[AclGraph] Adapt aclgraph into new graph dispatcher arch

[MengqingCao]

What this PR does / why we need it?

This pr adapt aclgraph into new graph dispatcher arch in vllm
breaks form:

• [Core] Allow full cudagraph with separate attention routines and orthogonal to compilation, add support for FA2 and FlashInfer vllm#20059
• [v1] Move block_hashes from KVCacheManager to Request.block_hashes (#19728) vllm#19728
• [Spec Decode] Make propose_draft_token_ids non-blocking for lower TTFT vllm#23041
• [V1] Logits processors extensibility vllm#19912

Does this PR introduce any user-facing change?

N/A

How was this patch tested?

CI passed with existing test.

• vLLM version: v0.10.0
• vLLM main: vllm-project/vllm@4d9c619

[gemini-code-assist]

Code Review

This pull request adapts aclgraph to the new graph dispatcher architecture in vLLM. It introduces ACLGraphWrapper for Ascend NPUs and integrates it into the model runner. The changes also include refactoring InputBatch to align with the new logits processor framework. My review found a critical logic error in capture_model that prevents graph capture, and an incomplete implementation in _capture_model that misses capturing graphs for decode mode.

[gemini-code-assist]

The _capture_model function only seems to capture ACL graphs for the mixed_mode. It's missing the logic to capture graphs for decode_mode (i.e., uniform decode batches). If a decode_mode is configured, the corresponding graphs will not be captured, which could lead to performance issues or errors at runtime. The implementation should also handle cudagraph_mode.decode_mode().

            aclgraph_mode = self.compilation_config.cudagraph_mode
            if aclgraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                aclgraph_runtime_mode = aclgraph_mode.mixed_mode()

                compilation_cases = list(reversed(self.aclgraph_batch_sizes))
                self._capture_aclgraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=aclgraph_runtime_mode,
                    uniform_decode=False)

            if aclgraph_mode.decode_mode() != CUDAGraphMode.NONE:
                aclgraph_runtime_mode = aclgraph_mode.decode_mode()

                compilation_cases = list(
                    reversed(self.aclgraph_dispatcher.get_decode_graph_keys()))
                self._capture_aclgraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=aclgraph_runtime_mode,
                    uniform_decode=True)

[MengqingCao]

we don't capture decode_mode here because the full graph is not supported currently.

[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.

[github-actions]

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

[github-actions]

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