feat: add CompilerConfig dataclass for structured compiler options

[AoyuQC]

Replace string-based compiler arguments with a type-safe CompilerConfig dataclass that provides discoverability and easy customization.

Changes:

• Add CompilerConfig dataclass with fields for common neuronx-cc options (lnc, model_type, auto_cast, enable_mixed_precision_accumulation, etc.)
• Add factory methods for_nkipy() and for_nki() with appropriate defaults
• Add get_default_compiler_args() helper to inspect default settings
• Add compiler_config parameter to @baremetal_jit, baremetal_run_traced_kernel, and DeviceKernel.compile_and_load()
• Export CompilerConfig and get_default_compiler_args from nkipy.runtime
• Add tutorial section demonstrating CompilerConfig usage

Backward compatible: legacy additional_compiler_args parameter still works.

Issue #, if available:
N/A

Description of changes:

Replace string-based compiler arguments with a type-safe CompilerConfig dataclass that provides discoverability and easy customization.

Changes:

• Add CompilerConfig dataclass with fields for common neuronx-cc options (lnc, model_type, auto_cast, enable_mixed_precision_accumulation, etc.)
• Add factory methods for_nkipy() and for_nki() with appropriate defaults
• Add get_default_compiler_args() helper to inspect default settings
• Add compiler_config parameter to @baremetal_jit, baremetal_run_traced_kernel(), and DeviceKernel.compile_and_load()
• Export CompilerConfig and get_default_compiler_args from nkipy.runtime
• Add tutorial section demonstrating CompilerConfig usage

Usage:

  from nkipy.runtime import CompilerConfig, baremetal_jit, get_default_compiler_args                                                                                                 
                                                                                                                                                                                     
  # View default compiler args                                                                                                                                                       
  print(get_default_compiler_args())  # "--lnc 1 --internal-tensorizer-opt-level=2"                                                                                                  
                                                                                                                                                                                     
  # Use with decorator                                                                                                                                                               
  @baremetal_jit(compiler_config=CompilerConfig.for_nkipy(model_type="transformer"))                                                                                                 
  def my_kernel(x): ...                                                                                                                                                              
                                                                                                                                                                                     
  # Use with DeviceKernel                                                                                                                                                            
  kernel = DeviceKernel.compile_and_load(                                                                                                                                            
      my_func, x, w,                                                                                                                                                                 
      compiler_config=CompilerConfig.for_nkipy(                                                                                                                                      
          lnc=2,                                                                                                                                                                     
          enable_mixed_precision_accumulation=True,                                                                                                                                  
      )                                                                                                                                                                              
  )   

Backward compatible: legacy additional_compiler_args parameter still works.

Runned Tests:

• Existing unit tests pass (uv run pytest tests/unit/)
• Linting passes (uv run ruff check .)
• Manual verification of CompilerConfig functionality
• Backward compatibility verified with legacy additional_compiler_args

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[liangfu]

remove?

[AoyuQC]

done, already removed options not documented in public api, they can still be passed via extra_args or through pure string mode

[liangfu]

LGTM

[vgene]

My understanding is that the most critical ask here is to not have --lnc=1 as the hardcoded flag. Let's use this as an opportunity to remove all hardcoded flags. I find the new dataclass unnecessary since expert users can directly use command line compiler flags. Keeping the dataclass adds maintenance complexity.

[vgene]

I prefer not to change this simple tutorial. As the first tutorial, we want to limit the information exposed to the user

[vgene]

Do we want to keep these compiler specific knowledge as a class in NKIPy? People who want to change these level of details are likely to know what they are doing, and they can use "--model-type transformer" directly. Supporting all these in NKIPy means we need to keep it in sync with the compiler

[vgene]

The NKI related compilation flags should be removed now. We don't support directly compiling NKI kernels

[vgene]

All internal-tensorizer-opt-level should be dropped from the code. Since we stop supporting compiling NKI kernel directly (support through kernel invocation), all NKI kernel flags should be dropped too.