fix(generator): use logarithmic cuda graph sizes

[simone-chen]

Overview:

Set cuda_graph_batch_sizes to a logarithmic series of graphs than capturing all graph sizes with step 1.

Details:

• Refactored the _eval function in the rule engine to support a wider range of python expressions (e.g. list comprehension) for better flexibility. Instead of using jinja's expression compiler, use asteval to evaluate the rules in Python.
  • Why: The jinja expression compiler supports very limited python expression and doesn't support the use case of creating a logarithmic series. Without this change, we must define the series in all extra_engine_args*.yaml.j2.

    >>> env = Environment()
    >>> expr = env.compile_expression("[x * 2 for x in numbers]")
    Traceback (most recent call last):
      File "<stdin>", line 1, in <module>
      File "/home/simonec/virtual_envs/venv_aic/lib/python3.12/site-packages/jinja2/environment.py", line 812, in compile_expression
        self.handle_exception(source=source)
      File "/home/simonec/virtual_envs/venv_aic/lib/python3.12/site-packages/jinja2/environment.py", line 942, in handle_exception
        raise rewrite_traceback_stack(source=source)
      File "<unknown>", line 1, in template
    jinja2.exceptions.TemplateSyntaxError: expected token ',', got 'for'
  

  • Intention: All computation logic should be centralized in .rule files instead of scattering in jinja templates, ensuring consistency and maintainability.
• Set cuda_graph_batch_sizes to a logarithmic series of graphs than capturing all graph sizes with step 1. This saves excessive resource consumption at runtime.
  • Ref: https://forums.developer.nvidia.com/t/best-practices-for-optimizing-pytorch-models-with-cuda-graphs/351177

Where should the reviewer start?

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• closes GitHub issue: #xxx

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[jasonqinzhou]

This part is pending in discussion in this PR: #245.
We need a meeting to finalize this fix.