-
-
Notifications
You must be signed in to change notification settings - Fork 4.3k
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
[Performance] Optimize get_seqs
#7051
Conversation
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge). To run full CI, you can do one of these:
🚀 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
lgtm!
@@ -458,25 +459,24 @@ def __init__( | |||
self.prompt_adapter_request = prompt_adapter_request | |||
self.encoder_seq = encoder_seq | |||
self.trace_headers = trace_headers | |||
self._first_seq = next(iter(self.seqs_dict.values())) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think you can still keep self._first_seq = seqs[0]
, and use it to replace self.seqs[0]
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think it doesn't hurt much to use seqs[0]
without caching it? _first_seq
was introduced to avoid the overhead of retrieving a value from the dictionary. I believe the overhead of seqs[0]
will be negligible even if it's Python.
Also, since the sequence can be removed, I feel more comfortable with self.seqs[0]
than caching the sequence.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Glad to see it helps performance.
This PR optimizes the overhead of
seq_group.get_seqs()
, which was reported by @youkaichao.The solution is simple: We maintain
seqs: List[Sequence]
in addition toseqs_dict: Dict[int, Sequence]
, and useseqs
for allget_seqs
calls.This leads to small performance boost (llama3 8B, 1xH100)
Throughput: 23.98 requests/s, 9914.65 tokens/s
Throughput: 24.52 requests/s, 10138.92 tokens/s