support triton of mrope

[shiyuan680]

What this PR does / why we need it?

this pr support use triton mrope like cuda_forward, which performance is equal to ascendc ops
this triton ops should use cann 8.5.0

Does this PR introduce any user-facing change?

How was this patch tested?

test in qwen3-vl-235b acc textvqa
native 81.82
npu triton 81.58
cuda triton 81.52

• vLLM version: v0.13.0
• vLLM main: vllm-project/vllm@2f4e654

[gemini-code-assist]

Code Review

This pull request adds support for Triton-based mRoPE, including a new end-to-end test, updates to unit tests, and the core implementation in AscendMRotaryEmbedding. While the feature is a good addition, I've found several critical issues that must be addressed. The new e2e test contains bugs in its reference implementation and test logic, which invalidates its results. More importantly, the AscendMRotaryEmbedding implementation has a critical flaw in its caching logic that will lead to incorrect computations by using stale data. I've also noted a minor issue in the unit test mocks. Please see the detailed comments for specifics and suggestions for fixes.

[gemini-code-assist]

There appears to be a copy-paste error here. sin_row is being assigned values from token_cos instead of token_sin. This makes the reference implementation incorrect, and the test will validate the Triton kernel against wrong values.

Suggested change

	sin_row[t_end:h_end] = token_cos[t_end:h_end, 1]
	sin_row[t_end:h_end] = token_sin[t_end:h_end, 1]

[gemini-code-assist]

The arguments passed to pytorch_forward_native are incorrect. The head_size parameter is receiving num_tokens, and the rotary_dim parameter is receiving head_size. The actual rotary_dim from the test parameters is not being used at all. This makes the test logic flawed and it will not correctly validate the Triton kernel's output.

Suggested change

	q_gold, k_gold = pytorch_forward_native(q_gold,
	k_gold,
	cos,
	sin,
	mrope_section,
	num_tokens,
	head_size,
	True)
	q_gold, k_gold = pytorch_forward_native(q_gold,
	k_gold,
	cos,
	sin,
	mrope_section,
	head_size,
	rotary_dim,
	True)

[gemini-code-assist]

These attributes are used in forward_triton to cache cos and sin values. However, this caching is stateful and incorrect because the values depend on positions, which can change between calls. This will lead to using stale cached data. These attributes should be removed, and the caching logic in forward_triton should be corrected to be stateless within the forward pass.

[gemini-code-assist]

The caching logic for cos and sin is incorrect. These tensors depend on positions, which can change on each call to forward_triton. Caching them as instance attributes (self.cos, self.sin) will cause subsequent calls with different positions to use stale values, leading to incorrect results. The cos and sin tensors should be computed on every call and not stored on self.

        cos_sin = self.cos_sin_cache[positions]  # type: ignore
        cos, sin = cos_sin.chunk(2, dim=-1)
        query_shape = query.shape
        key_shape = key.shape

        assert self.mrope_section

        q, k = triton_mrope(
            query,
            key,
            cos.contiguous(),
            sin.contiguous(),
            self.mrope_section,
            self.head_size,
            self.rotary_dim,
            self.mrope_interleaved,
        )

[gemini-code-assist]

This patch for vllm.triton_utils.HAS_TRITON is both redundant and incorrect. It's redundant because HAS_TRITON is already patched on line 457. It's incorrect because it uses return_value=True, which is meant for patching callables, but HAS_TRITON is a boolean constant. Please remove this line.

[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.

[weijinqian0]

remove init