Skip to content

upstream sync - #2

Open
LucasWilkinson wants to merge 29 commits into
vllm-project:main-archivefrom
deepseek-ai:main
Open

upstream sync#2
LucasWilkinson wants to merge 29 commits into
vllm-project:main-archivefrom
deepseek-ai:main

Conversation

@LucasWilkinson

Copy link
Copy Markdown
Collaborator

No description provided.

beginlner and others added 6 commits April 28, 2025 18:53
* Add more GPU architctures support

* Merge fmha and mla runner

* add varlen & non varlen support, and add incontiguous tensor support

* update readme

* add varlen api

---------

Co-authored-by: dianzhangc <dianzhangc@nvidia.com>
@jdebache

Copy link
Copy Markdown

Hey @LucasWilkinson, any insight on how we can resolve the failing sign-off checks? I have some work based on the latest version of FlashMLA that I would like to try out in vLLM, and so it would be convenient for me if this could be merged 😅 .

The SWITCH_MAX_SPLITS macro only enumerated split buckets up to 160;
any num_splits in (160, 256] fell through to FLASH_ASSERT(false) at
kernel launch. On GPUs with high SM counts the decode scheduler can
request more than 160 splits, so extend the ladder with 192/224/256
buckets. Purely additive.

Signed-off-by: Zachary Aristei <zaristei@nvidia.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

10 participants