vLLM shrinks a 970 MB output head to 300 MB and gains 12% on Strix Halo

Xiao Yu's merged patch compresses Qwen3.6's 16-bit lm_head at load, cuts median decode time by 2.82 ms, and publishes the perplexity cost beside the gain.

By · Published

A large white AMD logo sits over a dark blue circuit board with teal chips and metallic traces.

Xiao Yu's merged patch compresses Qwen3.6's 16-bit lm_head at load, cuts median decode time by 2.82 ms, and publishes the perplexity cost beside the gain.