hygon

国产加速卡-海光DCU实战（大模型训练、微调、推理 等）

K100AI 模型部署实践 —— 海光 K100-AI 上大模型的成功配置线拉起包（当前：Qwen3.8-27B 六条精选线；持续更新）

Reproducible single-GPU Qwen3.6-35B-A3B W8A8 inference tuning for Hygon K100AI (gfx928) on vLLM 0.18.1.

MiniMax H3 INT8 ConvRot optimization and dual-K100AI QKV/Attention inference for Hygon K100AI (gfx928)

dsh plugin for scnet.cn

MiniMax-H3 (135B BF16) cross-hardware port to 8x Hygon K100 AI (gfx928): six-card production system, FSDP-lite sharding, RCCL workarounds, gray-mush root cause fix, ops playbooks and verification methodology

Scoped DeepSeek V4 paged-MQA correctness repair for Hygon BW100 (gfx936), TP8 integration, and historical million-token validation.

Meta-skill for integrating HPC, AI4S and scientific applications with SCNet compute resources, Slurm scheduling and cluster packaging, including Hygon DCU compatibility.

Terminal monitor for Hygon DCU (hy-smi) with nvtop-style scrolling curves

A miniature of vLLM v1 with a vLLM-style Platform/op-family hardware abstraction layer. One engine, four backends: NVIDIA CUDA, Hygon DCU (DTK/HIP), Ascend NPU, Moore Threads MUSA. Real continuous batching, paged KV cache, TP, CUDA graph — runs Qwen3/Olmoe end-to-end

Agent skill for working on SCNet clusters