AIHacker Newsabout 3 hours ago
DeepSeek V4 Flash on a Single AMD MI300X
1 min read
DeepSeek-V4-Flash-0731 runs on one AMD MI300X via a pinned vLLM ROCm stack with patches for FP8 format, MoE routing, and kernel tuning. The 304B-parameter checkpoint achieves 7.9-8.5K tok/s uncached prefill. The MI300X's 192 GB HBM3 enables single-GPU deployment without quantization or offload.
Level
Hype check
Tap to vote and see what everyone thinks.
In this storyAMD
