[Feat.]: Support 310P device run qwen2.5/3 dense and qwen2.5vl models

[Tflowers-0129]

What this PR does / why we need it?

Does this PR introduce any user-facing change?

How was this patch tested?

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

[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 the 310P device, enabling it to run specific qwen models. The changes are well-structured, primarily introducing 310P-specific implementations within a new _310p directory and using conditional logic to activate them. My review identified a critical issue in the attention implementation where some attention states are unhandled, which could lead to incorrect behavior. Additionally, I've pointed out a high-severity issue concerning a hardcoded sequence length that might impact model compatibility. Addressing these points will improve the robustness and correctness of the 310P support.

[gemini-code-assist]

The forward_impl method only handles DecodeOnly and PrefillNoCache attention states. Other states from the AscendAttentionState enum, such as PrefillCacheHit, ChunkedPrefill, and SpecDecoding, are not handled. This will cause incorrect behavior when these attention states occur, as the function will return the output tensor without processing it. This is a critical bug that needs to be addressed. Additionally, the use of two separate if statements is incorrect; an if/elif structure should be used to ensure only one path is taken.

Suggested change

	def forward_impl(self, query, key, value, kv_cache, attn_metadata, output):
	if attn_metadata.attn_state == AscendAttentionState.DecodeOnly:
	output = self.forward_paged_attention(query, attn_metadata, output)
	if attn_metadata.attn_state == AscendAttentionState.PrefillNoCache:
	num_tokens = query.shape[0]
	q = query[:num_tokens]
	k = key[:num_tokens]
	v = value[:num_tokens]
	out = self._forward_prefill_310p_fallback(q, k, v, attn_metadata, output)
	output[:num_tokens] = out
	return output
	def forward_impl(self, query, key, value, kv_cache, attn_metadata, output):
	if attn_metadata.attn_state == AscendAttentionState.DecodeOnly:
	output = self.forward_paged_attention(query, attn_metadata, output)
	elif attn_metadata.attn_state == AscendAttentionState.PrefillNoCache:
	num_tokens = query.shape[0]
	q = query[:num_tokens]
	k = key[:num_tokens]
	v = value[:num_tokens]
	out = self._forward_prefill_310p_fallback(q, k, v, attn_metadata, output)
	output[:num_tokens] = out
	else:
	raise NotImplementedError(
	f"Attention state {attn_metadata.attn_state} is not yet supported in AscendAttentionBackendImpl310."
	)
	return output

[gemini-code-assist]

The methods get_splitfuse_attn_mask and get_attention_mask use a hardcoded sequence length of 2048. This could lead to incorrect behavior or errors for models with a different maximum sequence length. It would be more robust to derive this value from the model_config, for example, by using model_config.max_model_len.

Suggested change

	def get_splitfuse_attn_mask(self) -> torch.Tensor:
	return self._get_fp16_mask(2048)
	def get_attention_mask(self, model_config) -> torch.Tensor:
	if getattr(model_config, "runner_type", None) == "pooling":
	return self._base.get_attn_mask(2048, torch.bool)
	return self.get_splitfuse_attn_mask()
	def get_splitfuse_attn_mask(self, max_seq_len: int) -> torch.Tensor:
	return self._get_fp16_mask(max_seq_len)
	def get_attention_mask(self, model_config) -> torch.Tensor:
	# Fallback to 2048 if max_model_len is not available.
	max_seq_len = getattr(model_config, "max_model_len", 2048)
	if getattr(model_config, "runner_type", None) == "pooling":
	return self._base.get_attn_mask(max_seq_len, torch.bool)
	return self.get_splitfuse_attn_mask(max_seq_len)