[Update] update dev/perf branch to main branch commit 9c2df5 and cherry-pick mla_fix, mmd/fix/fwd_v3#1467

Merged

wuhuikx merged 115 commits intodev/perffrom

yuhua/dev/perf

Nov 24, 2025

Commits on Nov 23, 2025

integrate deep gemm (#1265)

lalala-sh
authored and
zhuyuhua-v
committed
add a tuned config and insert entries in untuned config (#1243)

authored and
zhuyuhua-v
committed
Enable large batch size and optimization of non-Ragged batching (#1269)

valechen
authored and
zhuyuhua-v
committed
add few more fw ds f4 untuned and tuned shapes for using asm kernel (#1298)

hongxiayang
authored and
zhuyuhua-v
committed
CI: Optimize autotuning pipeline and inital the docs (#1286)

gyohuangxin
authored and
zhuyuhua-v
committed
topk per row kernel (#1262)

authored and
zhuyuhua-v
committed
fix aot (#1279)

authored and
zhuyuhua-v
committed
Fix ATOM fp8 model quant fail issue in torch compile (#1299)

ZhangLirong-amd
authored and
zhuyuhua-v
committed
feat - pa_fwd support block map with stride in num_kv_heads_dim (#1301)

alibaba-miji
authored and
zhuyuhua-v
committed
Fix how to update accumulator for dot_scaled (#1297)

zhanglx13
authored and
zhuyuhua-v
committed
CI: Optimize autotuning pipeline docs (#1300)

authored and
zhuyuhua-v
committed
Fix the lint issue (#1307)

gyohuangxin
authored and
zhuyuhua-v
committed
fix fwd perf calc error (#1305)

minmengdie
authored and
zhuyuhua-v
committed
add the asm kernel performance of fwd and bwd (#1270)

minmengdie
authored and
zhuyuhua-v
committed
Fused TopK and Sigmoid kernel (#1251)

samremes
authored and
zhuyuhua-v
committed
Ar rms (#1290)

authored and
zhuyuhua-v
committed
Dsv32 cache (#1314)

authored and
zhuyuhua-v
committed
Fix displaying supported architectures (#1316)

HollowMan6
authored and
zhuyuhua-v
committed
using standalone pybind (#1317)

valarLip
authored and
zhuyuhua-v
committed
Enable mha bwd hd192_hd128 (#1308)

slippedJim
authored and
zhuyuhua-v
committed
CI: Add pre-check status check (#1252)

authored and
zhuyuhua-v
committed
[CK_TILE] fmha: Add backward pass support for padded inputs (#1212)

Jeff-Huang
authored and
zhuyuhua-v
committed
Mla splitkv enhance split alg inte (#1233)

authored and
zhuyuhua-v
committed
Fix gemm tuner error mi350 (#1313)

yzhou103
authored and
zhuyuhua-v
committed
CI: Skip triton setup in Aiter standard/multigpu tests and add retries when setting up triton (#1325)

gyohuangxin
authored and
zhuyuhua-v
committed
Fix global variable torch_fp8 initialization caused issue (#1322)

authored and
zhuyuhua-v
committed
Add transpose scale to the triton fused_rms_fp8_group_quant (#1291)

tjtanaa
authored and
zhuyuhua-v
committed
[Triton] 355 wip Llama FP4 triton fusion + TP8 triton decode shape tunning (#1315)

k50112113
authored and
zhuyuhua-v
committed
[TRITON] Kernel naming: add reusable constexpr repr helper (#1260)

Boss2002n
authored and
zhuyuhua-v
committed
Merge tuned file (#1327)

authored and
zhuyuhua-v
committed
fix graph_breaks by return tensor for bool op (#1333)

ZhangLirong-amd
authored and
zhuyuhua-v
committed
fix_bf16gemm_asm (#1329)

amd-ruitang3
authored and
zhuyuhua-v
committed
Improve Memory Usage in MLA (#1338)

ruanjm
authored and
zhuyuhua-v
committed
fix tune error caused by merge tuned_file (#1342)

yzhou103
authored and
zhuyuhua-v
committed
rm rocblas in tuner (#1337)

yzhou103
authored and
zhuyuhua-v
committed
[Triton] DS a16w8 GEMM and fused reduce_rms_fp8_group_quant (#1328)

k50112113
authored and
zhuyuhua-v
committed
Add block_m=16 for a8w8_ck_moe_blockscale (#1081)

authored and
zhuyuhua-v
committed
Add Fused RMSNorm + FP8 Per-tensor Static Quantization Triton Kernel (#1330)

farlukas
authored and
zhuyuhua-v
committed
[TRITON] GEMM kernels nomenclature changes (#1283)

Boss2002n
authored and
zhuyuhua-v
committed
Temporarily run aiter standard and multigpu tests on the TW cluster, will switch back once the mirror registry is ready. (#1359)

gyohuangxin
authored and
zhuyuhua-v
committed
[Triton] Disable failing lean attention tests (#1357)

cagrikymk
authored and
zhuyuhua-v
committed
update ck to fix fp4 gemm issue (#1361)

gino-lu
authored and
zhuyuhua-v
committed
add config (#1355)

valarLip
authored and
zhuyuhua-v
committed
add how_v3_bf16_cvt control to the Python API (#1351)

minmengdie
authored and
zhuyuhua-v
committed
[fix]: car 6 rank coredump (#1335)

authored and
zhuyuhua-v
committed
Wrapper_flash_attn_backward custom op to avoid functionalize fallback and fix guard logic (#1348)

authored and
zhuyuhua-v
committed
[TRITON] GEMM kernels nomenclature changes (#1292)

Boss2002n
authored and
zhuyuhua-v
committed
[TRITON] Initial implementations of sparse attention kernels (#1296)

cagrikymk
authored and
zhuyuhua-v
committed
[MI35X]cktile moe a16w4 support (#1341)

authored and
zhuyuhua-v
committed
[TRITON] Batched GEMM kernels nomenclature changes (#1293)

Boss2002n
authored and
zhuyuhua-v
committed
[TRITON] Instruction shape fix for Gluon gemm_a8w8_blockscale kernel (#1261)

authored and
zhuyuhua-v
committed
moe mxfp4 block_m = 64/128 (#1266)

authored and
zhuyuhua-v
committed
bug fix (#1370)

valarLip
authored and
zhuyuhua-v
committed
[opus] enhance opus utility (#1324)

carlushuang
authored and
zhuyuhua-v
committed
Fix issue in metadata v1.2 where batch size is too large (#1352)

ruanjm
authored and
zhuyuhua-v
committed
[GEMM][Config] add a8w8 block scale tuned config for deepseek-v3 (#1310)

authored and
zhuyuhua-v
committed
support all logit values (#1323)

authored and
zhuyuhua-v
committed
CI: Skip triton in Aiter standard and multigpu tests (#1374)

authored and
zhuyuhua-v
committed
add the performance data bar chart in the readme (#1372)

minmengdie
authored and
zhuyuhua-v
committed
force ds ptpc moe use 1 stage moe (#1373)

junhaha666
authored and
zhuyuhua-v
committed
[TRITON] MHA PA optimizations (#1245)

cagrikymk
authored and
zhuyuhua-v
committed
Enable fa multii target build on other arch (#1318)

authored and
zhuyuhua-v
committed
[Triton] DS FP4 triton fusion (#1371)

k50112113
authored and
zhuyuhua-v
committed
[TRITON] Simplify and optimize triton_kernels moe code and move it into aiter (#1326)

lburzawa
authored and
zhuyuhua-v
committed
Use torch.zeros_like instead of empty_like to prevent accruacy drop (#1387)

hubertlu-tw
authored and
zhuyuhua-v
committed
CI: Temporarily using old vllm nightly image (#1389)

gyohuangxin
authored and
zhuyuhua-v
committed
Revert "[Triton] DS FP4 triton fusion (#1371)" (#1392)

valarLip
authored and
zhuyuhua-v
committed
add a8w8 ptpc gemm config for dsv3 (#1382)

authored and
zhuyuhua-v
committed
Test the CI on both MI325 and MI355 (#1364)

gyohuangxin
authored and
zhuyuhua-v
committed
[Triton] change BF16 GEMM config filename (#1398)

k50112113
authored and
zhuyuhua-v
committed
Support distributed_init_method and DP in init_distributed (#1353)

ZhangLirong-amd
authored and
zhuyuhua-v
committed
FA V3(fp8) and paged Attention compressed (CI green) (#1065)

micmelesse
authored and
zhuyuhua-v
committed
is_shuffled (#1377)

authored and
zhuyuhua-v
committed
Ar rms new interface (#1401)

authored and
zhuyuhua-v
committed
minor fix for mi355 (#1408)

valarLip
authored and
zhuyuhua-v
committed
[Fix] Add sliding window feature for paged_attention_v1 (#1362)

authored and
zhuyuhua-v
committed
fused_qk_rope_cat_and_cache_mla: Fix Triton compilation error and batch size constraint and output tensor sizing (#1407)

xudonlyu
authored and
zhuyuhua-v
committed
max mla splits perbatch (#1390)

authored and
zhuyuhua-v
committed
topk_per_row_opt (#1394)

authored and
zhuyuhua-v
committed
Fix fused_rms_mxfp4_quant comment (#1369)

drewads
authored and
zhuyuhua-v
committed
leanAttn softmax fix for spurious data mismatch test failures (#1396)

valechen
authored and
zhuyuhua-v
committed
Add reduce_scatter api (#1413)

ZhangLirong-amd
authored and
zhuyuhua-v
committed
fix error in fmoe_tuner (#1405)

yzhou103
authored and
zhuyuhua-v
committed
optimize thread divergence (#1421)

carlushuang
authored and
zhuyuhua-v
committed
[TRITON] complex number multiplication that supports 3D ROPE triton kernel (#1061)

authored and
zhuyuhua-v
committed
Feat: pa_mqa_logits performance optimization & support kv_preshuffle + blocksize16/64 (#1424)

sjfeng1999
authored and
zhuyuhua-v
committed
[Config] add tuned moe and gemm config for qwen3 235b (#1378)

authored and
zhuyuhua-v
committed
fix repeated unnecessary device check (#1221)

b8zhong
authored and
zhuyuhua-v
committed
remove lru func in fake (#1429)

ZhangLirong-amd
authored and
zhuyuhua-v
committed
Temporarily disable the test on mi355 (#1437)

gyohuangxin
authored and
zhuyuhua-v
committed
Enable MI355 test on main branch

gyohuangxin
authored and
zhuyuhua-v
committed
CI: Aiter tests bug fix

gyohuangxin
authored and
zhuyuhua-v
committed
[M308] tune silu&act (#1404)

authored and
zhuyuhua-v
committed
add deepseek ep moe tune config (#1431)

junhaha666
authored and
zhuyuhua-v
committed
[TRITON] Moe a8w4 tuning (#1410)

lburzawa
authored and
zhuyuhua-v
committed
[TRITON] Apply config-aware naming (kernel_repr) to attention kernels (#1295)

Boss2002n
authored and
zhuyuhua-v
committed
[fix]: prebuild gen so (#1412)

authored and
zhuyuhua-v
committed
[TRITON] FP8 MQA optimizations (#1422)

cagrikymk
authored and
zhuyuhua-v
committed
redirect asm_moe_tkw1 call to fused_moe in order to force kernel tuning (#1366)

authored and
zhuyuhua-v
committed
CI: Move some tests to MI355 due to the network issue of TW cluster (#1446)

gyohuangxin
authored and
zhuyuhua-v
committed
CI: Move Triton tests from TW cluster to internal cluster (#1451)

gyohuangxin
authored and
zhuyuhua-v
committed
tune a8w8_blockscale&bpreshuffle for tencent (#1444)

LJ-underdog
authored and
zhuyuhua-v
committed
[fix]: add ar switch (#1376)

authored and
zhuyuhua-v
committed
update 3rdparty/composable_kernel to e31a7a4
zhuyuhua-v
committed
remote maxsize for mla lru cache

Zzz9990
authored and
zhuyuhua-v
committed
enable extenal num_kv_splits & num_kv_splits_ptr

Zzz9990
authored and
zhuyuhua-v
committed
fix output/lse is nan when kseq=0

minmengdie
authored and
zhuyuhua-v
committed
fix gfx950 128_128 fwd_v3

minmengdie
authored and
zhuyuhua-v
committed
update the k_seq=0 error in MI300 and MI308

minmengdie
authored and
zhuyuhua-v
committed
update the smoke test

minmengdie
authored and
zhuyuhua-v
committed
update the smoke test

minmengdie
authored and
zhuyuhua-v
committed
fix MI300 and MI308 err

minmengdie
authored and
zhuyuhua-v
committed
fix qseq >> kseq error MI300 and MI308

minmengdie
authored and
zhuyuhua-v
committed
fix qseq >> kseq error in MI355

minmengdie
authored and
zhuyuhua-v
committed
fix the MI300 error

minmengdie
authored and
zhuyuhua-v
committed