DeepSeek/DeepGEMM
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepGEMM synced 2025-05-06 05:34:22 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity
98 Commits 3 Branches 0 Tags 3.4 MiB
95e81b3dd6
Commit Graph

26 Commits

Author SHA1 Message Date
yukuai26
95e81b3dd6
Indivisible TMA (#90) 
Fix indivisible shapes for TMA multicast

---------

Co-authored-by: yukuai <yukuai@deepseek.com>
Co-authored-by: Chenggang Zhao <chenggangz@deepseek.com>
 2025-04-23 14:55:14 +08:00
yukuai26
891f35adf5
Support TMA multicast on B with m_grouped_gemm_contiguous. (#88) 2025-04-21 09:43:17 +08:00
Chenggang Zhao
37aa127451
Use swizzling instead of padding (#86) 
* Add swizzling params

* Add TMA D descriptor

* Always use STSMx2

* Swizzling draft

* Compatible with padding

* Fix bugs

* Optimize swizzle performance

* Optimize expression

* Optimize TMA issues

* Fix README

* Stricter assertions
 2025-04-14 15:20:58 +08:00
Chenggang Zhao
8041ed7164 Use 1D TMA store 2025-04-11 10:42:01 +08:00
Chenggang Zhao
a6524d411a Larger block N candidates 2025-04-09 10:11:43 +08:00
Chenggang Zhao
48a5f071be Clean up config heuristics 2025-04-09 10:01:15 +08:00
sazc
97575bf1c6 Performance: BlockTile 256x128 optimizations enable 1500+ TFLOPS FP8 performance on the H800-SXM platform 2025-04-08 17:42:23 +08:00
Chenggang Zhao
b4ecf9c3ff Fix TMA multicast bugs 2025-04-07 14:34:42 +08:00
Chenggang Zhao
bff5724ded Code format 2025-04-07 09:32:43 +08:00
Yi Zhang
776bd0cccc
add lru-cache to avoid repeated calculation 2025-04-04 12:44:26 +08:00
Chenggang Zhao
6db7e1863b Solve STSM bank conflict via padding and 3D TMA 2025-04-03 15:39:35 +08:00
Chenggang Zhao
25db8de345 Better performance 2025-03-25 17:34:06 +08:00
Chenggang Zhao
1999d553e5 Lower TMA requirement 2025-03-25 17:18:53 +08:00
Chenggang Zhao
3497428a5e Minor fix 2025-03-25 15:16:26 +08:00
Chenggang Zhao
7ffb118e54 Support multicasting on B 2025-03-25 14:56:42 +08:00
Chenggang Zhao
b922e64cb2 Support block size 160 2025-03-25 13:37:59 +08:00
sazc
46eb0d08fb Performance: Larger BlockTile optimizations enable 1470+ TFLOPS FP8 performance on the H800-SXM platform 2025-03-25 10:44:57 +08:00
Chenggang Zhao
5e4badc577 Fix type lint 2025-03-10 13:10:16 +08:00
sazc
bed67b234c Minor fix 2025-03-10 13:02:02 +08:00
sazc
ed278eddd3 formats: Optimize get_best_configs implementation 2025-03-10 12:56:14 +08:00
sazc
50cf26cc7c Performance: Configuration algorithms tuned to minimize the impact of tail effects, now up to 1402 TFLOPS 2025-03-10 11:45:05 +08:00
Chenggang Zhao
6c5da03ba9 Support more shapes 2025-02-28 10:04:59 +08:00
Chenggang Zhao
6e10cba207 Minor fix 2025-02-28 09:21:35 +08:00
Liang
fbec9e5eee
Update get_best_configs 
a better strategy to choose config
 2025-02-27 23:18:52 +08:00
AcraeaTerpsicore
96b31fd6bb
fix typo 2025-02-26 18:37:22 +08:00
Chenggang Zhao
a6d97a1c1b Initial commit 2025-02-25 22:52:41 +08:00