Add model manager that automatically manage model across processes #37113

AMOOOMA · 2025-12-15T21:54:30Z

Added Model Manager as a util class that offers managed access to models, the client can request models without having to worry about managing GPU OOMs.
Also added various tests that checks the functions of all classes.

Classes

GPUMonitor

start(): Begins background memory polling.
stop(): Stops polling.
reset_peak(): Resets peak usage tracking.
get_stats() -> (current, peak, total): Returns memory stats.

ResourceEstimator

is_unknown(model_tag: str) -> bool: Checks if model needs profiling.
get_estimate(model_tag: str, default_mb: float) -> float: Returns memory cost.
set_initial_estimate(model_tag: str, cost: float): Manually sets cost.
add_observation(active_snapshot, peak_memory): Updates cost model via NNLS solver.

ModelManager

acquire_model(tag: str, loader_func: Callable) -> Any: Gets model instance (handles isolation/concurrency).
release_model(tag: str, instance: Any): Returns model to pool.
force_reset(): Clears all models and caches.
shutdown(): Cleans up resources.

Thank you for your contribution! Follow this checklist to help us incorporate your contribution quickly and easily:

Mention the appropriate issue in your description (for example: addresses #123), if applicable. This will automatically add a link to the pull request in the issue. If you would like the issue to automatically close on merging the pull request, comment fixes #<ISSUE NUMBER> instead.
Update CHANGES.md with noteworthy changes.
If this contribution is large, please file an Apache Individual Contributor License Agreement.

See the Contributor Guide for more tips on how to make review process smoother.

To check the build health, please visit https://github.com/apache/beam/blob/master/.test-infra/BUILD_STATUS.md

GitHub Actions Tests Status (on master branch)

See CI.md for more information about GitHub Actions CI or the workflows README to see a list of phrases to trigger workflows.

gemini-code-assist · 2025-12-15T21:54:48Z

Summary of Changes

Hello @AMOOOMA, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces a sophisticated model management system for Apache Beam's ML inference capabilities. The core "ModelManager" class, supported by "GPUMonitor" and "ResourceEstimator", intelligently handles the lifecycle of machine learning models, particularly on GPU-accelerated environments. It aims to prevent out-of-memory errors by dynamically estimating model memory requirements, isolating unknown models for profiling, and implementing a demand-aware eviction strategy. This system ensures efficient and concurrent execution of diverse ML models within Beam pipelines, optimizing GPU resource utilization and improving overall stability.

Highlights

ModelManager Introduction: A new "ModelManager" class is added to provide managed access to ML models, handling GPU OOMs and optimizing resource usage.
GPUMonitor for Memory Tracking: Implements "GPUMonitor" to continuously poll and track GPU memory usage, including current, peak, and total memory.
ResourceEstimator for Cost Estimation: Introduces "ResourceEstimator" which uses Non-Negative Least Squares (NNLS) to dynamically estimate the memory cost of models and adapt to fluctuating usage.
Intelligent Model Eviction: The "ModelManager" includes an eviction strategy that prioritizes models based on demand, age, and surplus copies to free up GPU memory when needed.
Isolation Mode for Unknown Models: Unknown models are loaded in an isolated environment to accurately profile their memory footprint without affecting other active models.
Comprehensive Testing: New unit tests are added to validate the functionality of "GPUMonitor", "ResourceEstimator", and "ModelManager", covering capacity checks, isolation, concurrent execution, OOM recovery, and eviction logic.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

AMOOOMA · 2025-12-15T21:54:53Z

R: @damccorm

github-actions · 2025-12-15T21:59:57Z

Stopping reviewer notifications for this pull request: review requested by someone other than the bot, ceding control. If you'd like to restart, comment assign set of reviewers

sdks/python/apache_beam/ml/inference/model_manager.py

damccorm · 2026-01-13T20:02:38Z

sdks/python/apache_beam/ml/inference/model_manager.py

+      logger.info("Initial Profile for %s: %s MB", model_tag, cost)
+
+  def add_observation(
+      self, active_snapshot: Dict[str, int], peak_memory: float):


I'm having a little bit of a hard time following this. Is active_snapshot a map of model tags to the number of models loaded for that tag?

Yep! That's correct.

Could we update the log below to be clear on that? Ideally it would format the dict in a human readable way that makes that clear

Yes for sure, updated the print string.

damccorm

I've been looking at it for a while today and I am still having a hard time understanding the full scope of this PR (even with this being the second long look), and it will probably take a few more passes.

For now, added some things that will help me review better, but if there are pieces we can separate out further to make this more reviewable (either by splitting large functions apart or by pulling classes out of the PR), that would be quite helpful.

damccorm · 2026-01-26T22:02:13Z

sdks/python/apache_beam/ml/inference/model_manager.py

+      logger.info("Initial Profile for %s: %s MB", model_tag, cost)
+
+  def add_observation(
+      self, active_snapshot: Dict[str, int], peak_memory: float):


Could we update the log below to be clear on that? Ideally it would format the dict in a human readable way that makes that clear

damccorm · 2026-01-26T22:06:47Z

sdks/python/apache_beam/ml/inference/model_manager.py

+
+    return evicted_something
+
+  def _perform_eviction(self, key, tag, instance, score):


Can we add types to the args here and elsewhere?

damccorm · 2026-01-26T22:07:04Z

sdks/python/apache_beam/ml/inference/model_manager.py

+        del self._models[tag][i]
+        break
+
+    if hasattr(instance, "trackedModelProxy_unsafe_hard_delete"):


Does this ever not exist?

I originally was aiming for this model manager to be versatile and can store other objects too but I guess at this point that ship has sailed. We won't need this check anymore. Updated.

damccorm · 2026-01-26T22:12:11Z

sdks/python/apache_beam/ml/inference/model_manager.py

+      with self._load_lock:
+        logger.info("Loading Model: %s (Unknown: %s)", tag, is_unknown)
+        isolation_baseline_snap, _, _ = self._monitor.get_stats()
+        instance = TrackedModelProxy(loader_func())


What is loader_func here - is this spawning the model in process?

Yep! This depends on what the user pass in which in the RunInference case would be spawning a model in process with the new MultiProcessShared util.

damccorm · 2026-01-26T22:17:26Z

sdks/python/apache_beam/ml/inference/model_manager.py

+    ticket_num = next(self._ticket_counter)
+    my_id = object()
+
+    with self._cv:


Is thread contention going to be an issue here/do we need this? If we are using this model_manager in a multi_process_shared setting, the calls to this (and other functions) should be serialized.

I'm a bit worried that this serialization is going to cause broader perf issues

+1 I was also worried but as a first step this is kept in this way since model inference can take some time. I originally tried splitting up the read and load lock to not have loading models block normal operations to grab stuff from the idle pool but I think there's some race condition that can happen if we do release model and model loading at the same time, the snapshot taken can be broken in some cases. We can probably figure out a smarter way to unblock this perf issue though, I will add a comment and defer it to later for now.

damccorm · 2026-01-26T22:18:00Z

sdks/python/apache_beam/ml/inference/model_manager.py

+
+    with self._cv:
+      # FAST PATH
+      if self._pending_isolation_count == 0 and not self._isolation_mode:


Is self._pending_isolation_count ever non-zero?

Good catch! This is some old code from older iterations. Removed now.

damccorm · 2026-01-26T22:18:18Z

sdks/python/apache_beam/ml/inference/model_manager.py

+    self._total_active_jobs = 0
+    self._pending_reservations = 0.0
+
+    self._isolation_mode = False


Could you add comments describing what these are?

damccorm · 2026-01-26T22:18:56Z

sdks/python/apache_beam/ml/inference/model_manager.py

+  It handles:
+  1. LRU Caching of idle models.
+  2. Resource estimation and admission control (preventing OOM).
+  3. Dynamic eviction of low-priority models when space is needed.


Could you add more info on what makes a model low-priority here?

damccorm · 2026-01-26T22:20:24Z

sdks/python/apache_beam/ml/inference/model_manager.py

+  def all_models(self, tag) -> list[Any]:
+    return self._models[tag]
+
+  def acquire_model(self, tag: str, loader_func: Callable[[], Any]) -> Any:


This function is pretty large and I find it hard to follow; could you try breaking it up into subfunctions?

Yes! Sorry about this giant change I forgot it can be really hard to read if it's for the first time. Iterating it kind of just grow the code larger and larger but I was under the false impression that it's pretty much the same size as the first draft.

I added more comments and split this function into smaller pieces, hopefully that would help. Let me know if there's anything I can help! Thanks!

…code and cleanup some code logics

codecov · 2026-01-27T03:39:55Z

Codecov Report

❌ Patch coverage is 2.09790% with 420 lines in your changes missing coverage. Please review.
✅ Project coverage is 39.97%. Comparing base (3ce2abd) to head (8a856b0).
⚠️ Report is 233 commits behind head on master.

Files with missing lines	Patch %	Lines
...s/python/apache_beam/ml/inference/model_manager.py	2.09%	420 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff              @@
##             master   #37113      +/-   ##
============================================
- Coverage     40.38%   39.97%   -0.42%     
  Complexity     3476     3476              
============================================
  Files          1226     1222       -4     
  Lines        188339   187813     -526     
  Branches       3607     3607              
============================================
- Hits          76058    75074     -984     
- Misses       108883   109341     +458     
  Partials       3398     3398

Flag	Coverage Δ
python	`39.66% <2.09%> (-1.00%)`	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

Add model manager that automatically manage model across processes

10ee00a

AMOOOMA mentioned this pull request Dec 15, 2025

Add model manager that automatically allocate models and prevents OOMs #37103

Draft

3 tasks

github-actions bot added the python label Dec 15, 2025

damccorm reviewed Jan 13, 2026

View reviewed changes

Add pydoc and move gpu detection to start

b123a95

damccorm reviewed Jan 26, 2026

View reviewed changes

AMOOOMA added 2 commits January 27, 2026 03:13

Add comments and helper function to make it easier to understand the …

4f82caa

…code and cleanup some code logics

Add TODO for threading

8a856b0


		return evicted_something

		def _perform_eviction(self, key, tag, instance, score):

Add model manager that automatically manage model across processes #37113

Are you sure you want to change the base?

Add model manager that automatically manage model across processes #37113

Conversation

AMOOOMA commented Dec 15, 2025

Classes

GitHub Actions Tests Status (on master branch)

Uh oh!

gemini-code-assist bot commented Dec 15, 2025

Summary of Changes

Highlights

Footnotes

Uh oh!

AMOOOMA commented Dec 15, 2025

Uh oh!

github-actions bot commented Dec 15, 2025

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

damccorm left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

codecov bot commented Jan 27, 2026

Codecov Report

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants