Summary

Add ModelOpt FP8 support for Qwen Image diffusion transformers in the SGLang runtime and FP8 converter.

• make Qwen Image attention, MLP, and top-level projections quant-aware with full checkpoint prefixes
• add a Qwen Image / Qwen Image Edit BF16 fallback profile, including transformer_blocks.*.img_mlp.net.2 after image-quality ablation
• fix the FP8 converter so explicit BF16 fallback tensors are written before ModelOpt ignore-preservation skips the source tensor
• publish clean SGLang-native ModelOpt FP8 transformer overrides under the lmsys Hugging Face org
• document the validated Qwen Image and Qwen Image Edit ModelOpt FP8 checkpoint flow in docs/diffusion/quantization.md
• update the diffusion ModelOpt quant skill with the Qwen Image FP8 fallback and converter-ordering notes
• add Qwen Image and Qwen Image Edit ModelOpt FP8 cases to the B200 diffusion CI set

Published FP8 weights

• Qwen Image: https://huggingface.co/lmsys/qwen-image-modelopt-fp8-sglang-transformer
• Qwen Image Edit: https://huggingface.co/lmsys/qwen-image-edit-modelopt-fp8-sglang-transformer

Both repos are intentionally clean transformer override repos: README.md, config.json, and .safetensors shards only.

Validation

Validated on H100 rank0 (CUDA_VISIBLE_DEVICES=0) for the generated artifacts and benchmarks.

• SGLang main base used during validation: 6ecd6f84d
• ModelOpt main used for PTQ export: 26ae8da51
• python -m compileall -q python/sglang/multimodal_gen/runtime/models/dits/qwen_image.py python/sglang/multimodal_gen/tools/build_modelopt_fp8_transformer.py -> passed
• isort, black, and ruff check --select=F401,F821 --fix on changed Python files -> passed, no further changes
• 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 after adding B200 cases
• 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

Qwen Image 1024x1024, 50 steps

Prompt: A futuristic cyberpunk city at night, neon lights reflecting on wet streets

• BF16 image: visually normal
• Old FP8 default: severe dark/blurred quality regression
• Fixed FP8 (img_mlp.net.2 BF16 fallback): visually normal
• converter stats: 660 quantized weights, 1320 scale tensors, 186 BF16 fallback weights, 3 output shards
• benchmark with native sglang generate --backend=sglang --warmup:
  • E2E: 13589.20 ms -> 12159.39 ms, 10.5% faster
  • Denoising: 12928.76 ms -> 11437.40 ms, 11.5% faster

Qwen Image Edit 512x512, 8 steps

Prompt: A clean product photo of a small ceramic teapot on a wooden table, soft daylight, sharp details.

• BF16 and fixed FP8 images are visually aligned for the smoke edit workload
• converter stats: 660 quantized weights, 1320 scale tensors, 186 BF16 fallback weights, 3 output shards
• benchmark with native sglang generate --backend=sglang --warmup:
  • E2E: 6791.97 ms -> 6085.03 ms, 10.4% faster
  • Denoising: 5204.32 ms -> 4524.01 ms, 13.1% faster

B200 CI

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

• qwen_image_modelopt_fp8_t2i
• qwen_image_edit_modelopt_fp8_ti2i

Notes

• The ModelOpt Qwen model registration used for PTQ export was patched only in the H100 validation checkout and is not included here.
• The fixed profiler trace was captured for Qwen Image 1024x1024, 8 steps, --profile --num-profiled-timesteps=2.