DeepInfer
Production inference framework. Offline path for DeepSeek-V4.1-Flash.
- SM100 · Ascend 950
- Apache 2.0
DeepSeek DeepInfer for production inference
Offline inference engine for DeepSeek-V4.1-Flash on SM100 and Ascend 950. Separate from deepseek-recipe protocols and the DeepSeek Harness agent runtime.
Three layers: DeepInfer, deepseek-recipe, DeepSeek Harness
DeepInfer, deepseek-recipe, and DeepSeek Harness solve different problems. Prefill and Decode are phases of the same engine.
deepseek-recipe
Request conversion, prompts, tokenization. Does not run inference.
pip install deepseek-recipe- MIT
DeepSeek Harness
Agent runtime — covered on dshbase, not this site’s focus.
DeepInfer vs deepseek-recipe vs DSH
Official Prefill / Decode charts
From DeepInfer doc/images/. Production cluster figures (155T tokens/day, $38/GPU-hour, …) are not OSS install outcomes — see FAQ attribution and request pipeline.
Who should skip — and who should install
See also the hardware gate.
Skip if you…
Install if you…
- Run offline DeepSeek-V4.1-Flash
- Have Linux, Python 3.12, GCC ≥ 13, CUDA 13 or CANN 9.2
- Can follow the official README build