist
-
-- Reproduce the current SGLang kernel exactly in AKO first.
-- Run the custom microbench before making edits.
-- Record one representative `ncu` report on a real hot shape.
-- Note the baseline bottleneck in plain language.
-
-## Iteration Discipline
-
-- One optimization idea per iteration.
-- Re-benchmark after every code change.
-- Log the result in `ITERATIONS.md`.
-- Keep the best candidate easy to identify.
-
-Stop a direction early when:
-
-- 3 consecutive iterations do not beat the best runtime
-- correctness gets fragile
-- AKO-only gains stop transferring to real denoise runs
-
-## Real Validation Gate
-
-Before calling a kernel "done", validate all of:
-
-- syntax or import checks
-- targeted unit test or regression test
-- kernel or op-level benchmark
-- model-level denoise benchmark with perf dumps
-- one generated image if the PR needs production proof
-
-## PR Artifact Checklist
-
-Prepare these artifacts:
-
-- microbench table
-- denoise-stage table
-- end-to-end table
-- one `ncu` before or after pair
-- one short explanation of why the kernel got faster
-- one generated output image when applicable

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-ako4all-kernel/scripts/ensure_ako4all_clean.sh

@@ -1,84 +0,0 @@
-#!/usr/bin/env bash
-set -euo pipefail
-
-BASE_DIR="${1:-$PWD}"
-BASE_DIR="$(cd "$BASE_DIR" && pwd)"
-AKO_DIR="${BASE_DIR}/AKO4ALL"
-CANONICAL_UPSTREAM_URL="https://github.com/TongmingLAIC/AKO4ALL.git"
-UPSTREAM_URL="${AKO4ALL_UPSTREAM_URL:-$CANONICAL_UPSTREAM_URL}"
-CLONE_URL="${AKO4ALL_URL:-$UPSTREAM_URL}"
-
-say() {
-  printf '[ako4all] %s\n' "$*"
-}
-
-fail() {
-  printf '[ako4all] ERROR: %s\n' "$*" >&2
-  exit 1
-}
-
-if [[ ! -d "$AKO_DIR/.git" ]]; then
-  say "AKO4ALL not found under ${BASE_DIR}; cloning ${CLONE_URL}"
-  git clone "$CLONE_URL" "$AKO_DIR"
-fi
-
-cd "$AKO_DIR"
-
-if ! git remote get-url origin >/dev/null 2>&1; then
-  fail "AKO4ALL exists but has no origin remote."
-fi
-
-if ! git remote get-url upstream >/dev/null 2>&1; then
-  say "Adding missing upstream remote -> ${UPSTREAM_URL}"
-  git remote add upstream "$UPSTREAM_URL"
-fi
-
-git fetch upstream --prune
-git remote set-head upstream -a >/dev/null 2>&1 || true
-
-default_branch="${AKO4ALL_BRANCH:-}"
-if [[ -z "$default_branch" ]]; then
-  if upstream_head="$(git symbolic-ref --quiet --short refs/remotes/upstream/HEAD 2>/dev/null)"; then
-    default_branch="${upstream_head#upstream/}"
-  else
-    default_branch="main"
-  fi
-fi
-
-if [[ -n "$(git status --porcelain)" ]]; then
-  fail "AKO4ALL worktree is dirty. Clean all local changes before using this skill."
-fi
-
-if git show-ref --verify --quiet "refs/heads/${default_branch}"; then
-  git switch "$default_branch" >/dev/null
-else
-  git switch -c "$default_branch" --track "upstream/${default_branch}" >/dev/null
-fi
-
-git fetch upstream --prune
-
-local_head="$(git rev-parse HEAD)"
-upstream_head="$(git rev-parse "upstream/${default_branch}")"
-
-if [[ "$local_head" != "$upstream_head" ]]; then
-  if git merge-base --is-ancestor "$local_head" "$upstream_head"; then
-    say "Fast-forwarding ${default_branch} to upstream/${default_branch}"
-    git merge --ff-only "upstream/${default_branch}" >/dev/null
-  else
-    fail "Local ${default_branch} diverges from upstream/${default_branch}. Reset or re-clone AKO4ALL before continuing."
-  fi
-fi
-
-if [[ -n "$(git status --porcelain)" ]]; then
-  fail "AKO4ALL became dirty after sync; stop and inspect the repo."
-fi
-
-final_head="$(git rev-parse HEAD)"
-expected_head="$(git rev-parse "upstream/${default_branch}")"
-if [[ "$final_head" != "$expected_head" ]]; then
-  fail "AKO4ALL is not exactly at upstream/${default_branch}."
-fi
-
-say "Ready: ${AKO_DIR}"
-say "Branch: ${default_branch}"
-say "Commit: ${final_head}"

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-benchmark-profile/SKILL.md

@@ -12,7 +12,7 @@ This skill is diagnosis-first. It owns:
 - 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
-- handing confirmed kernel work to a specialized optimization skill such as [../sglang-diffusion-ako4all-kernel/SKILL.md](../sglang-diffusion-ako4all-kernel/SKILL.md)
+- 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.
 
@@ -42,7 +42,7 @@ If any benchmark, perf-dump, or `torch.profiler` command prints one of those sig
 
 ## Main Reference
 
-- [benchmark-and-profile.md](benchmark-and-profile.md) — canonical denoise benchmark, perf dump, and `torch.profiler` workflow; uses checked-in nightly-aligned presets plus skill-only stress recipes such as `LTX-2.3` one-stage/two-stage, HunyuanVideo, MOVA, Helios, JoyAI/FireRed image edit, and Hunyuan3D shape
+- [benchmark-and-profile.md](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 FLUX.2 Klein, Cosmos3, Ideogram4, ERNIE/GLM/SANA image models, FastWan2.2, `LTX-2.3` one-stage/two-stage/HQ, HunyuanVideo, MOVA, Helios, JoyAI/FireRed image edit, and Hunyuan3D shape
 - [existing-fast-paths.md](existing-fast-paths.md) — map bottlenecks to existing fused kernels, packed QKV paths, fused `QK norm + RoPE`, distributed overlap patterns, and open optimization PRs before proposing new code
 - [scripts/diffusion_skill_env.py](scripts/diffusion_skill_env.py) — preflight helper: repo root discovery via `sglang.__file__`, write-access probe, benchmark/profile output directories, idle GPU selection
 - [scripts/bench_diffusion_denoise.py](scripts/bench_diffusion_denoise.py) — end-to-end denoise benchmark preset runner via `sglang generate`; supports `--no-torch-compile`, validates nightly preset drift with `--validate-nightly-alignment`, and saves perf dumps by label for `compare_perf.py`
@@ -53,8 +53,11 @@ Before calling a diffusion hotspot "new", first classify it with `existing-fast-
 
 Always rule out these existing families first:
 - HunyuanVideo VAE GroupNorm+SiLU
+- LTX upsampler GroupNorm+SiLU
 - Z-Image residual-form modulation
 - fused diffusion `QK norm + RoPE`
+- LTX2 split RoPE
+- varlen USP attention pack/scatter
 - NVFP4 / Nunchaku packed QKV
 - Nunchaku fused GELU MLP
 - Ulysses / USP attention overlap

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-benchmark-profile/benchmark-and-profile.md

@@ -21,10 +21,10 @@ This guide intentionally stops at:
 - hotspot ranking
 - mapping hotspots to known fast paths
 
-If the hotspot survives this checklist, hand the work to
-`sglang-diffusion-ako4all-kernel` or another specialized kernel-optimization
-skill. Do not grow this skill back into a general Nsight or kernel-authoring
-guide.
+If the hotspot survives this checklist, package the perf dump, profiler trace,
+exact command, and shape/topology notes for the appropriate kernel, Nsight, or
+framework-specific optimization workflow. Do not grow this skill back into a
+general Nsight or kernel-authoring guide.
 
 ## Prerequisites
 
@@ -163,7 +163,8 @@ PYTHONPATH=python python3 "$BENCH_PY" \
   --output-dir "${BENCH_DIR}"
 ```
 
-Run the full preset sweep:
+Run the full preset sweep only when you have enough GPU time for both the
+nightly-aligned cases and the source-tracked extras:
 
 ```bash
 PYTHONPATH=python python3 "$BENCH_PY" \
@@ -172,7 +173,21 @@ PYTHONPATH=python python3 "$BENCH_PY" \
   --output-dir "${BENCH_DIR}"
 ```
 
-Nightly-aligned presets come first; skill-only presets stay available after them.
+Nightly-aligned presets come first, followed by current-source extras from the
+registry / GPU test cases, then broader skill-only stress presets.
+
+Use the preset categories this way:
+- **Nightly-aligned**: exact mirrors of
+  `scripts/ci/utils/diffusion/comparison_configs.json`; use these when the goal
+  is apples-to-apples comparison with CI / nightly coverage.
+- **Current-source extras**: models or request shapes with explicit support
+  evidence in the current registry, GPU cases, compatibility matrix, pipeline
+  files, or unit tests, but without a nightly comparison case yet.
+- **Skill-only stress / coverage presets**: extra profiling scenarios kept by
+  this skill to stress a topology, high-resolution path, multi-GPU mode, or
+  model-specific stage. These may be older than the latest registry additions,
+  so re-check the active source tree before treating them as support-matrix
+  commitments.
 
 | Preset | Model | Nightly | Notes |
 | --- | --- | --- | --- |
@@ -186,6 +201,19 @@ Nightly-aligned presets come first; skill-only presets stay available after them
 | `ltx2` | `Lightricks/LTX-2` | Yes: `ltx2_twostage_t2v` | `LTX2TwoStagePipeline`, 2 GPUs, CFG parallel, 768x512, 121 frames, seed 42 |
 | `ltx23-ti2v-two-stage` | `Lightricks/LTX-2.3` | Yes: `ltx2.3_twostage_ti2v_2gpus` | Nightly cat image, motion prompt, `LTX2TwoStagePipeline`, 2 GPUs, `--cfg-parallel-size 2`, 768x512, 121 frames, seed 42 |
 | `wan-i2v` | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | Yes: `wan22_i2v_a14b_720p` | Nightly cat image and motion prompt, 1280x720, 81 frames, 4 GPUs, CFG parallel, Ulysses degree 2, text encoder CPU offload and pinned CPU memory |
+| `qwen-image` | `Qwen/Qwen-Image` | No | Current-source extra covering the base Qwen-Image native path, separate from the nightly `Qwen-Image-2512` case |
+| `qwen-edit-2509` | `Qwen/Qwen-Image-Edit-2509` | No | Current-source extra for the pre-2511 edit-plus path; uses the cat image, 1024x1024 |
+| `zimage-base` | `Tongyi-MAI/Z-Image` | No | Current-source extra for non-turbo Z-Image; keep it separate from `zimage` / `Z-Image-Turbo` |
+| `flux2-klein` | `black-forest-labs/FLUX.2-klein-4B` | No | Current-source extra for the distilled FLUX.2 Klein path; gated repo, 1024x1024, DiT layerwise offload disabled |
+| `flux2-klein-base` | `black-forest-labs/FLUX.2-klein-base-4B` | No | Current-source extra for the undistilled FLUX.2 Klein Base path; gated repo, 1024x1024, DiT layerwise offload disabled |
+| `cosmos3-nano-t2i` | `nvidia/Cosmos3-Nano` | No | Current-source extra for the single-frame Cosmos3 image path; sets `SGLANG_DISABLE_COSMOS3_GUARDRAILS=1` in the helper environment |
+| `cosmos3-nano-t2v` | `nvidia/Cosmos3-Nano` | No | Current-source extra for a short Cosmos3 video path; sets `SGLANG_DISABLE_COSMOS3_GUARDRAILS=1` in the helper environment |
+| `ideogram4-fp8` | `ideogram-ai/ideogram-4-fp8` | No | Current-source extra matching the native Ideogram 4 FP8 pipeline; do not override steps/guidance directly because the sampling preset owns them |
+| `ernie-image-turbo` | `baidu/ERNIE-Image-Turbo` | No | Current-source extra for ERNIE-Image Turbo |
+| `glm-image` | `zai-org/GLM-Image` | No | Current-source extra for GLM-Image |
+| `sana-1.5-1.6b` | `Efficient-Large-Model/SANA1.5_1.6B_1024px_diffusers` | No | Current-source extra for a SANA native image path |
+| `fastwan22-ti2v-5b` | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | No | Current-source extra matching the FastWan2.2 TI2V registered path |
+| `ltx23-hq-two-stage` | `Lightricks/LTX-2.3` | No | Current-source extra for `LTX2TwoStageHQPipeline` with `snapshot` device mode; high-resolution and VRAM-heavy |
 | `ltx23-one-stage` | `Lightricks/LTX-2.3` | No | Skill-only extra preset for the native `LTX-2.3` one-stage baseline; 2 GPUs, 768x512, 121 frames, fps 24, 30 steps, guidance 3.0, seed 1234 |
 | `ltx23-two-stage` | `Lightricks/LTX-2.3` | No | Skill-only high-resolution stress preset for the native `LTX-2.3` two-stage path; uses `LTX2TwoStagePipeline`, 2 GPUs, 1536x1024, 121 frames, fps 24, 30 steps, guidance 3.0, seed 1234 |
 | `ltx23-two-stage-cfg-parallel` | `Lightricks/LTX-2.3` | No | Skill-only high-resolution CFG-parallel stress preset matching `ltx23-two-stage` plus `--cfg-parallel-size 2` |
@@ -461,6 +489,8 @@ the known mainline families.
 | `fused_inplace_qknorm_rope` missing, but separate qk norm plus rope show up | Check whether the fused diffusion `QK norm + RoPE` path should have engaged |
 | `to_q -> to_k -> to_v` on NVFP4 or Nunchaku FLUX-family checkpoints | Treat as a packed-QKV fast-path miss or checkpoint-format mismatch |
 | `fused_norm_tanh_mul_add*` missing on Z-Image | Treat as a missing mainline modulation path, not a new fusion request |
+| LTX-2 split RoPE appears as a long PyTorch elementwise chain | Check the `apply_ltx2_split_rotary_emb` Triton path and its shape guards |
+| masked attention spends time packing/unpacking Q/K/V | Check whether fused varlen USP pack/scatter should have engaged |
 | `all_to_all`, ring attention, or async A2A dominate | Classify against Ulysses, USP, or turbo-layer overlap first |
 | split `fc1 -> gelu -> quant -> fc2.lora_down` on Nunchaku FLUX | Treat as a missing fused GELU MLP path |
 | attention kernels dominate | Confirm backend, topology, and shape guards before proposing a new kernel |
@@ -475,7 +505,7 @@ Only after the hotspot survives the fast-path checklist:
 1. save a baseline perf dump
 2. save a representative `torch.profiler` trace
 3. note the exact model, shape, dtype, and GPU topology
-4. hand the work to `sglang-diffusion-ako4all-kernel` or another future specialized optimization skill
+4. hand the work to the appropriate kernel, Nsight, or framework-specific optimization workflow
 
 This skill intentionally stops here. It tells you whether you are looking at:
 - a missing existing optimization
@@ -490,4 +520,4 @@ This skill intentionally stops here. It tells you whether you are looking at:
 - [ ] one representative `torch.profiler` trace saved
 - [ ] hotspot classified against `existing-fast-paths.md`
 - [ ] reference image or video checked for correctness
-- [ ] any remaining kernel work handed to a specialized optimization skill
+- [ ] any remaining kernel work handed off with perf/profile evidence attached

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-benchmark-profile/existing-fast-paths.md

@@ -2,11 +2,13 @@
 
 Use this guide when mapping a diffusion bottleneck to an existing fused path or
 distributed overlap pattern in `sglang.multimodal_gen`. Prefer reuse and
-configuration first before handing the problem to a specialized kernel-optimization skill.
+configuration first before handing the problem to a kernel, Nsight, or
+framework-specific optimization workflow.
 
 **Key Files**
 - `python/sglang/multimodal_gen/runtime/layers/layernorm.py`
 - `python/sglang/multimodal_gen/runtime/layers/elementwise.py`
+- `python/sglang/multimodal_gen/runtime/layers/fused_scale_shift_gate.py`
 - `python/sglang/multimodal_gen/runtime/layers/rotary_embedding/utils.py`
 - `python/sglang/jit_kernel/diffusion/triton/scale_shift.py`
 - `python/sglang/jit_kernel/diffusion/group_norm_silu.py`
@@ -14,8 +16,14 @@ configuration first before handing the problem to a specialized kernel-optimizat
 - `python/sglang/jit_kernel/diffusion/triton/norm.py`
 - `python/sglang/jit_kernel/diffusion/triton/rmsnorm_onepass.py`
 - `python/sglang/jit_kernel/diffusion/triton/rotary.py`
+- `python/sglang/jit_kernel/diffusion/triton/ltx2_rotary.py`
+- `python/sglang/jit_kernel/diffusion/triton/varlen_pack_pad.py`
 - `python/sglang/jit_kernel/diffusion/cutedsl/scale_residual_norm_scale_shift.py`
+- `python/sglang/jit_kernel/tests/diffusion/test_qwen_image_modulation.py`
 - `python/sglang/jit_kernel/tests/diffusion/test_group_norm_silu.py`
+- `python/sglang/jit_kernel/tests/diffusion/test_varlen_pack_pad.py`
+- `python/sglang/jit_kernel/tests/diffusion/test_varlen_uspattn_equivalence.py`
+- `python/sglang/jit_kernel/benchmark/diffusion/bench_qwen_image_modulation.py`
 - `python/sglang/jit_kernel/benchmark/diffusion/bench_group_norm_silu.py`
 - `python/sglang/jit_kernel/norm.py`
 - `python/sglang/multimodal_gen/runtime/platforms/cuda.py`
@@ -24,10 +32,10 @@ configuration first before handing the problem to a specialized kernel-optimizat
 
 **Core Fusion Patterns**
 
-1. Scale/Shift elementwise fusion (AdaLN modulation)
-- Kernels: `fuse_scale_shift_kernel`, `fuse_scale_shift_gate_select01_kernel`
-- Locations: `elementwise.py`, `layernorm.py`, `qwen_image.py`, `triton/scale_shift.py`
-- Use cases: `x * (1 + scale) + shift` and `a * (k + b) + c`
+1. Scale/Shift elementwise and gate fusion (AdaLN modulation)
+- Kernels: `fuse_scale_shift_kernel`, `fuse_layernorm_scale_shift_gate_select01_kernel`, `fuse_residual_layernorm_scale_shift_gate_select01_kernel`
+- Locations: `elementwise.py`, `layernorm.py`, `fused_scale_shift_gate.py`, `qwen_image.py`, `triton/scale_shift.py`
+- Use cases: `x * (1 + scale) + shift`, `a * (k + b) + c`, and Qwen-style `(layernorm/residual layernorm) + scale/shift + gate select`.
 - Constraints: `x` must be CUDA and contiguous. `scale/shift` support 0D/1D/2D/3D/4D broadcast. 4D `[B, F, 1, C]` requires `L % F == 0`.
 - NPU fallback: `scale_shift.py` swaps to `npu_fallback` native path.
 - Validation: `python/sglang/jit_kernel/tests/diffusion/test_qwen_image_modulation.py`.
@@ -72,11 +80,18 @@ configuration first before handing the problem to a specialized kernel-optimizat
 - NPU fallback: `npu_fallback.apply_rotary_embedding_native`.
 - Validation: `python/sglang/jit_kernel/tests/test_rope.py`.
 
-7. HunyuanVideo VAE GroupNorm + SiLU fusion
+7. LTX2 split RoPE fusion
+- Kernel: `apply_ltx2_split_rotary_emb`
+- Locations: `triton/ltx2_rotary.py`, `runtime/models/dits/ltx_2.py`
+- Use case: LTX-2 split rotary embedding over `[B, S, num_heads * head_dim]` with separate `cos` and `sin` tensors.
+- Constraints: `cos` and `sin` shapes must match `[B, H, S, head_dim / 2]`, and `inner_dim == H * head_dim`.
+- Workflow rule: if LTX-2 traces show a large split-RoPE PyTorch chain, check whether the LTX2-specific Triton path was disabled by shape or dtype before proposing a new RoPE kernel.
+
+8. HunyuanVideo / LTX upsampler GroupNorm + SiLU fusion
 - Kernel: `triton_group_norm_silu`
-- Locations: `diffusion/group_norm_silu.py`, `triton/group_norm_silu.py`, `runtime/models/vaes/hunyuanvae.py`
+- Locations: `diffusion/group_norm_silu.py`, `triton/group_norm_silu.py`, `runtime/models/vaes/hunyuanvae.py`, `runtime/models/upsampler/latent_upsampler.py`
 - Use case: `activation(group_norm(x))` when the activation is non-inplace `nn.SiLU` and the GroupNorm is affine.
-- Enablement: mainline uses `apply_group_norm_silu(...)` in HunyuanVideo VAE paths by default; there is no env toggle. The wrapper dispatches to Triton only when guards pass.
+- Enablement: mainline uses `apply_group_norm_silu(...)` in HunyuanVideo VAE paths and LTX latent upsampler paths by default; there is no env toggle. The wrapper dispatches to Triton only when guards pass.
 - Constraints: CUDA inference path only; no grad, `x.requires_grad == False`, `nn.GroupNorm`, `nn.SiLU(inplace=False)`, affine norm with weight and bias. Unsupported cases fall back to native `activation(norm(x))`.
 - Validation: `python/sglang/jit_kernel/tests/diffusion/test_group_norm_silu.py`.
 - Microbench: `python/sglang/jit_kernel/benchmark/diffusion/bench_group_norm_silu.py`.
@@ -99,6 +114,12 @@ configuration first before handing the problem to a specialized kernel-optimizat
 - Location: `rotary_embedding/utils.py`
 - Behavior: `flashinfer.rope.apply_rope_with_cos_sin_cache_inplace` when available, otherwise Triton RoPE fallback.
 
+4. Varlen USP attention pack/scatter
+- Locations: `runtime/layers/attention/layer.py`, `triton/varlen_pack_pad.py`
+- Behavior: masked `USPAttention.forward` can gather dense Q/K/V into packed `[total_valid, H, D]` rows with `fused_pack_qkv`, run varlen attention, then scatter back with `fused_scatter_to_padded`.
+- Validation: `python/sglang/jit_kernel/tests/diffusion/test_varlen_pack_pad.py` and `test_varlen_uspattn_equivalence.py`.
+- Workflow rule: if a masked attention trace spends time in Python/advanced indexing pack or scatter, first check whether this fused varlen path should have engaged.
+
 **QK Norm Optimization**
 
 - Entry point: `apply_qk_norm` in `layernorm.py`.
@@ -125,7 +146,7 @@ configuration first before handing the problem to a specialized kernel-optimizat
   - Supported head dims: `64, 128, 256`.
 - Behavior: `apply_qk_norm_rope` prefers the fused JIT kernel when all guards pass; otherwise it falls back to `apply_qk_norm(...)` plus `apply_flashinfer_rope_qk_inplace(...)`.
 - Validation: `python/sglang/jit_kernel/tests/diffusion/test_qknorm_rope.py`.
-- Watchlist: PR #24025 adds LTX2-specific QK norm fusion work. Until merged, treat LTX2 traces that miss the generic fused path as an enablement/shape-guard issue first, not proof that the PR path exists locally.
+- Workflow rule: treat LTX2 traces that miss the generic fused path as an enablement/shape-guard issue first, and check the separate LTX2 split-RoPE path before proposing new attention-prep kernels.
 
 **Nunchaku Fused GELU MLP**
 
@@ -144,11 +165,13 @@ configuration first before handing the problem to a specialized kernel-optimizat
 
 **Common Entry Points in Diffusion Models**
 - AdaLN modulation: `LayerNormScaleShift`, `RMSNormScaleShift`, `ScaleResidual*` in `layernorm.py`.
-- Qwen-Image gating: `fuse_scale_shift_gate_select01_kernel` in `qwen_image.py`.
+- Qwen-Image gating: `fuse_layernorm_scale_shift_gate_select01_kernel` and `fuse_residual_layernorm_scale_shift_gate_select01_kernel` through `fused_scale_shift_gate.py` and `qwen_image.py`.
 - Z-Image residual-form modulation: `fused_norm_tanh_mul_add` and `fused_norm_tanh_mul_add_norm_scale` in `zimage.py`.
-- HunyuanVideo VAE GroupNorm+SiLU: `apply_group_norm_silu` in `hunyuanvae.py`; default-eligible when wrapper guards pass.
+- HunyuanVideo VAE and LTX upsampler GroupNorm+SiLU: `apply_group_norm_silu` in `hunyuanvae.py` and `latent_upsampler.py`; default-eligible when wrapper guards pass.
 - QK norm: `apply_qk_norm` used in `flux.py`, `flux_2.py`, `qwen_image.py`, `zimage.py`, `wanvideo.py`, `ltx_2.py`, `hunyuanvideo.py`.
 - QK norm + RoPE: `apply_qk_norm_rope` in `layernorm.py`; use this path when the model wants fused attention prep instead of separate QK norm and RoPE calls.
+- LTX2 split RoPE: `apply_ltx2_split_rotary_emb` in `ltx_2.py`.
+- Varlen USP attention: `fused_pack_qkv` and `fused_scatter_to_padded` in `attention/layer.py`.
 - Nunchaku fused GELU MLP: `_fused_gelu_mlp` in `flux.py` for quantized FLUX-family checkpoints.
 - NVFP4 / packed QKV attention: `to_qkv`, `to_added_qkv`, and `to_qkv_mlp_proj` in FLUX-family quantized paths.
 - RoPE: `_apply_rotary_emb` prefers Triton; Q/K RoPE prefers FlashInfer when present.
@@ -161,11 +184,11 @@ configuration first before handing the problem to a specialized kernel-optimizat
 - Dual-stream diffusion models: `use_dual_stream = True` in models such as `hunyuan3d.py` is an existing overlap family.
 - Workflow rule: if a hotspot is communication-heavy, rule out these in-repo overlap families before proposing a brand new overlap design.
 
-**Open PR Watchlist**
+**Historical PR Watchlist**
 
-As of 2026-05-02, these SGLang PRs were still open. Use them as upstream
-direction and prior art, not as current-main behavior. Re-check the PR state
-before relying on any file path or flag.
+These SGLang PRs are useful as upstream direction and prior art, not as
+current-main behavior. Re-check the PR state and the active source tree before
+relying on any file path, flag, or claim about whether the work has merged.
 
 - Norm, modulation, and packed projection fusions:
   - #24025 LTX2 QK norm fusion.
@@ -211,4 +234,4 @@ before relying on any file path or flag.
 - Keep CuTe compile cache keys aligned to `(dtype, ndim, D)`.
 - Avoid implicit broadcasts that force hidden `contiguous()` copies.
 - Preserve NPU and ROCm fallback paths.
-- If none of the families above match, package the evidence from the benchmark/profile skill and hand the kernel work to a specialized optimization skill such as `sglang-diffusion-ako4all-kernel`.
+- If none of the families above match, package the evidence from the benchmark/profile skill and hand the kernel work to the appropriate kernel, Nsight, or framework-specific optimization workflow.

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-benchmark-profile/scripts/bench_diffusion_denoise.py

@@ -12,7 +12,7 @@ Usage:
     # Tag the run for later compare_perf.py usage
     python3 python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-benchmark-profile/scripts/bench_diffusion_denoise.py --model flux --label tuned
 
-    # All 20 preset models
+    # All preset models
     python3 python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-benchmark-profile/scripts/bench_diffusion_denoise.py --all
 
     # Show preset order, model path, and nightly mapping
@@ -56,13 +56,18 @@ ASSET_DIR = ensure_dir(get_assets_dir(REPO_ROOT))
 NIGHTLY_CONFIG_PATH = (
     REPO_ROOT / "scripts" / "ci" / "utils" / "diffusion" / "comparison_configs.json"
 )
-GATED_MODELS = {"flux", "flux2"}
+GATED_MODELS = {
+    "flux",
+    "flux2",
+    "flux2-klein",
+    "flux2-klein-base",
+}
 DIFFUSERS_FALLBACK_SIGNALS = (
     "falling back to diffusers backend",
     "using diffusers backend",
     "loaded diffusers pipeline",
 )
-CATALOG_TABLE_WIDTH = 105
+CATALOG_TABLE_WIDTH = 140
 RESULTS_TABLE_WIDTH = 105
 NIGHTLY_PRESET_ORDER = (
     "flux",
@@ -80,7 +85,7 @@ NIGHTLY_PRESET_ORDER = (
 # ---------------------------------------------------------------------------
 # Model configs — kept in exact sync with benchmark-and-profile.md
 # Nightly-aligned presets mirror scripts/ci/utils/diffusion/comparison_configs.json
-# first, followed by skill-only extras.
+# first, followed by current-source extras and skill-only stress / coverage presets.
 # Each entry produces the same `sglang generate` command as shown in that doc.
 # ---------------------------------------------------------------------------
 MODELS = {
@@ -217,7 +222,137 @@ MODELS = {
             "--pin-cpu-memory",
         ],
     },
-    # 11. Skill-only extra preset
+    # Source-tracked extras from current registry / GPU test coverage.
+    "qwen-image": {
+        "path": "Qwen/Qwen-Image",
+        "prompt": "A futuristic cyberpunk city at night, neon lights reflecting on wet streets",
+        "extra_args": [
+            "--width=1024",
+            "--height=1024",
+        ],
+    },
+    # Requires: <repo>/inputs/diffusion_benchmark/figs/cat.png
+    "qwen-edit-2509": {
+        "path": "Qwen/Qwen-Image-Edit-2509",
+        "prompt": "Make the cat wear a red hat",
+        "image_path": str(ASSET_DIR / "cat.png"),
+        "extra_args": [
+            "--width=1024",
+            "--height=1024",
+        ],
+    },
+    "zimage-base": {
+        "path": "Tongyi-MAI/Z-Image",
+        "prompt": "A futuristic cyberpunk city at night, neon lights reflecting on wet streets",
+        "extra_args": [
+            "--width=1024",
+            "--height=1024",
+        ],
+    },
+    "flux2-klein": {
+        "path": "black-forest-labs/FLUX.2-klein-4B",
+        "prompt": "A futuristic cyberpunk city at night, neon lights reflecting on wet streets",
+        "extra_args": [
+            "--width=1024",
+            "--height=1024",
+            "--dit-layerwise-offload",
+            "false",
+        ],
+    },
+    "flux2-klein-base": {
+        "path": "black-forest-labs/FLUX.2-klein-base-4B",
+        "prompt": "A futuristic cyberpunk city at night, neon lights reflecting on wet streets",
+        "extra_args": [
+            "--width=1024",
+            "--height=1024",
+            "--dit-layerwise-offload",
+            "false",
+        ],
+    },
+    "cosmos3-nano-t2i": {
+        "path": "nvidia/Cosmos3-Nano",
+        "prompt": "A red cube on a white table, product photo.",
+        "env": {
+            "SGLANG_DISABLE_COSMOS3_GUARDRAILS": "1",
+        },
+        "extra_args": [
+            "--width=1024",
+            "--height=1024",
+            "--num-frames=1",
+            "--num-inference-steps=35",
+        ],
+    },
+    "cosmos3-nano-t2v": {
+        "path": "nvidia/Cosmos3-Nano",
+        "prompt": "A blue box slides across a clean warehouse floor.",
+        "env": {
+            "SGLANG_DISABLE_COSMOS3_GUARDRAILS": "1",
+        },
+        "extra_args": [
+            "--width=832",
+            "--height=480",
+            "--num-frames=9",
+            "--num-inference-steps=4",
+        ],
+    },
+    "ideogram4-fp8": {
+        "path": "ideogram-ai/ideogram-4-fp8",
+        "prompt": "A clean product poster for a new open-source inference engine",
+        "extra_args": [
+            "--width=1024",
+            "--height=1024",
+        ],
+    },
+    "ernie-image-turbo": {
+        "path": "baidu/ERNIE-Image-Turbo",
+        "prompt": "A futuristic cyberpunk city at night, neon lights reflecting on wet streets",
+        "extra_args": [
+            "--width=1024",
+            "--height=1024",
+        ],
+    },
+    "glm-image": {
+        "path": "zai-org/GLM-Image",
+        "prompt": "A futuristic cyberpunk city at night, neon lights reflecting on wet streets",
+        "extra_args": [
+            "--width=1024",
+            "--height=1024",
+        ],
+    },
+    "sana-1.5-1.6b": {
+        "path": "Efficient-Large-Model/SANA1.5_1.6B_1024px_diffusers",
+        "prompt": "A futuristic cyberpunk city at night, neon lights reflecting on wet streets",
+        "extra_args": [
+            "--width=1024",
+            "--height=1024",
+        ],
+    },
+    "fastwan22-ti2v-5b": {
+        "path": "FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers",
+        "prompt": "The cat starts walking slowly towards the camera.",
+        "image_path": str(ASSET_DIR / "cat.png"),
+        "extra_args": [
+            "--width=1280",
+            "--height=720",
+            "--num-frames=81",
+        ],
+    },
+    "ltx23-hq-two-stage": {
+        "path": "Lightricks/LTX-2.3",
+        "prompt": "A beautiful sunset over the ocean",
+        "env": {
+            "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
+            "SGLANG_LTX2_SNAPSHOT_RELEASE_EMPTY_CACHE": "true",
+        },
+        "extra_args": [
+            "--pipeline-class-name=LTX2TwoStageHQPipeline",
+            "--ltx2-two-stage-device-mode=snapshot",
+            "--width=1920",
+            "--height=1088",
+            "--num-frames=121",
+        ],
+    },
+    # Skill-only extra preset
     "ltx23-one-stage": {
         "path": "Lightricks/LTX-2.3",
         "prompt": "A beautiful sunset over the ocean",
@@ -233,7 +368,7 @@ MODELS = {
             "--num-gpus=2",
         ],
     },
-    # 12. Skill-only extra preset
+    # Skill-only extra preset
     "ltx23-two-stage": {
         "path": "Lightricks/LTX-2.3",
         "prompt": "A beautiful sunset over the ocean",
@@ -250,7 +385,7 @@ MODELS = {
             "--num-gpus=2",
         ],
     },
-    # 13. Skill-only extra preset
+    # Skill-only extra preset
     "ltx23-two-stage-cfg-parallel": {
         "path": "Lightricks/LTX-2.3",
         "prompt": "A beautiful sunset over the ocean",
@@ -268,7 +403,7 @@ MODELS = {
             "--cfg-parallel-size=2",
         ],
     },
-    # 14. Skill-only extra preset
+    # Skill-only extra preset
     "hunyuanvideo": {
         "path": "hunyuanvideo-community/HunyuanVideo",
         "prompt": "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window.",
@@ -281,7 +416,7 @@ MODELS = {
             "--num-inference-steps=30",
         ],
     },
-    # 15. Skill-only extra preset
+    # Skill-only extra preset
     # Requires: <repo>/inputs/diffusion_benchmark/figs/mova_single_person.jpg
     "mova-720p": {
         "path": "OpenMOSS-Team/MOVA-720p",
@@ -297,7 +432,7 @@ MODELS = {
             "--num-inference-steps=2",
         ],
     },
-    # 16. Skill-only extra preset
+    # Skill-only extra preset
     "helios": {
         "path": "BestWishYsh/Helios-Base",
         "prompt": "A curious raccoon",
@@ -315,14 +450,13 @@ MODELS = {
             "false",
         ],
     },
-    # 16. Skill-only extra preset
+    # Skill-only extra preset
     # Requires: <repo>/inputs/diffusion_benchmark/figs/cat.png
     "joyai-edit": {
         "path": "jdopensource/JoyAI-Image-Edit-Diffusers",
         "prompt": "Make the cat wear a red hat",
         "image_path": str(ASSET_DIR / "cat.png"),
         "extra_args": [
-            "--backend=sglang",
             "--width=1024",
             "--height=1024",
             "--num-inference-steps=40",
@@ -336,14 +470,13 @@ MODELS = {
             "--ulysses-degree=1",
         ],
     },
-    # 17. Skill-only extra preset
+    # Skill-only extra preset
     # Requires: <repo>/inputs/diffusion_benchmark/figs/cat.png
     "firered-edit-1.0": {
         "path": "FireRedTeam/FireRed-Image-Edit-1.0",
         "prompt": "Make the cat wear a red hat",
         "image_path": str(ASSET_DIR / "cat.png"),
         "extra_args": [
-            "--backend=sglang",
             "--width=1024",
             "--height=1024",
             "--num-inference-steps=40",
@@ -357,14 +490,13 @@ MODELS = {
             "--ulysses-degree=1",
         ],
     },
-    # 18. Skill-only extra preset
+    # Skill-only extra preset
     # Requires: <repo>/inputs/diffusion_benchmark/figs/cat.png
     "firered-edit-1.1": {
         "path": "FireRedTeam/FireRed-Image-Edit-1.1",
         "prompt": "Make the cat wear a red hat",
         "image_path": str(ASSET_DIR / "cat.png"),
         "extra_args": [
-            "--backend=sglang",
             "--width=1024",
             "--height=1024",
             "--num-inference-steps=40",
@@ -378,7 +510,7 @@ MODELS = {
             "--ulysses-degree=1",
         ],
     },
-    # 19. Skill-only extra preset
+    # Skill-only extra preset
     # Requires: <repo>/inputs/diffusion_benchmark/figs/cat.png
     "hunyuan3d-shape": {
         "path": "tencent/Hunyuan3D-2",
@@ -388,7 +520,6 @@ MODELS = {
             "paint_enable": False,
         },
         "extra_args": [
-            "--backend=sglang",
             "--num-inference-steps=50",
             "--guidance-scale=5.0",
             "--dit-layerwise-offload",
@@ -544,17 +675,17 @@ def print_model_catalog():
     """Print preset order, model path, and whether each preset maps to nightly."""
     print()
     print("=" * CATALOG_TABLE_WIDTH)
-    print("MODEL PRESETS — Nightly-aligned first, skill-only extras after")
+    print("MODEL PRESETS — Nightly-aligned, then current-source and skill-only extras")
     print("=" * CATALOG_TABLE_WIDTH)
-    print(f"{'Preset':<24} {'Nightly':<28} {'Model Path':<46} {'GPUs':>4}")
+    print(f"{'Preset':<32} {'Nightly':<30} {'Model Path':<66} {'GPUs':>4}")
     print("-" * CATALOG_TABLE_WIDTH)
     for model_key, cfg in MODELS.items():
         print(
-            f"{model_key:<24} {model_nightly_case_id(model_key):<28} {cfg['path']:<46} {required_gpus_for_model(model_key):>4}"
+            f"{model_key:<32} {model_nightly_case_id(model_key):<30} {cfg['path']:<66} {required_gpus_for_model(model_key):>4}"
         )
     print("-" * CATALOG_TABLE_WIDTH)
     print(
-        "Nightly column shows the comparison_configs.json case id; '-' means skill-only."
+        "Nightly column shows the comparison_configs.json case id; '-' means no nightly mapping."
     )
 
 
@@ -575,6 +706,7 @@ def build_sglang_cmd(
     cmd = [
         "sglang",
         "generate",
+        "--backend=sglang",
         f"--model-path={cfg['path']}",
         f"--prompt={cfg['prompt']}",
     ]
@@ -631,6 +763,9 @@ def run_benchmark_once(
 
     env = os.environ.copy()
     env.setdefault("FLASHINFER_DISABLE_VERSION_CHECK", "1")
+    cfg = MODELS[model_key]
+    for key, value in cfg.get("env", {}).items():
+        env.setdefault(key, str(value))
     if env.get("HF_TOKEN") and not env.get("HUGGINGFACE_HUB_TOKEN"):
         env["HUGGINGFACE_HUB_TOKEN"] = env["HF_TOKEN"]
 

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-modelopt-quant/SKILL.md

@@ -54,6 +54,7 @@ This repo now contains:
 - diffusion-side NVFP4 loading from ModelOpt exports
 - FLUX.2 packed-QKV detection that distinguishes packed NVFP4 checkpoints from standard diffusers exports
 - automatic protection against incompatible FP8 CPU offload while keeping layerwise DiT offload available
+- separate online diffusion quantization paths such as `--quantization fp8` / `mxfp4`; keep those out of this ModelOpt PTQ/export workflow unless the user explicitly asks for runtime quantization
 - FP8 transformer build:
   [`python/sglang/multimodal_gen/tools/build_modelopt_fp8_transformer.py`](../../../tools/build_modelopt_fp8_transformer.py)
 - NVFP4 mixed transformer build:
@@ -67,6 +68,11 @@ Validated documentation and CI coverage currently center on these ModelOpt diffu
 - NVFP4: FLUX.1-dev, FLUX.2-dev, Wan2.2
 
 Treat a new family, a new precision, or a new checkpoint layout as unsupported until it has a documented matrix row and a matching validation story.
+Current B200 CI also contains an Ideogram4 NVFP4 native load case
+(`ideogram4_nvfp4_t2i` via `Comfy-Org/Ideogram-4`). Treat that as source
+evidence for an existing NVFP4 path, but do not expand the ModelOpt support
+matrix to Ideogram4 unless `docs/diffusion/quantization.md` is updated with the
+exact checkpoint, loader path, quality check, and benchmark scope.
 Before writing CLI examples, re-read the active branch's `docs/diffusion/quantization.md`: FLUX.2 NVFP4 is an official `black-forest-labs/*` repo rather than a `lmsys/*` converted repo, and its preferred flag depends on the current documented loader flow. Use `--transformer-path` for a component override directory with `config.json`; use `--transformer-weights-path` when the repo or path should be probed as raw weights.
 
 B200 CI coverage can include loose BF16-vs-quantized quality checks. Inspect the active branch's `run_suite.py` before assuming they are part of the suite; mainline and feature branches may differ. Those checks are intended to catch blank, corrupted, or obviously divergent images, not exact image parity.
@@ -78,9 +84,9 @@ you are explicitly testing a historical branch.
 
 ## Related PR Watchlist
 
-As of 2026-05-04, these related SGLang PRs are relevant to ModelOpt diffusion
-support. Treat unmerged items as future support or migration work until the
-docs/CI matrix is updated.
+These related SGLang PRs are useful as ModelOpt diffusion support history.
+Re-check the PR state and the active source tree before treating any item as
+current behavior, and keep the docs/CI matrix as the support boundary.
 
 - #23155 added Qwen Image ModelOpt FP8 support.
 - #23199 adds HunyuanVideo ModelOpt FP8 support.

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-performance/SKILL.md

@@ -252,21 +252,27 @@ Use these as first commands to benchmark, not as universal winners.
 | Model family | First performance shape | Starting flags | Notes |
 |---|---|---|---|
 | FLUX.1 / FLUX.2 image | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup --dit-layerwise-offload false` | `black-forest-labs/FLUX.*` repos are gated; for FP8/NVFP4 use validated `--transformer-path` or `--transformer-weights-path` flows from the quant skill. |
+| FLUX.2 Klein / Klein Base | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup --dit-layerwise-offload false` | Current registry has `black-forest-labs/FLUX.2-klein-4B`, `FLUX.2-klein-9B`, and base variants. Klein is step-distilled; Klein Base is not. |
 | Qwen-Image / Qwen-Image-Edit | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup`; optionally native `SGLANG_CACHE_DIT_ENABLED=true` | Cache-DiT is lossy. For edit tasks, keep reference image, seed, and output size fixed. |
-| Z-Image-Turbo | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup` | Mainline has Z-Image tanh/gate norm fusions; PR #21912 tracks FP8 plus CUDA Graph work. |
+| Z-Image / Z-Image-Turbo | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup` | Keep base Z-Image separate from Turbo: base uses 50-step CFG defaults, Turbo uses 9-step zero-CFG defaults. Mainline has Z-Image tanh/gate norm fusions. |
 | Wan2.2 A14B T2V/I2V | 1280x720, 81 frames | Nightly: `--num-gpus 4 --enable-cfg-parallel --ulysses-degree 2 --text-encoder-cpu-offload --pin-cpu-memory` | For lowest latency, also benchmark pure Ulysses on the same GPUs. |
 | Wan2.2 TI2V 5B | 1280x720, 81 frames, 1 GPU | `--enable-torch-compile --warmup` | Keep the input image and motion prompt fixed when comparing sparse attention or Cache-DiT. |
-| LTX-2 / LTX-2.3 | 768x512, 121 frames, runtime-default steps/guidance, 2 GPUs | `--pipeline-class-name LTX2TwoStagePipeline --enable-torch-compile --warmup`; LTX-2 uses `--enable-cfg-parallel`, LTX-2.3 TI2V uses `--cfg-parallel-size 2` | Use the benchmark/profile skill presets for exact nightly alignment. PRs #22441, #24025, and #23736 track additional LTX2 perf/parallel work. |
+| Wan2.1 / FastWan / TurboWan variants | 480p or 720p video, family defaults | `--enable-torch-compile --warmup`; add `--ulysses-degree` / CFG parallel only after measuring | Current registry includes Wan2.1, FastWan2.1, FastWan2.2 TI2V, TurboWan2.1, TurboWan2.2 I2V, and Wan2.1-Fun InP. Use the compatibility matrix and benchmark presets before choosing topology. |
+| Cosmos3 Nano / Super | T2I: 1024x1024 with `--num-frames 1`; T2V/I2V: 480p/720p video | `SGLANG_DISABLE_COSMOS3_GUARDRAILS=1` for benchmark isolation; `--enable-torch-compile --warmup` | One checkpoint serves T2I/T2V/I2V. Mode is request-driven: `num_frames == 1` means T2I, `--image-path` means I2V. |
+| Ideogram 4 FP8/NVFP4 | 1024x1024, native preset defaults | `--enable-torch-compile --warmup` | Do not set `--num-inference-steps` or `--guidance-scale` directly unless you also update the Ideogram preset; sampling params derive them from `preset`. |
+| ERNIE-Image / GLM-Image / SANA / SD3 | 1024-class image, family defaults | `--enable-torch-compile --warmup`; disable offload only after checking VRAM | Treat these as current native image families. Start with benchmark/profile presets for ERNIE, GLM, and SANA; use registry/config defaults for SD3 unless you add a new preset. |
+| LTX-2 / LTX-2.3 | 768x512 or HQ 1920x1088, 121 frames | `--pipeline-class-name LTX2TwoStagePipeline --enable-torch-compile --warmup`; HQ uses `LTX2TwoStageHQPipeline` and `--ltx2-two-stage-device-mode snapshot` by default | Use benchmark/profile presets for nightly alignment, one-stage, high-resolution stress, and HQ. Device mode choices are `original`, `snapshot`, and `resident`; `resident` is fastest but uses more VRAM. |
 | HunyuanVideo | 848x480 or 720p class video | `--text-encoder-cpu-offload --pin-cpu-memory --enable-torch-compile --warmup` | Check VAE decode separately. GroupNorm+SiLU is default-eligible in mainline when wrapper guards pass; use `bench_group_norm_silu.py` when VAE residual blocks are hot. |
 | JoyAI-Image-Edit | 1024-class TI2I, 40 steps, guidance 4.0 | `--backend=sglang --num-gpus 2 --enable-cfg-parallel --ulysses-degree 1 --enable-torch-compile --warmup --dit-layerwise-offload false --dit-cpu-offload false` | Newly supported image-edit path. Keep the input image, prompt, seed, and output size fixed; 2-GPU CFG parallel is the validated H100 starting point. |
 | FireRed-Image-Edit 1.0 / 1.1 | 1024x1024 image edit, 40 steps, guidance 4.0 | `--backend=sglang --num-gpus 2 --enable-cfg-parallel --ulysses-degree 1 --enable-torch-compile --warmup --dit-layerwise-offload false --dit-cpu-offload false` | Uses the native `QwenImageEditPlusPipeline` path. 2-GPU CFG parallel is the validated H100 starting point; benchmark 1.0 and 1.1 separately because checkpoint differences can change denoise latency. |
 | Hunyuan3D-2 shape | Shape generation, 50 steps, guidance 5.0 | `--backend=sglang --enable-torch-compile --warmup --dit-layerwise-offload false --dit-cpu-offload false` | Focus on `Hunyuan3DShapeDenoisingStage`; keep mesh export/paint timings separate from denoise. |
-| MOVA / Helios | Use the benchmark/profile presets first | `--enable-torch-compile --warmup`; pin offload flags explicitly | PR #20530 tracks MOVA fused RMSNorm+RoPE; PR #24059 tracks Helios fused norm modulation. |
+| MOVA / Helios / LingBot World | Use the benchmark/profile presets or server test cases first | `--enable-torch-compile --warmup`; pin offload and topology flags explicitly | These video/realtime families have model-specific stages and condition handling. Keep prompt/image/action inputs fixed and prefer perf dumps over wall time alone. |
 
-## Open PR Watchlist
+## Historical PR Watchlist
 
-As of 2026-05-02, these performance PRs were open. Treat them as direction and
-prior art until merged:
+Treat these performance PRs as direction and prior art only. Re-check the PR
+state and the active source tree before relying on any path, flag, or claim
+about whether the work has merged:
 
 - Fusion/kernel: #24025 LTX2 QK norm, #24059 Helios norm modulation, #24117 Z-Image packed QKV, #19488 Wan elementwise cross-block fusion, #19249 Z-Image gate/norm fusion, #20429 Qwen-Image layernorm/modulation, #20530 MOVA RMSNorm+RoPE.
 - VAE/decode: #22531 LTX2 parallel VAE, #20927 batched tiled VAE decode.
@@ -279,6 +285,6 @@ prior art until merged:
 - **Offload tuning**: after the first request, the runtime logs peak GPU memory and which components could stay resident. Use this to decide which `--*-cpu-offload` flags to disable.
 - **Backend selection**: `--backend sglang` (default, auto-detected) enables native optimizations (fused kernels, SP, native Cache-DiT env knobs, etc.). `--backend diffusers` falls back to Diffusers pipelines and is the path that accepts `--cache-dit-config` plus diffusers attention backend names.
 - **Wan2.2-I2V sizing**: explicit `--width/--height` on `Wan2.2-I2V-A14B` control the target area while preserving the condition-image aspect ratio.
-- **Mainline diffusion fast paths**: before proposing a new kernel or overlap scheme, check `sglang-diffusion-benchmark-profile/existing-fast-paths.md`. It covers GroupNorm+SiLU, Z-Image residual-form modulation, fused diffusion `QK norm + RoPE`, packed QKV/NVFP4 expectations, and existing multi-GPU overlap families such as Ulysses / USP and turbo-layer async all-to-all.
+- **Mainline diffusion fast paths**: before proposing a new kernel or overlap scheme, check `sglang-diffusion-benchmark-profile/existing-fast-paths.md`. It covers GroupNorm+SiLU, Z-Image residual-form modulation, fused diffusion `QK norm + RoPE`, LTX2 split RoPE, varlen USP pack/scatter, packed QKV/NVFP4 expectations, and existing multi-GPU overlap families such as Ulysses / USP and turbo-layer async all-to-all.
 - **NVFP4 trace interpretation**: on FLUX.2 NVFP4 and Nunchaku-style checkpoints, packed QKV is expected. SGLang intentionally uses fused projection modules such as `to_qkv` / `to_added_qkv` instead of separate `to_q` / `to_k` / `to_v`, so a split-QKV trace usually means the quantized path did not engage rather than a brand new fusion opportunity.
-- **Hotspot workflow split**: use `sglang-diffusion-benchmark-profile` to prove and classify a slowdown with perf dumps plus `torch.profiler`; hand concrete kernel work to `sglang-diffusion-ako4all-kernel` or another specialized optimization skill instead of expanding the benchmark skill.
+- **Hotspot workflow split**: use `sglang-diffusion-benchmark-profile` to prove and classify a slowdown with perf dumps plus `torch.profiler`; hand concrete kernel work off with the perf/profile evidence attached instead of expanding the benchmark skill.