Skip to content

[Python][RunInference] vLLM: prefix-aware batching and capacity-based autoscaling #40908

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.

Sign up for GitHub

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

Draft
AMOOOMA wants to merge 4 commits into
base: master
Choose a base branch
from
Draft

[Python][RunInference] vLLM: prefix-aware batching and capacity-based autoscaling #40908

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
07aac31
[RunInference] Add prefix-aware batching across workers
AMOOOMA Oct 30, 2026
63bcd39
[RunInference] Add a capacity-based autoscaling signal for model-serv…
AMOOOMA Oct 30, 2026
fd13ad7
[RunInference] Report the model identifier to model handlers
AMOOOMA Oct 30, 2026
24ee97a
[RunInference] vLLM: prefix-aware batching and capacity-based autosca…
AMOOOMA Oct 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions CHANGES.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,6 +70,8 @@

* X feature added (Java/Python) ([#X](https://github.com/apache/beam/issues/X)).
* (Python) Added support for Protobuf 7.x ([#40580](https://github.com/apache/beam/issues/40580)).
* (Python) `BatchElements` can group elements by key across workers (`key_fn`, `num_shards`), and RunInference model handlers can use it through `batch_key_fn` / `batch_key_num_shards`. `apache_beam.ml.inference.prefix_batching` adds a `PrefixBatchingConfig` and prompt/chat prefix extractors so LLM handlers can batch requests that share a prompt prefix and hit the server's prefix cache ([#40905](https://github.com/apache/beam/pull/40905)).
* (Python) vLLM model handlers accept a `prefix_batching_config` to batch requests that share a prompt prefix, and publish a capacity-based autoscaling signal (`num_loaded_models_<model_identifier>`, estimated from vLLM's completion rate and in-flight requests) so Dataflow's GPU autoscaling scales vLLM workers with load; tune or disable it with `autoscaling_config` / `enable_autoscaling_metric`. `apache_beam.ml.inference.autoscaling_metrics` provides the generic pieces for other model-server handlers ([#40908](https://github.com/apache/beam/pull/40908)).

## Breaking Changes

Expand Down
Loading
Loading