[5/N][refactor]add torchair rotary ops

[hust17yixuan]

What this PR does / why we need it?

Move torchair related rotary ops into torchair dir to make the code clear. Next step we'll remove all torchair related code outside of torchair rotary ops.

Does this PR introduce any user-facing change?

No

How was this patch tested?

vLLM version: main
vLLM main: vllm-project/vllm@ab9f2cf

• vLLM version: v0.10.1.1
• vLLM main: vllm-project/vllm@81eea3d

[gemini-code-assist]

Code Review

This pull request refactors the torchair-related rotary embedding operations by moving them into a dedicated torchair/ops directory, which improves code organization. The changes also involve monkey-patching RotaryEmbedding and DeepseekScalingRotaryEmbedding for torchair-specific implementations. I've identified a critical bug in the rotary embedding cache resizing logic where the new sequence length was ignored. This could lead to incorrect computations for sequences longer than the pre-allocated cache size. A fix is suggested to address this issue.

[gemini-code-assist]

The seq_len parameter in __set_cos_sin_cache is ignored. At line 258, torch.arange uses self.max_position_embeddings instead of the provided seq_len to create the t tensor. This means that when this function is called from rope_forward to extend the cache for a sequence length greater than self.max_position_embeddings, the cache is not actually resized. This will lead to incorrect rotary embeddings for longer sequences.

To fix this, seq_len should be used to create the t tensor, and self.max_position_embeddings should be updated to reflect the new cache size.

Suggested change

	def __set_cos_sin_cache(self, seq_len, device, dtype):
	inv_freq = 1.0 / (self.base**(torch.arange(
	0, self.rotary_dim, 2, device=device, dtype=torch.float32) *
	(1 / self.rotary_dim)))
	self.register_buffer("inv_freq", inv_freq)
	t = torch.arange(self.max_position_embeddings,
	device=self.inv_freq.device,
	dtype=torch.float32)
	freqs = torch.einsum("i,j->ij", t, self.inv_freq)
	emb = torch.cat((freqs, freqs), dim=-1)
	self.register_buffer("cos", emb.cos().to(dtype=dtype), persistent=False)
	self.register_buffer("sin", emb.sin().to(dtype=dtype), persistent=False)
	self.embed = F.embedding
	def __set_cos_sin_cache(self, seq_len, device, dtype):
	self.max_position_embeddings = seq_len
	inv_freq = 1.0 / (self.base**(torch.arange(
	0, self.rotary_dim, 2, device=device, dtype=torch.float32) *
	(1 / self.rotary_dim)))
	self.register_buffer("inv_freq", inv_freq)
	t = torch.arange(seq_len,
	device=self.inv_freq.device,
	dtype=torch.float32)
	freqs = torch.einsum("i,j->ij", t, self.inv_freq)
	emb = torch.cat((freqs, freqs), dim=-1)
	self.register_buffer("cos", emb.cos().to(dtype=dtype), persistent=False)
	self.register_buffer("sin", emb.sin().to(dtype=dtype), persistent=False)
	self.embed = F.embedding

[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.

[hust17yixuan]

The CI failed at isort, but use pre-commit to fix locally and the CI has no err message. this PR may don 't have problems in lint, but blocked by some CI problems.

[codecov]

Codecov Report

❌ Patch coverage is 72.39264% with 90 lines in your changes missing coverage. Please review.
✅ Project coverage is 73.02%. Comparing base (3a5fc5e) to head (c4323f4).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
...m_ascend/torchair/ops/torchair_rotary_embedding.py	42.06%	84 Missing ⚠️
vllm_ascend/torchair/utils.py	14.28%	6 Missing ⚠️

❌ Your patch status has failed because the patch coverage (72.39%) is below the target coverage (80.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2559      +/-   ##
==========================================
- Coverage   73.03%   73.02%   -0.01%     
==========================================
  Files         149      151       +2     
  Lines       21515    21841     +326     
==========================================
+ Hits        15714    15950     +236     
- Misses       5801     5891      +90     

Flag	Coverage Δ
unittests	73.02% <72.39%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.