Hzfengsy opened a new pull request, #15381: URL: https://github.com/apache/tvm/pull/15381
This PR introduces a new rule, especially for GEMV and decode-GEMV workloads. The current `DecodeGEMV` rule works for common reduction rules, so in this PR, we did the following renaming: 1. `DecodeGEMV` -> `Reduction` 2. `Reduction` -> `GeneralReduction` In the future, we would refactor `Reduction` and `GeneralReduction` into one rule. With this PR, the performance of LLMs decoding stage improves: | Model | Platform | Before | After | | --|--|-|-| | LLaMA-7b | cuda @RTX-3080 | 81.0 tok/s | 107.6 tok/s| | LLaMA-7b | vulkan @RTX-3080 | 59.9 tok/s |64.2 tok/s| | LLaMA-7b | metal @Apple-M1-Pro | 23.3 tok/s | 26.5 tok/s| | RedPajama-3B | cuda @RTX-3080 | 108.6 tok/s | 159.8 tok/s | | RedPajama-3B | vulkan @RTX-30800 | 72.9 tok/s | 78.8 tok/s | | RedPajama-3B | metal @Apple-M1-Pro | 37.0 tok/s | 40.3 tok/s | cc @spectrometerHBH @vinx13 -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
