Skip to content

Actions: flash-algo/flash-sparse-attention

Actions

Auto assign reviewers and assignees

Actions

Loading...
Loading

Show workflow options

Create status badge

Loading
auto-assign.yml
41 workflow runs
41 workflow runs

Filter by Event

Filter by Status

Filter by Branch

Filter by Actor

[BUG FIX] Unify masking utilities and improve performance Auto assign reviewers and assignees #41: Pull request #209 opened by LoserCheems
10s
10s
Chore/sync after move Auto assign reviewers and assignees #40: Pull request #208 opened by LoserCheems
12s
12s
Fix documentation and references for Flash Sparse Attention Auto assign reviewers and assignees #39: Pull request #207 opened by LoserCheems
12s
12s
[FEATURE SUPPORT] Triton special compact dynamic-mask attention: 1.6× faster fwd+bwd, numerically equivalent Auto assign reviewers and assignees #38: Pull request #206 opened by LoserCheems
10s
10s
Refactor attention block smoothing for consistency Auto assign reviewers and assignees #37: Pull request #205 opened by LoserCheems
23s
23s
Add selectable masking strategies for attention Auto assign reviewers and assignees #36: Pull request #204 opened by LoserCheems
1h 51m 25s
1h 51m 25s
[FEATURE SUPPORT] Move scaling out of streaming loops, bias-initialized acc_s, and fix dQ double-scaling Auto assign reviewers and assignees #35: Pull request #203 opened by LoserCheems
12s
12s
Add block-wise smoothing to attention mask Auto assign reviewers and assignees #34: Pull request #201 opened by LoserCheems
11s
11s
Optimize triton version: GQA, mask/bias broadcasting, skip inactive tiles, and stability fixes Auto assign reviewers and assignees #33: Pull request #200 opened by LoserCheems
21s
21s
Fix attention bias calculation and dbias handling Auto assign reviewers and assignees #32: Pull request #199 opened by LoserCheems
17s
17s
Update documentation to use mask utility in examples Auto assign reviewers and assignees #31: Pull request #198 opened by LoserCheems
20s
20s
[FEATURE SUPPORT] Centralize dynamic mask creation for FDMA Auto assign reviewers and assignees #30: Pull request #197 opened by LoserCheems
19s
19s
[FEATURE SUPPORT] Robust dBias accumulation for seqlen_q_bias == 1 Auto assign reviewers and assignees #29: Pull request #194 opened by LoserCheems
17s
17s
Enhance bias gradient accumulation in backward pass Auto assign reviewers and assignees #28: Pull request #193 opened by LoserCheems
11s
11s
Fix attention_mask and attention_bias shape descriptions and remove redundant checks Auto assign reviewers and assignees #27: Pull request #192 opened by LoserCheems
10s
10s
Refactor bias initialization and enhance bias computation in FlashDMAttnFunc Auto assign reviewers and assignees #26: Pull request #191 opened by LoserCheems
12s
12s
[FEATURE SUPPORT] Broadcastable 4D mask/bias, 128‑rounded key length, stride‑0 broadcasting, and dbias reductions Auto assign reviewers and assignees #25: Pull request #190 opened by LoserCheems
16s
16s
[FEATURE SUPPORT] Variable-Length Attention with Padding-Free Execution Auto assign reviewers and assignees #24: Pull request #188 opened by LoserCheems
9s
9s
[FEATURE] Support total_k-head broadcastable mask/bias in varlen batch inference Auto assign reviewers and assignees #23: Pull request #187 opened by Copilot AI
Action required
Action required
Add issue/PR templates Auto assign reviewers and assignees #22: Pull request #186 opened by LoserCheems
11s
11s
Implement variable-length attention with mask and bias support Auto assign reviewers and assignees #21: Pull request #185 opened by LoserCheems
11s
11s
[BUG FIX] Optimize top-k mask construction: prevent unsafe gradient flow and eliminate unnecessary memory allocations Auto assign reviewers and assignees #20: Pull request #184 opened by LoserCheems
51m 53s
51m 53s
[BUG FIX] Fix mask/bias memory access and vectorization issues in kernels Auto assign reviewers and assignees #19: Pull request #182 opened by LoserCheems
11s
11s
Fix INF issue in bf16 backward pass with safer value clamping Auto assign reviewers and assignees #18: Pull request #181 opened by Copilot AI
7s
7s
[BUG FIX] SM80 NaN in bias.grad when both mask and bias are enabled Auto assign reviewers and assignees #17: Pull request #179 opened by LoserCheems
9s
9s