Add LlamaMode matmul for decode #632
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Added a new target 'next-matmul-llama' to the makefile for processing MLIR files with various transformations and compiling to LLVM.
Contributor
Author
|
Below are the DeepSeek end-to-end test results. Multiple tests were conducted, and two screenshots are selected here for demonstration: the left side shows the new matmul implementation, while the right side shows BLIS matmul. The average time per token was reduced by 0.05 seconds. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Added Llama Mode matmul for the decode stage.Implemented the handwritten-style next-matmul-llama.mlir and the corresponding Pass,MatMulLlamaOptimize.cpp.Also added the test case matmul-vectorization-llama.mlir.It has been successfully registered in buddy-opt and can be directly invoked using the-matmul-vectorization-llama option.
Below is the performance comparison for the deepseekR1 decode stage use case,with the upper part representing the BLIS mode and the lower part representing the Llama mode.
That‘s all