Add parallel MLA inference tests, fix corresponding issues#3787
Draft
janEbert wants to merge 7 commits intoNVIDIA:mainfrom
Draft
Add parallel MLA inference tests, fix corresponding issues#3787janEbert wants to merge 7 commits intoNVIDIA:mainfrom
janEbert wants to merge 7 commits intoNVIDIA:mainfrom
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
MLA used to not be tested for parallel dynamic inference. This PR implements appropriate test parameters into
TestDynamicInferenceEngine.test_parallel_inferenceto test MLA with and without latent caching (although MLA currently only works with latent caching).The test itself would fail without further changes, so the PR contributes several fixes to the code and improvements like making FlashMLA optional. For example, FlashMLA expects a block size of exactly 64 while FlashAttention expects a block size divisible by 256. In the generative phase, FlashMLA is not used, only FlashAttention, so any inference work other than decode-only tasks would not work with MLA without the changes in this PR.
Since the added tests are smoke-tests only, it's important to also add functional tests in future revisions to ensure correctness of the results.
Finally, tests for Mamba with the inference-optimized Transformer implementation are enabled since they pass without issues.