Summary

• add Wan2.2 ModelOpt NVFP4 support on top of #22672
• keep a global --transformer-weights-path override scoped to the primary transformer so transformer_2 stays on the base BF16 checkpoint unless explicitly overridden
• make scheduler loading tolerate newer config fields such as shift_terminal when the resolved SGLang scheduler class does not accept them
• honor SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND in diffusion ModelOpt FP4 GEMM selection so Blackwell bring-up can force the validated FlashInfer path
• document the validated Wan2.2 NVFP4 launch recipe

Validation

• official ModelOpt FP4 export for Wan-AI/Wan2.2-T2V-A14B-Diffusers, with only the primary transformer quantized and transformer_2 kept BF16
• B200 no-compile generation with base BF16 model + --transformer-weights-path override succeeded after forcing SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND=cudnn
• fixed-config B200 comparison on 832x480 / 17 frames / 2 steps:
  • main BF16 no-compile: E2E 55.89s, DenoisingStage 53.51s
  • this branch NVFP4 no-compile: E2E 25.72s, DenoisingStage 23.46s
  • delta: E2E -54.0%, DenoisingStage -56.2%
• warmup compile check on the same PR branch/config:
  • warmup eager: E2E 17.87s, DenoisingStage 16.20s
  • warmup compile: E2E 22.64s, DenoisingStage 20.96s
  • compile was slower on this setup (+26.7% E2E, +29.4% denoise)

Notes

• this PR is intentionally stacked on #22672
• the validated Blackwell bring-up path currently uses FlashInfer FP4 GEMM via SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND=cudnn
• local B200 artifacts (videos, perf dumps, traces, summaries) were collected separately outside the repo