test: Add Torch AOTI Tests

[whoisj]

This change:

• Creates a new L0_torch_aoti test suit.
• Adds complex Torch AOTI model generation to qa/common/gen_qa_models.py.
• Cleans up existion AOTI model generation in qa/common/gen_qa_models.py.
• Enabled torchvision AOTI model generation in qa/common/gen_qa_model_repository.

[Copilot]

Pull request overview

Adds a new QA L0 suite to exercise PyTorch Torch AOTI models (including complex/nested I/O) and enables generation of the required QA model repository artifacts (including TorchVision AOTI ResNet50).

Changes:

• Introduces qa/L0_torch_aoti test suite (shell harness + HTTP client unittest coverage for complex/simple/torchvision AOTI models).
• Extends qa/common/gen_qa_models.py to generate complex Torch AOTI models/configs and updates Torch AOTI/TorchVision AOTI packaging/config generation.
• Enables TorchVision AOTI model generation in qa/common/gen_qa_model_repository.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 8 comments.

File	Description
qa/L0_torch_aoti/torch_aoti_infer_test.py	Adds HTTP-based unittest coverage for complex/named/index Torch AOTI models, simple dtype variants, and a TorchVision AOTI model.
qa/L0_torch_aoti/test.sh	Adds the L0 harness that stages models, starts Triton, runs the Python tests, and cleans up.
qa/common/gen_qa_models.py	Adds complex Torch AOTI generation/configs; refactors paths; updates TorchVision AOTI generation.
qa/common/gen_qa_model_repository	Enables TorchVision AOTI QA model repository generation.

---

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

[yinggeh]

Why not simply create a numpy to torch dtype map?

[whoisj]

what do you mean? This function generates sample inputs, not translate types. Sorry, just confused.

[yinggeh]

Something like

np_to_torch_dtype_dict = {
    np.int8: torch.int8,
    ...
}

...

if input_dtype in np_to_torch_dtype_dict:
    input0 = torch.zeros(input_shape, dtype=np_to_torch_dtype_dict[input_dtype], device=device)
    input1 = torch.zeros(input_shape, dtype=np_to_torch_dtype_dict[input_dtype], device=device)
else:
    print(
        f"{_color_yellow}warning: dtype {input_dtype} is unsupported; falling back to torch.int32{_color_reset}"
    )
    input0 = torch.zeros(input_shape, dtype=torch.int32, device=device)
    input1 = torch.zeros(input_shape, dtype=torch.int32, device=device)

[yinggeh]

What's the reasoning behind this change? How do you tell backend is working correctly for a zero-value output tensor?

[whoisj]

these are just sample values used by PyTorch to compile the model. These aren't used for inference.

[yinggeh]

Suggested change

	def create_torch_aoti_complex_modelconfig(
	def create_torch_aoti_complex_model_config(

[whoisj]

all other similar functions are named modelconfig not model_config, I'm just following the existing pattern.

[yinggeh]

I see. "Model config" should really be two words.

[whoisj]

agreed. I can change them all, what do you think?

[yinggeh]

For now I am fine with only changing this new method.