Repository navigation
Continuous batching: find the sharded key/value projections under any decoder path - #50366
Open
qgallouedec wants to merge 1 commit into
Open
qgallouedec wants to merge 1 commit into
qgallouedec wants to merge 1 commit into
Conversation
… decoder path A vision-language model's TP plan keys start with model.language_model., so the engine took its key/value heads as replicated and sized the cache with all of them on every rank: Qwen3.5 at tp=2 fails on the first decode step.
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
This branch has not been deployed
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What does this PR do?
Under tensor parallelism, continuous batching decides whether the key/value heads are split across ranks by looking up
layers.*.self_attn.k_projormodel.layers.*.self_attn.k_projin the TP plan. A vision-language model's keys start withmodel.language_model., so the engine sizes the cache with every head on every rank and the first forward fails. Found running Qwen3.5-2B and 4B (VLM checkpoints) at tp=2 on B200:k has wrong shape at dim 2on flash, a shape mismatch on sdpa. The lookup now matches the key suffix.