[Kernel] add custom op DispatchGmmCombineDecode

[GuoRen868]

What this PR does / why we need it?

add custom opapi DispatchGmmCombineDecode for A3, include kernel inpl, python Api, pytest.

vLLM version: v0.11.0
vLLM main: vllm-project/vllm@24d6314

• vLLM version: v0.12.0
• vLLM main: vllm-project/vllm@ad32e3e

[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 introduces a new custom operator DispatchGmmCombineDecode for the Ascend platform. The changes include the operator definition, kernel implementation, build scripts, and PyTorch bindings. My review has identified a few critical issues. There is a significant issue in the shell script csrc/build_aclnn.sh regarding environment variable setup which could cause silent failures. Another critical bug is in csrc/pytorch_npu_helper.hpp where tensor strides are calculated incorrectly, which will fail for non-contiguous tensors. Additionally, there's a confusing duplicated field in csrc/custom_ops/kernels/dispatch_gmm_combine_decode/op_kernel/dispatch_gmm_combine_decode_tiling.h that should be corrected to improve maintainability.

[gemini-code-assist]

The source command on this line will only affect the environment of the script's execution shell. When this script is executed, it runs in a sub-shell, and any environment variables set within it are lost when the script finishes. If the intention is to modify the environment of the calling shell, this script should be sourced (e.g., source csrc/build_aclnn.sh) rather than executed. The #!/bin/bash shebang is misleading if the script is meant to be sourced. This can lead to silent failures in the environment setup.

[gemini-code-assist]

The calculation of tensor strides is incorrect as it assumes the tensor is contiguous. This will lead to incorrect memory access and data corruption for non-contiguous tensors. You should use the tensor's actual strides and storage offset provided by PyTorch via at_tensor.strides() and at_tensor.storage_offset().

    const auto dimNum = at_tensor.dim();
    aclFormat format = ACL_FORMAT_ND;

    // 适配dispatch_gmm_combine_decode算子的weight入参
    if (acl_data_type == ACL_INT8 && dimNum == 3) {
        format = ACL_FORMAT_FRACTAL_NZ;
    }

    auto acl_tensor =
        aclCreateTensor(at_tensor.sizes().data(), dimNum, acl_data_type, at_tensor.strides().data(),
                        at_tensor.storage_offset(), format, at_tensor.sizes().data(), dimNum,
                        const_cast<void *>(at_tensor.storage().data()));

[gemini-code-assist]

There appears to be a duplicated field aivNum and a confusing comment. The struct DispatchGmmCombineDecodeInfo has aicNum with comment // aivNum and aivNum with comment // aivNum. This is likely a copy-paste error and can lead to confusion and bugs. Please clarify the purpose of each field and correct the comments. For example, aicNum should probably be for AI Core count and aivNum for AI Vector count.

Suggested change

	uint32_t aicNum; // aivNum
	uint32_t aivNum; // aivNum
	uint32_t aicNum; // aicNum
	uint32_t aivNum; // aivNum

[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.

[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.