DeepSeek-V4-Flash-0731-FP8-DSpark

FP8 re-packaging of deepseek-ai/DeepSeek-V4-Flash-0731 for Hopper (H200/H100).

Routed MoE experts converted MXFP4/FP4 -> block FP8 via lossless

cast_e2m1fn_to_e4m3fn (tmp/scripts/convert_hf_mxfp4_experts_to_fp8.py).

expert_dtype removed from config.json; other fields (including dspark_* if

present) kept.

On Hopper set SGLANG_DSV4_FP4_EXPERTS=0 when serving.