Morrowmake/dsv41-flash-cmp170hx-recipe
DeepSeek-V4.1-Flash on 4× NVIDIA CMP 170HX with vLLM: up to 319 tok/s for one user, up to 765 tok/s at eight, 262K context, one-command container install
GitHub repository with 16 stars and 4 forks.
Language: Python
Topics: ampere, cmp-170hx, deepseek, exl3, local-llm, moe, speculative-decoding, vllm