DeepSeek/DeepGEMM
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepGEMM synced 2025-05-05 08:54:22 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Update roadmap

Browse Source
This commit is contained in:
Chenggang Zhao 2025-04-09 11:41:36 +08:00
parent fed3e4d701
commit 677143be64
1 changed files with 18 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
README.md
View File

@ -10,6 +10,24 @@ Despite its lightweight design, DeepGEMM's performance matches or exceeds expert
- 2025.04.09: DeepGEMM now achieves up to **1520 TFLOPS** on H800! See [#74](https://github.com/deepseek-ai/DeepGEMM/pull/74), [#78](https://github.com/deepseek-ai/DeepGEMM/pull/78), and [#81](https://github.com/deepseek-ai/DeepGEMM/pull/81) for details.
## Roadmap
- [ ] More correctness tests for grouped-contiguous layout
- [ ] Shared memory swizzling for output instead of padding
- [ ] Larger block size on N (up to 256)
- [ ] MoE scheduler with TMA multicast compatibility
- [ ] Weight gradient kernels for dense models
- [ ] Weight gradient kernels for MoE models
- [ ] Utility kernels for MoE models (as a pre-built CUDA library)
- [ ] CUDA PDL support
- [ ] More scaling granularity support via templates
- [ ] Larger TMA multicast size for some shapes
- [ ] MMA template refactor with CUTLASS
- [ ] Optimizations for unaligned shapes
- [ ] Optimizations for power efficiency
- [ ] Remove shape limitations on N and K
- [ ] BF16 kernels
## Performance
We test all shapes potentially used in DeepSeek-V3/R1 inference (including both prefilling and decoding, but without tensor parallelism) on H800 SXM5 with NVCC 12.8. All speedup metrics are calculated in comparison to our internally and carefully optimized implementation based on CUTLASS 3.6.