deepseek-ai / FlashMLA

Optimized GPU and NPU attention kernels from DeepSeek for sparse prefill and FP8/FP4 KV-cache decoding in DeepSeek-V4.1 inference.