[CPU] Introduce GatherMatmul operation to optimize MoE pattern #32450
+2,942
−185
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
github-actions
bot
added
category: Core
2. initMoE2GeMMSubgraph builder is moved to a separate file 3. initMoE3GeMMSubgraph
maxnick
force-pushed
the
cpu_moe_op_support
branch
from
7d01283 to
a6d5711
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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.
Details:
In this PR we introduce yet another operation "GatherMatmu", which essentially does gemv operations over the current tokens and the active experts.
As the first step, we perform gemv operation using the dnnl::inner_product. But obviously this solution is suboptimal, as it doesn't give a fine grain control over parallelization, and in the case of many tokens being processed by a specific expert (prefill), having gemm operation may be more optimal as the tokens may be batched and we can do SIMD level parallelization by tokens as well.
Also this PR contains all the essential transformations that allow to enable a few common MoE patterns.
MoE pattern matcher is based on #32183
Related oneDNN fork PR: openvinotoolkit/oneDNN#292
Tickets: