Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Prefer longer servers for fine-tuning, exclude unreachable #448

Merged
merged 2 commits into from
Aug 7, 2023

Conversation

borzunov
Copy link
Collaborator

@borzunov borzunov commented Aug 7, 2023

We choose longer servers to minimize the number of hops but leave some randomization to distribute the load. We also exclude servers known to be unreachable.

@@ -50,7 +50,7 @@ class SequenceManagerConfig:
ban_timeout: float = 15 # when a remote peer fails to respond, prevent routing to that peer for this many seconds
active_adapter: Optional[str] = None # name of active LoRA adapter (usually, Hugging Face repo)

max_pinged: int = 5 # max servers to ping from each sequence side, per update
max_pinged: int = 3 # max servers to ping from each sequence side, per update
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

3 servers for start, middle, and end of sequence per update. It's stored for 5 min, so we'll have info about up to 45 servers after a while.

span_weights = np.array(
[span.length if client_server_rtts.get(span.peer_id) != np.inf else eps for span in candidate_spans],
dtype=np.float64,
)
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If not pinged yet, we'll have (None != np.inf) == False here

@borzunov borzunov merged commit 2a15077 into main Aug 7, 2023
7 checks passed
@borzunov borzunov deleted the faster-fine-tuning branch August 7, 2023 17:43
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