fix: fix Ascend attention metadata

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>

Author: yiz-liu

tests/e2e/singlecard/test_aclgraph.py

@@ -53,11 +53,10 @@ def test_models_with_aclgraph(
     # while running pytest
     if full_graph:
         vllm_model = LLM(model,
-                            compilation_config={
-                                "full_cuda_graph": True,
-                                "cudagraph_capture_sizes":
-                                [1, 4, 16, 64, 256]
-                            })
+                         compilation_config={
+                             "full_cuda_graph": True,
+                             "cudagraph_capture_sizes": [1, 4, 16, 64, 256]
+                         })
     else:
         vllm_model = LLM(model, max_model_len=1024)
     vllm_aclgraph_outputs = vllm_model.generate(prompts, sampling_params)

vllm_ascend/attention/attention_v1.py

@@ -32,8 +32,9 @@
 from vllm_ascend.attention.utils import \
     AscendCommonAttentionMetadata as CommonAttentionMetadata
 from vllm_ascend.ops.attention import vanilla_chunked_prefill
-from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ, aligned_16, get_graph_params, is_310p,
-                               nd_to_nz_2d, nd_to_nz_spec)
+from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ, aligned_16,
+                               get_graph_params, is_310p, nd_to_nz_2d,
+                               nd_to_nz_spec)
 from vllm_ascend.worker.npu_input_batch import InputBatch
 
 
@@ -135,7 +136,7 @@ class AscendMetadata:
     # tokens + new tokens (is None if it is a decoding).
     # (batch_size,)
     seq_lens: torch.Tensor = None
-    seq_lens_list: Optional[list[int]]
+    seq_lens_list: Optional[list[int]] = None
     query_start_loc: torch.Tensor = None
     query_lens: torch.Tensor = None
     # Maximum query length in the batch (None for decoding).
@@ -163,10 +164,7 @@ def reorder_batch(self, input_batch: "InputBatch",
                       scheduler_output: "SchedulerOutput") -> bool:
         return False
 
-    def build(self,
-              num_reqs,
-              num_actual_tokens,
-              max_query_len,
+    def build(self, num_reqs, num_actual_tokens, max_query_len,
               common_attn_metadata: CommonAttentionMetadata):
 
         block_table = self.runner.input_batch.block_table[0].get_device_tensor(
@@ -225,7 +223,6 @@ def build_dummy_metadata(self, num_actual_tokens, num_reqs,
                 num_reqs=num_reqs,
                 num_actual_tokens=num_actual_tokens,
                 max_query_len=num_scheduled_tokens.max(),
-                common_prefix_len=0,
                 common_attn_metadata=common_attn_metadata,
             )
         else:

vllm_ascend/attention/mla_v1.py

@@ -17,8 +17,6 @@
 from vllm_ascend import envs
 from vllm_ascend.ascend_config import get_ascend_config
 from vllm_ascend.attention.attention_v1 import AscendAttentionState
-from vllm_ascend.attention.utils import \
-    AscendCommonAttentionMetadata as CommonAttentionMetadata
 from vllm_ascend.multistream.base import MSAttentionMetadataSplitConfig
 from vllm_ascend.multistream.context import get_multistream_comm_context
 from vllm_ascend.multistream.ms_split import model_input_split_v1_mla_attn

vllm_ascend/worker/model_runner_v1.py

@@ -270,8 +270,6 @@ def __init__(self, vllm_config: VllmConfig, device: torch.device):
         self.query_lens = torch.zeros(self.max_num_reqs,
                                       dtype=torch.int32,
                                       device=self.device)
-        # None in the first PP rank. The rest are set after load_model.
-        self.intermediate_tensors: Optional[IntermediateTensors] = None
 
         self.uses_mrope = self.model_config.uses_mrope
         # Only relevant for models using M-RoPE (e.g, Qwen2-VL)