Morph MORPH
SIGN IN SIGN UP
dhruvhead / flash-linear-attention UNCLAIMED
0 0 0 Python

ACTIVITY

Commits Reviews
week month year
COMMITS
13
in the last week
CONTRIBUTORS
7
active
STARS
0
total
FORKS
0
total
TOP CONTRIBUTORS
M
morluto
4 commits
C
ChunyuWei
2 commits
A
Ascendyh
2 commits
S
sunyi0505
2 commits
M
Michael R
1 commit
Z
Zhiyuan Li
1 commit
A
Annie Guo
1 commit
RECENT COMMITS
M
[Attn] Support for heterogeneous hybrid-attention configurations (#1051)
Michael R
d1ce073
C
[CI] add ascend-a2-benchmark-ci and fix ascend-a2-ci (#1052)
ChunyuWei
89af578
C
[Ops] Add triton-ascend backend for KDA kernels (#1047)
ChunyuWei
33fcbdf
Z
[Cleanup] Remove Ascend chunk_fwd_o kernels superseded by #1049 (#1050)
Zhiyuan Li
e537b7c
A
[Perf] Fuse Ascend chunk_fwd_o inter+intra into unified 1D core-grid kernel (#1049)
Ascendyh
7e188d7
S
[Perf] Fuse ds in Ascend chunk_bwd_dqkwg to avoid recomputing do@v.T (#1048)
sunyi0505
cac37a9
A
[Perf] Add opt-in TileLang RWKV6 intra kernel (#1045)
Annie Guo
c70f11c
S
[Perf][Misc] Add LayerNormGated benchmarks and BT-tile Ascend fused_norm_gate (#1044)
sunyi0505
f3acf5d
M
[Perf] Cache and bucket Wall attention autotuning (#1041)
morluto
cbb0a72
M
[Fix] Support value-split Wall backward with local deltas (#1042)
morluto
cff3451
A
[Ops][GDN][Perf] Rewrite recompute_w_u_fwd kernel for Triton Ascend backend (#1021)
Ascendyh
ee23e19
M
[Refactor] Centralize generation errors for unsupported cache strategies (#1035)
morluto
fe8fce9
M
[Fix] Avoid concurrent LSE stores in value-split attention (#1033)
morluto
2cec2a3