Revert "[Kernel] changing fused moe kernel chunk size default to 32k (#7995)" #207
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
…llm-project#7995)" This reverts commit 34a0e96.
|
PS this is specifically to address SWDEV-486909 |
divakar-amd
approved these changes
Sep 25, 2024
There was a problem hiding this comment.
Looks good! Bigger chunk size is even better for us for 2 reasons:
- It has "slightly" better latency.
- We don't have OOM problem on mi30X. So, no reason for smaller (32k) chunk size
Upstream PR reference: vllm-project#7995
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Outdated suggestions cannot be applied.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.
Reverting the change that causes a regression on mixtral 8x22 with fp8 quantization (both dynamic and pre-quantized) on MI300 and H100
There was also a mismatch in the default value between the 2 sections that this also fixes.