This repository provides configuration and patches to run DeepSeek-V4-Flash-0731 on a single AMD MI300X in production. The checkpoint runs as shipped without additional weight quantization or offload. Results from the pinned stack (vLLM ROCm nightly 0.26.1rc1.dev229+g124154a88.rocm723, AITER 0.1.19) show single-stream decode at 168.6 tok/s and prefill at approximately 7.9–8.5K tok/s. The setup handles 8 concurrent streams at 542 tok/s aggregate and bursts of 64 streams at 830 tok/s aggregate with no OOM or engine errors. Context of 256K is validated, and the architecture supports 1M. Weights occupy 156.67 Gi B in HBM with no additional quantization or weight offload.

The MI300X's 192 GB HBM3 and 5.3 TB/s bandwidth enable simple single-GPU deployment. The entire 304B-parameter model fits in HBM without PCIe streaming or layer offload. One card handles 2–8 concurrent streams and bursts of up to 64 streams.

The repository includes correctness overlays for the pinned ROCm nightly, a validated serving configuration with probabilistic DSpark drafting, AITER GEMM tuning tables for gfx942 shapes, and a hybrid KV strategy with 20 GB GPU cache plus 96 Gi B CPU offload. Deployment requires one MI300X, ~235 Gi B RAM, and ~500 GB disk.