DeepSeek 开源 DeepInfer：仅支持 SM100 与昇腾 950

Robert Youssef @rryssf原推发布于 2026.10.21

在 X 上打开原推

DeepSeek's new engine doesn't support H100s. On October 20, DeepSeek open-sourced DeepInfer, the inference framework behind its DeepSeek-V4.1-Flash service. It supports two kinds of hardware: NVIDIA SM100 (Blackwell) GPUs and Huawei Ascend 950 NPUs. Hopper isn't on the list. It runs one model, too: DeepSeek-V4.1-Flash. The default checkpoint, DeepSeek-V4.1-Flash-1017, is MIT-licensed. The announcement pulled 2.3M views in about seven hours as "our production stack." What shipped is an offline engine. The README lists what a full service needs and leaves out: a KV cache system, the split of vision encoding, prefill and decode onto separate machines, and load balancing across replicas. DeepSeek says its KV cache and load balancing code is "tightly coupled to our internal infrastructure." The kernels are real: FlashMLA, DeepGEMM, DeepEP, plus MoE expert parallelism and DSpark speculative decoding, where a confidence head picks which draft tokens get verified. Two pieces are useful today. Structured outputs let you constrain generation to a regex, a JSON schema or an EBNF grammar through so_args in SamplingParams. Deterministic sampling gives identical outputs for identical inputs: set DEEPINFER_BATCH_INVARIANT=1 and DEEPINFER_ENABLE_MTP_VARLEN=0, pass a fixed seq_seed. NVIDIA only. Read the benchmark charts carefully. They used 4 SM100 GPUs for prefill and 64 for decode, and the decode chart assumes a 75% draft acceptance rate. The raw script counts every draft token as accepted. DeepSeek's own smallest efficient deployment is 64 GPUs of prefill plus 64 of decode. Before you build, run nvidia-smi --query-gpu=compute_cap --format=csv. SM100 prints 10.0. An H100 prints 9.0.