would there be an official test/benchmark for a fused kernel of dsa?
- Dominant language
- Python
- Stars
- 7.4k
- Forks
- 745
- Avg merge
- 1d 1h
- Merged PRs (30d)
- 104
Description
I found independent kernels for indexer, topk selector and sparse mla (https://github.com/tile-ai/tilelang/blob/main/examples/deepseek_v32/test_tilelang_example_deepseek_v32.py), for which i am very grateful, but i did not see a fused kernel anywhere
is it because the implementation will highly depend on the setup?
would tilelang team consider providing such a kernel for benchmark and testing?
Contributor guide
Research direction
Start with examples/deepseek_v32/test_tilelang_example_deepseek_v32.py and review the existing independent indexer, topk selector, and sparse MLA kernels. Determine the expected fused-kernel scope, supported setups, and benchmark or testing criteria from the surrounding implementation and issue discussion. Done means the team has an agreed design and a reproducible fused kernel evaluation path.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- performance, testing
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100