Pinned Loading

1. sdpa-attention-benchmark sdpa-attention-benchmark Public

   Benchmark PyTorch SDPA backends (math vs flash) on RTX 4060 Ti with Nsight Systems profiling

   Python 2

2. flashattn-cuda-metal flashattn-cuda-metal Public

   FlashAttention CUDA kernel implementation and Metal port (RTX 4060 Ti, Apple M4 Pro)

   Cuda 3

3. fused-qkv-int8-attention fused-qkv-int8-attention Public

   Fused INT8 KV-cache dequantization + FlashAttention-style tiled decode attention CUDA kernel on A100

   Python