Skip to content

Conversation

@harrisonyhq
Copy link
Collaborator

Purpose

  1. Support broadcast on ascend due to different shape of mla model k cache (nope, rope);
  2. Avoid create empty tensors everytime when need to broadcast, use buffer to decrease create time cost.

Modifications

See file changes.

Test

Model Platform TP Input length Concurrency Broadcast Odirect TTFT
DeepSeek-V2-Lite CUDA 4 4K 50 False False 855 ms
DeepSeek-V2-Lite CUDA 4 4K 50 True False 523 ms

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.

1 participant