All issues

Search results

ffpa-attn CUDA Backend 特性支持现状技术报告

Status: Open.

#344 In xlite-dev/ffpa-attn;

· DefTruth opened on Aug 28, 2026

[CUDA][RFC] Supprt tensor_layout BNHD for Split-D and Split-D M4N2 paths (FP16/FP8/FP4)

Status: Open.

#336 In xlite-dev/ffpa-attn;

· DefTruth opened on Aug 26, 2026

[FP8][RFC] Optimize the performance of hybrid attention (FP16 + FP8) path

Status: Open.

#331 In xlite-dev/ffpa-attn;

· DefTruth opened on Aug 24, 2026

[FP4][RFC] Support attn_mask (align with SDPA) for CUDABackend FP4

Status: Open.

#327 In xlite-dev/ffpa-attn;

· DefTruth opened on Aug 24, 2026

[FP8][RFC] Support attn_mask (align with SDPA) for CUDABackend FP8

Status: Open.

#326 In xlite-dev/ffpa-attn;

· DefTruth opened on Aug 24, 2026

[FP4][RFC] Support PV FP8 with MMA ACC F16 and inst buffer

Status: Open.

#325 In xlite-dev/ffpa-attn;

· DefTruth opened on Aug 23, 2026

varlen backward hangs on SM90/H100 at head_dim=512 with torch 2.13/triton 3.7.1 (returns correctly on torch 2.11/triton 3.6)

Status: Open.

#298 In xlite-dev/ffpa-attn;

· dengoswei opened on Aug 4, 2026

SM100 (and SM<90) support for varlen backward (for training)

Status: Open.

#276 In xlite-dev/ffpa-attn;

· nathan-az opened on Jul 14, 2026

Making CUTLASS/CuTeDSL optional for Windows

Status: Open.

#271 In xlite-dev/ffpa-attn;

· zzlol63 opened on Jun 29, 2026

[RFC] ffpa-attn 0.3.0 roadmap

Status: Open.

#269 In xlite-dev/ffpa-attn;

· DefTruth opened on Jun 19, 2026