DeepSeek for Compute and Chip Adaptation — the five layers you actually need

Running DeepSeek on non-NVIDIA silicon. The active work is not on CXMT or SMIC — it is Ascend, Hygon, AMD NPU and Intel.

Models

Toolchain

Reference impls

Hardware

1. Models

Model	Downloads
FastFlowLM/Deepseek-R1-Distill-Llama-8B-NPU2	811
FastFlowLM/DeepSeek-R1-0528-Qwen3-8B-NPU2	736
OpenFlowLM/DeepSeek-R1-0528-Qwen3-8B-NPU2	394
amd/DeepSeek-R1-Distill-Qwen-7B-onnx-ryzenai-npu	212
amd/DeepSeek-R1-Distill-Llama-8B-onnx-ryzenai-npu	183
Atomic-Germ/DeepSeek-R1-0528-Qwen3-8B-NPU2	175
Atomic-Germ/Deepseek-R1-Distill-Llama-8B-NPU2	137
amd/DeepSeek-R1-Distill-Qwen-1.5B-onnx-ryzenai-npu	101
OpenFlowLM/Deepseek-R1-Distill-Llama-8B-NPU2	97
amd/DeepSeek-R1-Distill-Qwen-7B_rai_1.7.1_npu_4K	31
llmware/DeepSeek-R1-Distill-Qwen-7B-ov-int4-npu	24
helenai/DeepSeek-R1-Distill-Qwen-7B-ov-int4-npu	23

Project	★	What it does

Project	★	What it does

What runs on what. These are working examples, not vendor claims.

Project	★	Hardware target
ipex-llm	8,844	Accelerate local LLM inference and finetuning (LLaMA, Mistral, ChatGLM, Qwen, DeepSeek, Mi
gpustack	5,978	A GPU cluster manager for high-performance AI model serving (vLLM, SGLang) and on-demand S
cube-studio	5,087	cube studio开源云原生一站式机器学习/深度学习/大模型AI平台，mlops算法链路全流程，算力租赁平台，notebook在线开发，拖拉拽任务流pipeline编排，多机多
cube-studio	2,592	cubestudio开源云原生一站式机器学习/深度学习/大模型AI平台/MaaS/mlops/人工智能平台/训推平台，算法全链路流程，多租户，算力租赁平台，token中转，拖拉拽任
FastFlowLM	2,019	Run LLMs on AMD Ryzen™ AI NPUs in minutes; purpose-built and deeply optimized for the AMD
unified-cache-management	352	Persist and reuse KV Cache to speedup your LLM.
dcu-in-action	119	国产加速卡-海光DCU实战（大模型训练、微调、推理等）
deepseek-v4.1-flash-ascend910B	58	DeepSeek-V4.1-Flash W4A8 optimized vLLM deployment for Ascend 910B (A2) / 910C (A3): patch
PocketLLM	55	Inference engine for large language models on consumer GPUs, with deep optimization for ol
SLAI-T-Rex	42	SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD

Ascend vLLM adaptation has a version matrix — mismatched vllm-ascend and CANN versions simply will not start.
Domestic quantization formats are not portable. W4A8 and W8A8 differ per vendor; weights cannot be moved directly.
Hygon DCU runs ROCm, whose ecosystem is thinner than CUDA — many kernels must be written by hand.
AMD NPU (Ryzen AI) only runs small models and needs ONNX conversion, not native support.
Collective communication differs: NCCL must be replaced on non-NVIDIA platforms.