Popular repositories Loading

1. mi210-llm-stack mi210-llm-stack Public

   AMD MI210 (gfx90a/CDNA2) LLM inference optimization — TurboQuant, KIVI, per-layer KV types, FlashAttention, MoE expert caching

   Shell 7

2. aiter-cdna2 aiter-cdna2 Public

   Run AMD AITER's hand-written ASM kernels on CDNA2 / gfx90a (MI210, MI250). 242 of 1,422 kernels translated, with the tests and benchmarks to prove which actually run.

   Python 7

3. cve-af-alg-block cve-af-alg-block Public

   Shell 3

4. cve-ptrace-mm-null-block cve-ptrace-mm-null-block Public

   SystemTap mitigation for __ptrace_may_access mm==NULL privilege escalation

   Python 2

5. krea2-intel-arc-b580 krea2-intel-arc-b580 Public

   Run Krea 2 (12B DiT) on an Intel Arc B580 in Docker — working PyTorch-XPU/ComfyUI setup for Battlemage on kernel 7.x, plus the fix for the Level-Zero abort, benchmarks, and troubleshooting

   Dockerfile 2

6. vllm-int8-moe-rocm vllm-int8-moe-rocm Public

   vLLM refuses INT8 MoE on every AMD GPU because of a CUDA-only check. One-line fix + benchmark harness. On MI210: 3.20s TTFT vs 5.07s for AWQ-Int4.

   Python 2