Always use 64 as the block size of moe_align kernel to avoid lds out of limit #303
Conversation
This file contains 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
charlifu
changed the title
gshtras
reviewed
Dec 4, 2024
| if (threadIdx.x < num_experts) { | ||
| tokens_cnts[index(num_experts, 0, threadIdx.x)] = 0; | ||
| for (int eid = threadIdx.x; eid < num_experts; eid += blockDim.x) { | ||
| tokens_cnts[index(num_experts, 0, eid)] = 0; | ||
| for (int i = 1; i <= blockDim.x; ++i) { |
There was a problem hiding this comment.
Does this have to be a nested loop?
Could a reduce operation be more efficient here?
There was a problem hiding this comment.
Each thread is doing its own reduce op on an independent data array. In this case warp reduce primitive does not apply.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
The moe_align kernel uses the max value between num of experts and warp size as the gpu block size. This leads to lds out of limit when handling model with big number of experts.
This PR sets the block size to always be warp size, and changes the kernel to handle the situation when thread number is smaller than number of experts.