Feat: Add new efficiency test

[anyangml]

The original design uses fixed datasets for LAMs. However, since different LAMs have varying GPU memory footprints during inference, they can trigger out-of-memory (OOM) errors with different structure sizes.

This PR introduces a new procedure with the following changes:

• The maximum number of atoms that a LAM can handle on a given device is dynamically determined using binary search.
• The primitive cells of the test structures are expanded on-the-fly to approach the maximum allowed number of atoms, ensuring convergence.

TODO:

• add UTs

[codecov]

Codecov Report

Attention: Patch coverage is 95.06173% with 4 lines in your changes missing coverage. Please review.

Project coverage is 65.99%. Comparing base (18630d9) to head (c7e4b9c).
Report is 14 commits behind head on main.

Files with missing lines	Patch %	Lines
...alculator/inference_efficiency/efficiency_utils.py	93.22%	3 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #264      +/-   ##
==========================================
+ Coverage   64.40%   65.99%   +1.59%     
==========================================
  Files          33       35       +2     
  Lines        1458     1538      +80     
  Branches      170      182      +12     
==========================================
+ Hits          939     1015      +76     
- Misses        482      485       +3     
- Partials       37       38       +1     

☔ 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.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

Pull Request Overview

This PR introduces a dynamic method to determine the maximum number of atoms processable by a model for the efficiency test and expands test structures on-the-fly accordingly. Key changes include:

• Adding a new OOM test atom and dynamic binary search for maximum atoms.
• Updating the inference functions to apply on-the-fly structure expansion.
• Introducing utility functions in efficiency_utils.py for factorization and binary search operations.

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
lambench/tasks/calculator/inference_efficiency/inference_efficiency.py	Updated inference function to dynamically expand atoms and include a binary search for max atoms.
lambench/tasks/calculator/inference_efficiency/efficiency_utils.py	Added utility functions for binary search, factorization, and OOM error handling.

[Copilot]

Pull Request Overview

This PR adds a new efficiency testing procedure to dynamically determine the maximum number of atoms a LAM can process without running into OOM errors. Key changes include:

• Adding unit tests for efficiency utilities (find_even_factors and binary_search_max_natoms) in the tests folder.
• Integrating binary search logic for maximum natoms into the inference workflow and updating the run_one_inference API.
• Modifying metrics calculation code that applies cell-level error computation.

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
tests/tasks/calculator/test_efficiency_utils.py	Introduces tests for factor finding and binary search max_natoms functionality.
lambench/tasks/calculator/inference_efficiency/inference_efficiency.py	Updates run_inference and run_one_inference to leverage dynamic natom scaling using binary search.
lambench/tasks/calculator/inference_efficiency/efficiency_utils.py	Provides new helper functions for OOM detection, dynamic scaling, and factor balancing.
lambench/metrics/vishelper/metrics_calculations.py	Adjusts the method to apply an instability error function over DataFrame cells.

Comments suppressed due to low confidence (2)

lambench/tasks/calculator/inference_efficiency/efficiency_utils.py:9

• Typo found in the docstring: 'Perfrom' should be corrected to 'Perform'.

    Perfrom force field prediction for one system, return energy, forces and stress.

lambench/tasks/calculator/inference_efficiency/efficiency_utils.py:81

• [nitpick] The parameter name 'safe_guard' could be more descriptive; consider renaming it to 'max_iterations' for improved clarity.

def binary_search_max_natoms(model: ASEModel, atoms: Atoms, upper_limit: int = 1000, safe_guard: int = 15) -> int: