Skip to content

Comments

reduce kv transfer process to num of tp for pd.#758

Closed
kingder wants to merge 6 commits intoModelTC:mainfrom
kingder:kv_trans_opt
Closed

reduce kv transfer process to num of tp for pd.#758
kingder wants to merge 6 commits intoModelTC:mainfrom
kingder:kv_trans_opt

Conversation

@kingder
Copy link
Collaborator

@kingder kingder commented Mar 7, 2025

combine kv transfer process per p/d pair to a single process per p or d for reducing vram per process.
still ~300M vram needed per p/d pair for nccl communicator.

@kingder kingder changed the title single kv transfer process for pd. reduce kv transfer process to num of tp for pd. Mar 11, 2025
@kingder kingder force-pushed the kv_trans_opt branch 2 times, most recently from 30b10a0 to aa3a044 Compare March 19, 2025 08:08
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.

2 participants