Add HunyuanVideo ModelOpt FP8 diffusion support

[BBuf]

Summary

Add HunyuanVideo ModelOpt FP8 diffusion support and publish the SGLang-native transformer override under the lmsys Hugging Face org.

• add HunyuanVideo ModelOpt FP8 runtime support
• document the HunyuanVideo ModelOpt FP8 checkpoint flow in the diffusion quantization docs
• update the diffusion ModelOpt quant skill with the HunyuanVideo FP8 path
• add the HunyuanVideo ModelOpt FP8 case to the B200 diffusion CI set

Published FP8 weights

• HunyuanVideo: https://huggingface.co/lmsys/hunyuanvideo-modelopt-fp8-sglang-transformer

The repo is intentionally clean: README.md, config.json, and .safetensors shards only.

H100 Validation

Updated H100 validation used the sglang-diffusion-benchmark-profile HunyuanVideo command shape. This supersedes the earlier short-video validation.

Run setup:

• Host/GPU: H100 rank0, CUDA_VISIBLE_DEVICES=0
• Backend: --backend=sglang; logs show Using pipeline from model_index.json: HunyuanVideoPipeline, no diffusers fallback markers observed
• Model: 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.
• Skill preset args preserved: --text-encoder-cpu-offload --pin-cpu-memory --num-frames=65 --width=848 --height=480 --num-inference-steps=30 --save-output --warmup --enable-torch-compile --seed=42
• 5s adjustment: added --fps=13, so the output is exactly 65 frames / 13 fps = 5.000s
• FP8 delta: same command plus --transformer-path lmsys/hunyuanvideo-modelopt-fp8-sglang-transformer
• Profiler delta: same generation settings, replacing save/perf output with --profile --num-profiled-timesteps=5 --no-save-output and setting SGLANG_DIFFUSION_TORCH_PROFILER_DIR

Benchmark, warmup excluded:

Metric	BF16	FP8	Delta	Speedup
E2E latency	59.546 s	54.748 s	-4.798 s (-8.1%)	1.09x
Denoising stage	42.542 s	37.980 s	-4.562 s (-10.7%)	1.12x
Avg denoise step	1.4180 s	1.2659 s	-0.1521 s	1.12x
Decoding stage	16.692 s	16.458 s	-0.233 s (-1.4%)	1.01x
Text encoding	0.308 s	0.306 s	-0.002 s (-0.7%)	1.01x

Profiler kernel share, 5 profiled denoise timesteps. Profiler timings include profiling overhead and are not used as benchmark latency numbers.

Precision	Total CUDA op time	Top CUDA/kernel shares
BF16	17.055 s	cudaMemcpyAsync 41.54%; FlashAttention 31.99%; BF16 GEMM nvjet_tst_192x208_64x4_2x1_v_bz_coopB_bias_TNT 9.77%; BF16 GEMM nvjet_tst_192x208_64x4_1x2_h_bz_coopB_bias_TNT 8.16%; BF16 GEMM nvjet_tst_256x152_64x4_1x2_h_bz_coopA_bias_TNT 2.11%
FP8	15.324 s	cudaMemcpyAsync 40.62%; FlashAttention 36.80%; FP8 Cutlass GEMM 12.83%; triton_poi_fused_cat_gelu_view_0 1.93%; _static_quant_fp8 1.37%

Local Validation After B200 CI Update

• git diff --check -> passed
• python3 -m py_compile python/sglang/multimodal_gen/test/server/testcase_configs.py python/sglang/multimodal_gen/test/server/gpu_cases.py -> passed
• python3 -m black --check python/sglang/multimodal_gen/test/server/testcase_configs.py python/sglang/multimodal_gen/test/server/gpu_cases.py -> passed
• python3 -m ruff check --select=F401,F821 python/sglang/multimodal_gen/test/server/testcase_configs.py python/sglang/multimodal_gen/test/server/gpu_cases.py -> passed

B200 CI

Added to ONE_GPU_MODELOPT_CASES for multimodal-gen-test-1-b200:

• hunyuanvideo_modelopt_fp8_t2v

One caveat from the FP8 log: the CLI keeps the same offload flags as the skill preset, but the ModelOpt FP8 runtime currently forces dit_cpu_offload off while preserving layerwise offload behavior for restored FP8 tensor strides.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[BBuf]

/tag-and-rerun-ci

[BBuf]

/tag-and-rerun-ci

[mickqian]

duplicated contents?

[BBuf]

Updated this PR to use the new clean lmsys ModelOpt diffusion repo.

• HunyuanVideo: https://huggingface.co/lmsys/hunyuanvideo-modelopt-fp8-sglang-transformer
• Added hunyuanvideo_modelopt_fp8_t2v to ONE_GPU_MODELOPT_CASES for multimodal-gen-test-1-b200.
• The shared lmsys collection is https://huggingface.co/collections/lmsys/diffusion-modelopt-69f06a1740c02269e36bf285

[BBuf]

/tag-and-rerun-ci

[BBuf]

/tag-and-rerun-ci

[BBuf]

/tag-and-rerun-ci

[BBuf]

/tag-and-rerun-ci

[BBuf]

https://github.com/sgl-project/sglang/actions/runs/25367205803/job/74399478580?pr=23199