Support keyword-only parameters and default values in decorated functions

[bastianhagedorn]

Fix **kwargs handling and support keyword-only params and default values

Problem

_sanitize_configs and run_profile_session used len(sig.parameters) to count function parameters, which includes *args and **kwargs. This caused a spurious validation error when the decorated function had **kwargs in its signature. Keyword-only parameters (after *) were also unsupported since configs were passed via func(*c) which can't populate keyword-only args.

Additionally, parameters with default values could not be omitted from configs — users had to always provide values for every parameter, even those with sensible defaults.

Fix

• **kwargs/*args handling: Exclude VAR_POSITIONAL and VAR_KEYWORD parameters from all parameter counting and config validation (4 locations across core.py, extraction.py, transformation.py).
• Keyword-only parameter support: New _bind_config_to_signature helper maps config values to parameters in declaration order, passing keyword-only params as **kwargs to the function call.
• Default value support: New _pad_config_with_defaults fills in missing trailing config values with defaults from the function signature. This happens in _sanitize_configs so the rest of the pipeline always sees full-length configs.
• Transformation fixes: Value column is now explicitly cast to numeric after explode (was staying as object dtype), and groupby uses dropna=False to handle None/NaN default values in parameter columns.

Tests

• test_function_with_kwargs — **kwargs tolerated in signature
• test_function_with_keyword_only_params — def f(x, *, y) works
• test_function_with_args_and_keyword_only — def f(x, *args, y, **kwargs) works
• test_function_with_default_parameter — updated: omitting defaulted param now works
• test_default_values_omitted — def f(x, y, z=64) with 2-element config
• test_default_values_overridden — explicitly providing a defaulted param
• test_keyword_only_default_omitted — def f(x, *, y=32) with 1-element config
• test_too_few_config_args_rejected — negative: fewer args than required params
• test_too_many_config_args_rejected — negative: more args than total params

Docs

• core_concepts.rst: documents supported parameter kinds (regular, keyword-only, defaults, *args/**kwargs)
• known_issues.rst: notes that *args/**kwargs are tolerated but ignored
• _sanitize_configs docstring updated

[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.

[copy-pr-bot]

/ok to test

@bastianhagedorn, there was an error processing your request: E1

See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/1/

[bastianhagedorn]

/ok to test 19e7f56

[acollins3]

Shame we can't support *args and **kwargs. This is just because there is no way to map a config tuple to *args and **kwargs right?

Maybe we could improve this in a follow up MR. Instead of just passing configs as simple tuples, we could pass a Config object whose arguments can be fully mapped to the profiled function (including to *args and **kwargs).

For example something like this:

@nsight.analyze.kernel(configs=[
    Config(1, 2, foo=3, bar=4)
])
def benchmark(x, y, **kwargs):
    ....

The arguments passed to the Config contructor could be mapped directly to those in the profiled funciton.

And to maintain backwards compatibility, (x,y,z) can be implicitly converted to Config(x,y,z)