Repository Issue Activity (beta)

#169 dual gemm

#172 Use `pyproject.toml`

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

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

#179 Do you open to add SM80 support

#66 Ampere architecture FlashMLA bring-up

#180 Make FlashMLA a libtorch and cpython stable extension

#149 [Question] DSA VS MLA Prefill Benchmark On H100

#171 CUTLASS Internal Error during run on SM100 with long sequences (seq_len=1M)

#168 flash_mla.flash_mla_with_kvcache have plans to support FP8 for the q data in the future or FP8 insufficient to support model accuracy?

#166 [Question]Questions regarding INTERLEAVE vs. SW128 Layouts for SM90 Sparse Attention Decode

#125 Can MHA be used in the DSA prefill stage?

#165 [Question] why not using FP8 rope in sparse FP8 decoding?

#164 mla

#161 [CUDA/CUTLASS] Improvements for varlen option derivation, input validation, can_implement errors, and workspace handling

#159 RuntimeError: CUBLAS_STATUS_INVALID_VALUE in ref_sparse_attn_decode on H800 (Hopper)

#155 [Question] What is MODEL1?

#158 [Bug/Correctness] Hardcoded device_id=0 + missing CUDAGuard can break multi-GPU correctness (wrong hw_info / stream mismatch)

#153 Compilation Error: exceeds maximum register limit.

#154 Missing the included header file: #include <span>

#148 FlashMLA for Blackwell architecture (B200)

#126 does flash_mla_with_kvcache work only in paged mode?

#119 Built FlashMLA for Windows and Nvidia sm_120 (workstation/50s) cards

#124 support for rtx 6000 sm120

#121 Build error on cuda13 arm64