AMD MI300X Deployment Runs DeepSeek V4-Flash at 168 tok/s Single-Stream, Ships FP8 Correctness Fixes

An open-source deployment kit fits DeepSeek V4-Flash's 304B parameters unquantized on a single AMD MI300X, delivering 168.6 tok/s single-stream decode and 7.9-8.5K tok/s prefill across 64 concurrent streams. The repo publishes correctness patches for AMD's FNUZ FP8 format, gfx942 kernel tuning tables, a 20GB GPU KV cache + 96GB CPU offload split, and DSpark-7 speculative decoding.