[Misc] Clean up uesless code in attention (#1933)

Before do attention module refactor, we can do some code cleanup to make
the next step easier.

What this PR does:

1. remove uesless `common_prefix_len` for attention builder
2. remove uesless `is_only_prefill` and `num_input_tokens` in attention
metadata.
3. remove `CommonAttentionMetadata` and ues `query_start_loc` instead,
`CommonAttentionMetadata` is over designed and uesless
4. update the attention backend input parameters to keep the same as
vLLM.
5. Rename attention name to the same style with `ASCEND` prefix

- vLLM version: v0.9.2
- vLLM main:
107111a859

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

vllm_ascend/worker/eagle_proposer_v1.py

@@ -132,7 +132,6 @@ class EagleProposer:
             num_reqs=batch_size,
             num_actual_tokens=num_tokens,
             max_query_len=max_query_len,
-            common_prefix_len=0,
         )
         if self.use_cuda_graph and \
             num_tokens <= self.cudagraph_batch_sizes[-1]:

vllm_ascend/worker/model_runner_v1.py

@@ -75,8 +75,7 @@ from vllm_ascend.attention.attention_mask import AttentionMaskBuilder
 from vllm_ascend.attention.attention_v1 import (AscendAttentionState,
                                                 AscendMetadata)
 from vllm_ascend.attention.attention_v1_torchair import AscendTorchairMetadata
-from vllm_ascend.attention.mla_v1 import (AscendMLAMetadata,
-                                          CommonAttentionMetadata)
+from vllm_ascend.attention.mla_v1 import AscendMLAMetadata
 from vllm_ascend.platform import NPUPlatform
 from vllm_ascend.sample.rejection_sampler import AscendRejectionSampler
 from vllm_ascend.torchair.utils import (check_torchair_cache_exist,
@@ -694,15 +693,10 @@ class NPUModelRunner(LoRAModelRunnerMixin):
         # in the same group share the same metadata.
         for kv_cache_group_id, kv_cache_group_spec in enumerate(
                 self.kv_cache_config.kv_cache_groups):
-
-            # Prepare for cascade attention if enabled & beneficial.
-            common_prefix_len = 0
-
             attn_metadata_i = self.attn_metadata_builder.build(
                 num_reqs=num_reqs,
                 num_actual_tokens=total_num_scheduled_tokens,
                 max_query_len=max_num_scheduled_tokens,
-                common_prefix_len=common_prefix_len,
             )
             for layer_name in kv_cache_group_spec.layer_names:
                 attn_metadata[layer_name] = attn_metadata_i
@@ -1049,27 +1043,22 @@ class NPUModelRunner(LoRAModelRunnerMixin):
             extra_builder_kwargs['graph_pad_size'] = graph_pad_size
 
         if self.vllm_config.model_config.use_mla:
-            query_start_loc = self.query_start_loc[:num_reqs + 1]
-            seq_lens = self.seq_lens[:num_reqs]
-            common_attn_metadata = CommonAttentionMetadata(
-                query_start_loc=query_start_loc, seq_lens=seq_lens)
+            extra_builder_kwargs[
+                "query_start_loc"] = self.query_start_loc[:num_reqs + 1]
             attn_metadata = self.attn_metadata_builder.build(  # type: ignore
                 num_reqs=num_reqs,
                 num_actual_tokens=total_num_scheduled_tokens,
                 max_query_len=max_num_scheduled_tokens,
-                common_attn_metadata=common_attn_metadata,
-                common_prefix_len=None,
                 **extra_builder_kwargs,
             )
+            attn_metadata.num_input_tokens = num_input_tokens
         else:
             attn_metadata = self.attn_metadata_builder.build(  # type: ignore
                 num_reqs=num_reqs,
                 num_actual_tokens=total_num_scheduled_tokens,
                 max_query_len=max_num_scheduled_tokens,
-                common_prefix_len=None,
                 **extra_builder_kwargs,
             )
-        attn_metadata.num_input_tokens = num_input_tokens
 
         # Prepare input_ids
         token_indices = (positions_np +

vllm_ascend/worker/mtp_proposer_v1.py

@@ -8,7 +8,6 @@ from vllm.model_executor.model_loader.utils import (
     process_weights_after_loading, set_default_torch_dtype)
 from vllm.v1.sample.metadata import SamplingMetadata
 
-from vllm_ascend.attention.mla_v1 import CommonAttentionMetadata
 from vllm_ascend.models.deepseek_mtp import CustomDeepSeekMTP
 
 
@@ -100,11 +99,6 @@ class MtpProposer:
         query_lens = cu_num_tokens[1:] - cu_num_tokens[:-1]
         max_query_len = query_lens.max().item()
 
-        seq_lens = (target_positions[last_token_indices] + 1)
-
-        common_attn_metadata = CommonAttentionMetadata(
-            query_start_loc=cu_num_tokens, seq_lens=seq_lens)
-
         # FIXME: reorder_batch() needs to be called before build()
         # because fields of attn_metadata_builder needs to be updated.
         # However, currently reorder_batch() takes input_batch and
@@ -120,8 +114,7 @@ class MtpProposer:
             num_reqs=batch_size,
             num_actual_tokens=num_tokens,
             max_query_len=max_query_len,
-            common_prefix_len=0,
-            common_attn_metadata=common_attn_metadata,
+            query_start_loc=cu_num_tokens,
         )
 
         with set_forward_context(attn_metadata, self.vllm_config):