Search results

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

Status: Open.

#192 In deepseek-ai/FlashMLA;

· MogicianWu opened on Jun 26, 2026

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

Status: Open.

#190 In deepseek-ai/FlashMLA;

· echo-timeless opened on Jun 17, 2026

Make FlashMLA a libtorch and cpython stable extension

Status: Open.

#180 In deepseek-ai/FlashMLA;

· janeyx99 opened on Apr 27, 2026

Do you open to add SM80 support

Status: Open.

#179 In deepseek-ai/FlashMLA;

· haosdent opened on Apr 27, 2026

Use pyproject.toml

Status: Open.

#172 In deepseek-ai/FlashMLA;

· sakgoyal opened on Mar 27, 2026

dual gemm

Status: Open.

#169 In deepseek-ai/FlashMLA;

· WalrusWTQ opened on Mar 6, 2026

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?

Status: Open.

#168 In deepseek-ai/FlashMLA;

· zhangfengwei opened on Mar 6, 2026

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

Status: Open.

#166 In deepseek-ai/FlashMLA;

· pengwubj opened on Feb 25, 2026

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

Status: Open.

#165 In deepseek-ai/FlashMLA;

· lyppg opened on Feb 18, 2026

mla

Status: Open.

#164 In deepseek-ai/FlashMLA;

· TZWX-0 opened on Feb 12, 2026

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

Status: Open.

#161 In deepseek-ai/FlashMLA;

· red1239109-cmd opened on Feb 8, 2026

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

Status: Open.

#158 In deepseek-ai/FlashMLA;

· red1239109-cmd opened on Jan 25, 2026