Add Qwen3.5 FP8 and NVFP4

[mmangkad]

No description provided.

[mmangkad]

cc @JustinTong0323

[mmangkad]

@JustinTong0323 I can also add now nvfp4 here if that sounds good

[kedarpotdar-nv]

Hi @mmangkad I am contributing FP4 and FP8 configs to InferenceX here - SemiAnalysisAI/InferenceX#820

Have added a couple recommendations for best perf based on our experiments. Thanks!

[kedarpotdar-nv]

cc: @faradawn

[mmangkad]

Hi @kedarpotdar-nv, thanks for your recommendations. I added --enable-flashinfer-allreduce-fusion to all configs as it should be good for SM90+ and tp > 2. I also added --kv-cache-dtype fp8_e4m3 for FP4 configs now, but this should be automatically resolved already in my PR sgl-project/sglang#19291.

Also there might still be changes here after sgl-project/sglang#19391 when spec v2 is already enabled.

[hlu1]

Why not just use the default value? flashinfer_cutlass is the current best due to an issue in flashinfer, but when the flashinfer issue is fixed, this will not be the best option.