[BugFix] Fix default kv-cache-dtype default for DeepseekV3.2 #25988
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
Signed-off-by: Lucas Wilkinson <[email protected]>
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request addresses a bug where the specific configuration for DeepseekV3.2 models was not being applied. The changes involve renaming the configuration class to DeepseekV32ForCausalLM and updating the model configuration map accordingly. The logic for applying the custom KV cache settings has also been simplified. While the changes are generally good, I've identified a potential issue in the handling of the bfloat16 cache data type that could lead to unexpected behavior.
LucasWilkinson
changed the title
IwakuraRein
reviewed
Sep 30, 2025
Signed-off-by: Lucas Wilkinson <[email protected]>
simon-mo
approved these changes
Oct 1, 2025
simon-mo
pushed a commit
that referenced
this pull request
Oct 1, 2025
Signed-off-by: Lucas Wilkinson <[email protected]> Signed-off-by: simon-mo <[email protected]>
pdasigi
pushed a commit
to pdasigi/vllm
that referenced
this pull request
Oct 2, 2025
…oject#25988) Signed-off-by: Lucas Wilkinson <[email protected]>
yewentao256
pushed a commit
that referenced
this pull request
Oct 3, 2025
Signed-off-by: Lucas Wilkinson <[email protected]> Signed-off-by: yewentao256 <[email protected]>
tomeras91
pushed a commit
to tomeras91/vllm
that referenced
this pull request
Oct 6, 2025
…oject#25988) Signed-off-by: Lucas Wilkinson <[email protected]> Signed-off-by: Tomer Asida <[email protected]>
xuebwang-amd
pushed a commit
to xuebwang-amd/vllm
that referenced
this pull request
Oct 10, 2025
…oject#25988) Signed-off-by: Lucas Wilkinson <[email protected]> Signed-off-by: xuebwang-amd <[email protected]>
choprahetarth
pushed a commit
to Tandemn-Labs/vllm
that referenced
this pull request
Oct 11, 2025
…oject#25988) Signed-off-by: Lucas Wilkinson <[email protected]> Signed-off-by: simon-mo <[email protected]>
shyeh25
pushed a commit
to shyeh25/vllm
that referenced
this pull request
Oct 14, 2025
Signed-off-by: Lucas Wilkinson <[email protected]> Signed-off-by: simon-mo <[email protected]>
lywa1998
pushed a commit
to lywa1998/vllm
that referenced
this pull request
Oct 20, 2025
…oject#25988) Signed-off-by: Lucas Wilkinson <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
In the final release DeepseekV32 was being registered separately so the config override was not getting picked-up anymore