Refactor: extract _prepare_eager_model() from CoreML export main()

[lucylq]

Extract eager model preparation logic (dtype conversion, linear splitting,
quantization, graph breaks) into a reusable _prepare_model() helper.

This enables applying the same transformations consistently to multiple
models, which is used in multimethod when creating multiple lora adapters.

No functional change — pure refactor.

Generated with Claude.

[lucylq]

Stack from ghstack (oldest at bottom):

• [wip] Add CI test for CoreML LoRA multimethod export #18354
• Add --method argument to CoreML static LLM runner #18355
• Add LoRA multimethod export to CoreML static LLM export #18347
• Add skip_names and LoRALinear guard to replace_linear_with_split_linear #18346
• Add LoRA support to StaticAttention for split_mha=False #18345
• -> Refactor: extract _prepare_eager_model() from CoreML export main() #18343
• Update lora def #18342

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18343

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure, 2 Unrelated Failures

As of commit 108fa03 with merge base 02bad9d ():

NEW FAILURE - The following job has failed:

• pull / android / run-emulator (gh)
  The process '/usr/local/lib/android/sdk/platform-tools/adb' failed with exit code 224

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / unittest / windows / windows-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• pull / unittest-editable / windows / windows-job (gh) (trunk failure)
  ##[error]The operation was canceled.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[Copilot]

Pull request overview

Refactors the CoreML static Llama export script by extracting model-preparation steps into a reusable helper, enabling consistent application of dtype conversion, splitting, quantization, and graph-break insertion (useful for multi-method exports).

Changes:

• Added _prepare_model(model, args, float_dtype) to encapsulate dtype + transformation steps.
• Replaced inlined preparation logic in main() with a single call to _prepare_model(...).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The _prepare_model docstring says it applies splitting/quantization/graph breaks, but the function also changes dtype and sets eval mode. Consider updating the docstring to reflect the full set of side effects (and that it mutates the module in-place).

Suggested change

	"""Apply splitting, quantization, and graph breaks to a model."""
	"""
	Prepare an eager model for export by applying dtype conversion, eval mode,
	splitting, quantization, and optional graph breaks.
	This function mutates the given ``model`` in-place:
	* moves it to ``float_dtype``,
	* sets it to evaluation mode,
	* optionally splits linear layers,
	* optionally quantizes embeddings and linear layers, and
	* optionally wraps the first/last transformer blocks with graph breaks.
	The same (mutated) ``model`` instance is returned.
	"""

[Copilot]

args.embedding_quantize is parsed via .split(",") and validated with assert. Since this is user-controlled CLI input, consider doing explicit validation and raising a ValueError/argparse.ArgumentTypeError with a clear message (and avoid assert, which can be stripped with python -O).

Suggested change

	bitwidth, group_size = args.embedding_quantize.split(",")
	bitwidth = int(bitwidth)
	group_size = int(group_size)
	assert bitwidth in [4, 8], "CoreML only supports 4-bit and 8-bit quantization"
	try:
	bitwidth_str, group_size_str = args.embedding_quantize.split(",")
	except ValueError as e:
	raise ValueError(
	f"Invalid value for --embedding_quantize: {args.embedding_quantize!r}. "
	"Expected format is 'BITWIDTH,GROUP_SIZE', e.g. '4,32'."
	) from e
	try:
	bitwidth = int(bitwidth_str)
	group_size = int(group_size_str)
	except ValueError as e:
	raise ValueError(
	f"Invalid value for --embedding_quantize: {args.embedding_quantize!r}. "
	"BITWIDTH and GROUP_SIZE must be integers, e.g. '4,32'."
	) from e
	if bitwidth not in (4, 8):
	raise ValueError(
	f"Unsupported BITWIDTH {bitwidth} in --embedding_quantize. "
	"CoreML only supports 4-bit and 8-bit quantization."
	)