support FULL_AND_PIECEWISE graph mode.

[momo609]

What this PR does / why we need it?

support FULL_AND_PIECEWISE graph mode.

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.11.0rc3
• vLLM main: vllm-project/vllm@17c540a

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request adds support for a new FULL_AND_PIECEWISE graph mode. My review focuses on a few critical issues, including a duplicated test function that would prevent one test from running, typos in the new graph mode's name, and incorrect tensor conversions. I've provided suggestions to fix these issues to ensure the new functionality is correctly implemented and tested.

[gemini-code-assist]

This test function has the same name as the one above it, test_models_distributed_Qwen3_MOE_TP2_WITH_FULLGRAPH, which will cause the first test to be overwritten and not run. Please rename this function to be unique, for example, test_models_distributed_Qwen3_MOE_TP2_WITH_FULL_AND_PIECEWISE_GRAPH.

Additionally, there is a typo in the cudagraph_mode value on line 88: FULL_AND_PIECEIWSE should be FULL_AND_PIECEWISE.

Suggested change

	def test_models_distributed_Qwen3_MOE_TP2_WITH_FULLGRAPH():
	if 'HCCL_OP_EXPANSION_MODE' in os.environ:
	del os.environ['HCCL_OP_EXPANSION_MODE']
	prompts = [
	"Hello, my name is", "The president of the United States is",
	"The capital of France is", "The future of AI is"
	]
	model = "Qwen/Qwen3-30B-A3B"
	sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
	with VllmRunner(model,
	max_model_len=1024,
	tensor_parallel_size=2,
	enforce_eager=False,
	compilation_config={"cudagraph_mode":
	"FULL_AND_PIECEIWSE"}) as runner:
	def test_models_distributed_Qwen3_MOE_TP2_WITH_FULL_AND_PIECEWISE_GRAPH():
	if 'HCCL_OP_EXPANSION_MODE' in os.environ:
	del os.environ['HCCL_OP_EXPANSION_MODE']
	prompts = [
	"Hello, my name is", "The president of the United States is",
	"The capital of France is", "The future of AI is"
	]
	model = "Qwen/Qwen3-30B-A3B"
	sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
	with VllmRunner(model,
	max_model_len=1024,
	tensor_parallel_size=2,
	enforce_eager=False,
	compilation_config={"cudagraph_mode":
	"FULL_AND_PIECEWISE"}) as runner:

[gemini-code-assist]

There are a few issues in this block:

1. There's a typo in CUDAGraphMode.FULL_AND_PIECEIWSE; it should likely be CUDAGraphMode.FULL_AND_PIECEWISE.
2. The f-string for logging is malformed and prints cudagraph_mode twice, resulting in a confusing log message.
3. The assertion message on line 239 is now outdated as it only mentions CUDAGraphMode.PIECEWISE, but the condition now also accepts FULL_AND_PIECEWISE.

I've provided a suggestion to fix these issues and improve readability by using an in clause.

        elif compilation_config.cudagraph_mode in (
            CUDAGraphMode.PIECEWISE, CUDAGraphMode.FULL_AND_PIECEWISE):
            logger.info(
                f"{compilation_config.cudagraph_mode.name} compilation enabled on NPU. "
                "use_inductor not supported - using only ACL Graph mode.")
            assert compilation_config.level == CompilationLevel.PIECEWISE, \
                f"When enabling {compilation_config.cudagraph_mode.name} aclgraph, please make sure compilation_config.level is PIECEWISE."

[gemini-code-assist]

Using torch.Tensor() to convert a NumPy array is discouraged. It copies the data and defaults to creating a torch.float32 tensor. Since self.query_start_loc_cpu is an IntTensor, this will cause an implicit float-to-int conversion. It's better to use torch.from_numpy() to avoid an unnecessary copy and ensure correct data types.

Suggested change

	self.query_start_loc_cpu[1:num_reqs + 1] = torch.Tensor(cu_num_tokens)
	self.query_start_loc_cpu[1:num_reqs + 1] = torch.from_numpy(cu_num_tokens)

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.