Br a5 dev 1110

[w630497878]

What this PR does / why we need it?

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.11.0
• vLLM main: vllm-project/vllm@2918c1b

[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 support for a new hardware target, designated as 'A5', across various components of the vLLM Ascend backend. While the changes are extensive, there are numerous critical issues, including typos, undefined variables, incorrect API usage, and logical errors that will prevent the code from running. These issues are present in almost every modified file and need to be addressed to ensure correctness and functionality on the new hardware path. My review provides specific comments and suggestions to fix these critical problems.

[gemini-code-assist]

There is a typo in the return statement. retunr should be return.

Suggested change

	retunr None
	return None

[gemini-code-assist]

The variable device_ip_ is not defined in this scope, which will cause a NameError. It is defined in the else block but not in the if is_A5() block. You need to extract device_ip from device_info.

Suggested change

	device_ip=device_ip_,
	device_ip=device_info["device_ip"],

[gemini-code-assist]

This code block contains several critical errors that will cause it to fail at runtime:

• Line 714: Typo hidden_szie should be hidden_size.
• Line 715 & 727: Typo batch_szie should be batch_size.
• Line 716: hidden_size is used but was defined with a typo as hidden_szie.
• Line 720: Typo torch_nup should be torch_npu.
• Line 723: The value variable is not defined in this scope. It should likely be derived from self.value_cache similar to how key is derived from self.key_cache.
• Line 724: attn_metadata does not have an attribute seq_len. It should probably be seq_lens.
• Line 730: Typo inpt_layout should be input_layout.

                if is_A5(): 
                    batch_size = attn_metadata.query_lens.shape[0]
                    hidden_size = self.num_heads * self.head_size
                    query = query[:batch_size]
                    query = query.view(batch_size, 1, hidden_size)
                    block_size = self.key_cache.shape[1]
                    key = self.key_cache.flatten(2, 3).contiguous()
                    value = self.value_cache.flatten(2, 3).contiguous()
                    ori_output = output
                    output, _ = torch_npu.npu_fused_infer_attention_score_v2(
                        query=query,
                        key=key,
                        value=value,
                        actual_seq_kvlen=attn_metadata.seq_lens,
                        num_query_heads=self.num_heads,
                        num_key_value_heads=self.num_kv_heads,
                        block_table=attn_metadata.block_tables[:batch_size],
                        block_size=block_size,
                        softmax_scale=self.scale,
                        input_layout="BSH"
                    )

[gemini-code-assist]

There are a couple of issues in this block:

1. On line 805, there is an indentation error for block_size, which will cause a SyntaxError.
2. On line 807, imput_layout is a typo and should be input_layout.

        if is_A5():
            output, _ = torch_npu.npu_fused_infer_attention_score_v2(
                query=query,
                key=self.key_cache.flatten(2,3).contiguous(),
                value=self.value_cache.flatten(2,3).contiguous(),
                atten_mask=attn_metadata.attn_mask,
                actual_seq_qlen=attn_metadata.actual_seq_lengths_q,
                actual_seq_kvlen=attn_metadata.seq_lens_list,
                num_query_heads=self.num_heads,
                num_key_value_heads=self.num_kv_heads,
                block_table=attn_metadata.block_tables[:attn_metadata.query_lens.shape[0]],
                block_size=self.key_cache.shape[1],
                softmax_scale=self.scale,
                input_layout="TND"
            )

[gemini-code-assist]

There are two critical issues here:

1. The is_A5 function is not imported, which will lead to a NameError.
2. The function forward_oot is expected to return a tuple (query, key). However, this if block has an early return without a value, which implicitly returns None. This will cause a TypeError when the caller tries to unpack the result. The comment indicates this is a placeholder, but it should at least return the original query and key to avoid crashing.

Suggested change

	if is_A5(): # A5不支持npu_mrope算子，这里需要使用小算子替换
	return
	if is_A5(): # A5不支持npu_mrope算子，这里需要使用小算子替换
	return query, key

[gemini-code-assist]

The function is_A5 is used here but it is not imported in this file. This will cause a NameError at runtime.

[gemini-code-assist]

This block has several critical issues:

1. The is_A5 function is not imported, which will cause a NameError.
2. On line 997, there is a typo mas_seq_len which should be max_seq_len.
3. On line 998, there is a typo self.block_szie which should be self.block_size.
4. On line 1000, seq_lens is reassigned but not used afterwards in this block, making the operation ineffective.

Suggested change

	if is_A5():
	mas_seq_len = max(seq_lens, default=0)
	max_seq_len = (max_seq_len + self.block_szie - 1) // self.block_size * self.block_size
	new_element = torch.tensor([max_seq_len])
	seq_lens = torch.cat([seq_lens, new_element], dim =0)
	return self.attn_mask_builder.get_attn_mask(max_seq_len, self.dtype, self.device).to(torch.bool)
	if is_A5():
	max_seq_len = max(seq_lens.tolist(), default=0)
	max_seq_len = (max_seq_len + self.block_size - 1) // self.block_size * self.block_size
	return self.attn_mask_builder.get_attn_mask(max_seq_len, self.dtype, self.device).to(torch.bool)

[gemini-code-assist]

The variable model_type is used here but it is not defined in the scope of the _select_moe_comm_method function. It needs to be passed as an argument to the function.

[github-actions]

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

[linfeng-yuan]

Please remove chinese characters and same below.

[github-actions]

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