FlashMLA · Issues· 128 open

[Portability][MSVC] GNU always_inline attributes on SM90 device lambdas fail to compile

[Portability][MSVC] Replace compound-literal-style dependent-type initialization in dense MLA decode

[Portability][MSVC] Replace __int128_t in shared-memory load/store paths 3

Sparse MLA decode is not batch-invariant: a row's output depends on its batchmates (1-2 ULP), flipping argmax at near-tie tokens

PyBind11 interface: How should function signature and CUDA error handling be handled for registered kernels?

10 rounds of hand-written fused attention that never beat a 3-kernel cuBLAS stack — trajectory, where each round died, and a correctness trap

DeepSeek-Coder-V2-Lite-Instruct 16B running locally on an RTX 2050 laptop

flashmla dense decoding test lack fp8 kv cache case

dual gemm

Use `pyproject.toml`

Sparse MLA decode (V3.2 / FP8 KV): throughput cost of the B200 accuracy fix (5aa668c) scales with topk

flash mla kernal是否支持同一batch下不同query的动态token数？

Do you open to add SM80 support

Ampere architecture FlashMLA bring-up

Make FlashMLA a libtorch and cpython stable extension