NVIDIA/cutlass
1
0
Fork 0
mirror of https://github.com/NVIDIA/cutlass.git synced 2026-05-12 09:15:56 +00:00
Code Issues Packages Projects Releases Wiki Activity
853 Commits 21 Branches 52 Tags
772fbb264e0fe823d62dec10a44a67e1503afdfc
Commit Graph

853 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
Linfeng Zheng
772fbb264e [CLI] add cutedsl fp16 gemm tutorial from 2 to 6 (#3106) 
* [CLI] add fp16 gemm tutorial from 2 to 6

* [CLI] refine comments
 2026-03-17 10:11:55 +08:00
Blake Ledden
087c84df83 docs: Fix float16 documentation in elementwise_add notebook (#2949) (#3047) 
The notebook uses float16 tensors but the vectorized kernel documentation
incorrectly describes elements as 32-bit and uses 4-element vectorization.
Updated to correctly state 16-bit elements with 8-element vectorization
for proper 128-bit loads/stores.

Signed-off-by: Blake Ledden <bledden@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
 2026-03-12 10:29:46 +08:00
dePaul Miller
73c59c055c Support for Group GEMM in CUTLASS Profiler for Geforce and Spark (#3092) 
Co-authored-by: dePaul Miller <23461061+depaulmillz@users.noreply.github.com>
 2026-03-06 20:36:29 -05:00
Johnsonms
e5fcd125a5 [fix] Boolean.__dsl_and__ emits arith.andi directly for i1 operands (#3087) 
Before this fix, combining two Boolean (i1) DSL values with Python `and`
triggered a verbose i1→i32→i1 round-trip in __dsl_and__:
  arith.extui  (×3), arith.select, arith.cmpi ne (×2) — 6 extra MLIR ops.

Add a fast path: when both operands are Boolean, delegate directly to
__and__, emitting a single arith.andi %a, %b : i1 — identical to `&`.

Both operators were already semantically equivalent; this fix makes the
generated MLIR identical as well.

Includes:
- repro_dsl_and_bool.py  — minimal standalone reproducer / bug-report script
- test_dsl_and_fix.py    — pytest tests verifying the fixed behaviour
 2026-03-05 17:20:26 +08:00
TLescoatTFX
a93d86ec83 Fix finding cuDNN (#2890) 
* Fix finding cuDNN

* Search $CUDNN_PATH first, not last
 2026-03-05 09:51:37 +08:00
David W.H. Swenson
49e54f2b23 fix: add_help=False in temporary parser (#2721) 2026-03-02 15:33:42 +08:00
drazi
b9847690c5 Merge pull request #3028 from SzymonOzog/patch-3 
Add option to not suffix prints with new line
 2026-02-28 10:11:05 +08:00
Junkai-Wu
3bb6e28d3c v4.4.1 update (#3079) 2026-02-27 13:59:21 -05:00
Tianqi Zhang (张天启)
c651d660d2 fix typo (#3012) 2026-02-27 16:25:35 +08:00
Ziang Li
518327d631 Fix error in Blackwell document of referring to Mxf4 format as NVF4 (#2977) 
* Update blackwell_functionality.md

Fixed the error of referring to Mxf4 format as NVF4.

* Correct NVF4 output matrix reference in documentation
 2026-02-27 16:25:16 +08:00
StevenYangCC
de67bb7a42 Fix example in CuTe tutorials (#2752) 2026-02-27 16:24:34 +08:00
Neil Kichler
edf2f82c00 Fix register index bug in mma.sync.aligned.m16n8k16 (#2740) 2026-02-27 16:24:18 +08:00
mnehete32
79345359a7 Fix debug typo in sgemm_2.cu and sgemm_sm70.cu (#2678) 2026-02-27 16:23:59 +08:00
zkyue
8b9b3d78df fix typo in documentation (#2671) 2026-02-27 16:23:37 +08:00
Gabriel Wu
fc5bbc2dab Fix typo in cute.nvgpu.warpgroup.mma doc (#2548) 2026-02-27 16:22:55 +08:00
Junkai-Wu
057635de5c Remove redundant dsl example. (#3074) 2026-02-26 08:10:59 -05:00
Junkai-Wu
c213bfdfc1 Remove redundant dsl examples. (#3071) v4.4.0 2026-02-25 22:42:01 -05:00
Haicheng Wu
954503d44c Bump version to 4.4.0 2026-02-25 00:04:04 -05:00
Haicheng Wu
6c4200f1bc Bump version from 4.3.5 to 4.4.0 2026-02-25 00:03:23 -05:00
Haicheng Wu
de93e8a4ac Bump version from 4.3.5 to 4.4.0 2026-02-25 00:03:04 -05:00
Haicheng Wu
b92b9f0d37 Bump version from 4.3.5 to 4.4.0 2026-02-25 00:02:41 -05:00
Haicheng Wu
2aedca6f5e Bump CUTLASS version to 4.4.0 2026-02-25 00:01:56 -05:00
Haicheng Wu
6450964b57 Update README 2026-02-24 23:55:55 -05:00
Haicheng Wu
284449fa5b Revise chagnelog 2026-02-24 23:54:56 -05:00
Haicheng Wu
0853d81d70 Revise README 2026-02-24 15:32:17 -05:00
Linfeng Zheng
3476ddb7bd remove mixed_input_fmha_prefill (#3041) 2026-02-18 07:59:01 -05:00
Yihan Chen
291300ffff [CuTeDSL] implment a cta-level norm example (both layernorm and rmsnorm) (#3009) 
* kernel impl

* add copyright
 2026-02-14 17:54:03 +08:00
aragorn-guan
f9a5f76b7a Replace fence proxy to the latest routine code in examples/distributed/all_reduce_tma.py (#3027) 2026-02-14 17:51:20 +08:00
drazi
ec7e6cb17b Merge pull request #2971 from rsmallblue/tvm-ffi 
[CuTeDSL]fix tvm-ffi path in from_dlpack
 2026-02-14 14:14:10 +08:00
Yuan Xiaolan
395ab575f6 Merge branch 'main' into tvm-ffi 2026-02-14 13:35:28 +08:00
Junkai-Wu
d4bbf728ca v4.4 tag release update. (#3032) 2026-02-13 23:27:58 -05:00
Szymon Ożóg
beb80e04e1 Add option to not suffix prints with new line 2026-02-13 15:56:50 +01:00
drazi
01687cfba1 Merge pull request #3004 from tridao/add-sub-packed-f32x2 
[CuTeDSL] Add sub_packed_f32x2 operation
 2026-02-13 20:46:26 +08:00
drazi
5c42d0f28c Merge pull request #3021 from tridao/clc_no_multicast 
[Cute-DSL] Add option for issue_clc_query without multicast
 2026-02-13 20:45:52 +08:00
drazi
1d36152f34 Merge pull request #3022 from tridao/nvvm_fmin 
[Cute-DSL] Add cute.arch.fmin by calling nvvm
 2026-02-13 20:45:08 +08:00
Tri Dao
244e8d00d5 [Cute-DSL] Add cute.arch.fmin by calling nvvm 2026-02-11 14:23:09 -05:00
Tri Dao
5b83b34afd [Cute-DSL] Add option for issue_clc_query without multicast 2026-02-11 14:19:29 -05:00
aragorn-guan
8dbce01473 [CuTeDSL] Distributed example, using TMA load to access remote memory rank-by-rank, reducing in cta, broadcast result to all ranks by multimem TMA store (#2970) 2026-02-11 11:54:00 +08:00
drazi
71aa7a0abc Merge pull request #2919 from pbelevich/patch-1 
Refactor binary_op functions to remove unused result parameter
 2026-02-11 11:48:58 +08:00
Tri Dao
51935551fb [CuTeDSL] Add sub_packed_f32x2 operation 
Add subtraction operation for packed f32x2 values, following the same
pattern as the existing add_packed_f32x2 and mul_packed_f32x2 operations.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-02-04 21:18:46 +07:00
Junkai-Wu
6b3e607b85 v4.4 release update v2. (#2999) 2026-02-03 20:48:31 -05:00
yuanxiaolan
de161925a5 pass in stream=-1 2026-02-03 11:59:14 +08:00
yuanxiaolan
de198b2419 fix tvm-ffi path in from_dlpack 2026-02-03 11:59:13 +08:00
Hua Huang
1cfbb53a23 [CuTeDSL] Fix: SM100 block-scale gemm overlapping accumulator (#2995) 
* Fix: SM100 block-scale gemm overlapping accumulator

Signed-off-by: Hua Huang <huah@nvidia.com>

* Also include threads_per_warp fix

Signed-off-by: Hua Huang <huah@nvidia.com>

---------

Signed-off-by: Hua Huang <huah@nvidia.com>
 2026-02-03 11:01:41 +08:00
dongxiao
a4eb0e05f6 fix performance inssues in cute-dsl examples for 4.4-ctk13.1 release (#2988) 
* fix grouped gemm

* fix mixed input gemm

* fix mixed input grouped gemm

* fix version checking

* use advanced compiler options

* fix comment

* rename advanced compiler configs to adcanced compiler control

* fix comment

* fix name

* fix name
 2026-01-30 13:31:04 +08:00
myu-guo
d252b01300 fix performance regression in cute-dsl examples for 4.4-ctk13.1 release (#2990) 
* fix regression with cu13.1

* update
 2026-01-30 13:30:49 +08:00
Xiao Song
acb45938e9 Update nvvm API call from nvvm enum to str (#2985) 2026-01-27 17:28:29 +08:00
Xiao Song
7a14467776 update api usage (#2969) 2026-01-27 15:33:22 +08:00
drazi
51f82812ec Merge pull request #2891 from ColinPeppler/main 
docs: note when DSL dumps are populated
 2026-01-26 17:38:27 -08:00
Junkai-Wu
9fba3195f9 v4.4 update. (#2979) 2026-01-24 11:46:17 -05:00