[BugFix]Fix incorrect get_current_vllm_config (#5121)

### What this PR does / why we need it?
This PR fixes some incorrect `get_current_vllm_config` calling, which
creates empty vllm_config instead.

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: Angazenn <supperccell@163.com>

vllm_ascend/ascend_config.py

@@ -154,8 +154,7 @@ class AscendConfig:
         # npu_fused_infer_attention_score in some cases. We allow to execute
         # _npu_paged_attention in this cases. This should be removed once
         # npu_fused_infer_attention_score performs better on all scenarios.
-        self.pa_shape_list = additional_config.get("pa_shape_list",
-                                                   [1, 2, 3, 4])
+        self.pa_shape_list = additional_config.get("pa_shape_list", [])
 
         kv_cfg = vllm_config.kv_transfer_config
         if kv_cfg is not None and not getattr(kv_cfg, "_engine_id_patched",

vllm_ascend/attention/attention_v1.py

@@ -367,6 +367,7 @@ class AscendAttentionBackendImpl(AttentionImpl):
         kv_sharing_target_layer_name: Optional[str],
         **kwargs,
     ) -> None:
+        self.vllm_config = get_current_vllm_config()
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)
@@ -723,7 +724,7 @@ class AscendAttentionBackendImpl(AttentionImpl):
     ):
         num_tokens = query.shape[0]
         if (attn_metadata.attn_state == AscendAttentionState.DecodeOnly
-                and using_paged_attention(num_tokens)
+                and using_paged_attention(num_tokens, self.vllm_config)
                 and self.sliding_window is None):
             output = self.forward_paged_attention(query, attn_metadata, output)
         else:

vllm_ascend/attention/utils.py

@@ -1,10 +1,9 @@
 from dataclasses import dataclass
-from functools import lru_cache
 from typing import Any, List, Optional
 
 import torch
 import torch.nn.functional as F
-from vllm.config import get_current_vllm_config
+from vllm.config import VllmConfig
 from vllm.distributed.kv_transfer import (get_kv_transfer_group,
                                           has_kv_transfer_group,
                                           is_v1_kv_transfer_group)
@@ -14,9 +13,7 @@ from vllm_ascend.utils import (AscendDeviceType, get_ascend_config,
                                get_ascend_device_type)
 
 
-@lru_cache
-def using_paged_attention(runtime_shape: int) -> bool:
-    vllm_config = get_current_vllm_config()
+def using_paged_attention(runtime_shape: int, vllm_config: VllmConfig) -> bool:
     if vllm_config.speculative_config is not None:
         return False
     if get_ascend_device_type() == AscendDeviceType.A5:

vllm_ascend/compilation/acl_graph.py

@@ -296,8 +296,9 @@ def _update_attn_fia_params(update_stream, forward_context, runtime_shape):
             event.record(update_stream)
 
 
-def update_attn_params(update_stream, forward_context, runtime_shape):
-    if using_paged_attention(runtime_shape):
+def update_attn_params(update_stream, forward_context, runtime_shape,
+                       vllm_config):
+    if using_paged_attention(runtime_shape, vllm_config):
         _update_attn_pa_params(update_stream, forward_context, runtime_shape)
     else:
         _update_attn_fia_params(update_stream, forward_context, runtime_shape)

vllm_ascend/compilation/passes/qknorm_rope_fusion_pass.py

@@ -23,8 +23,7 @@ from torch._inductor.pattern_matcher import (PatternMatcherPass,
                                              PatternPrettyPrinter)
 from vllm.attention.layer import Attention
 from vllm.compilation.vllm_inductor_pass import VllmInductorPass
-from vllm.config import (VllmConfig, get_current_vllm_config,
-                         get_layers_from_vllm_config)
+from vllm.config import VllmConfig, get_layers_from_vllm_config
 
 
 class QKNormRopeFusionPattern:
@@ -42,7 +41,6 @@ class QKNormRopeFusionPattern:
         self.q_size = self.num_heads * self.head_dim
         self.kv_size = self.num_kv_heads * self.head_dim
         self.eps = eps
-        vllm_config = get_current_vllm_config()
         self.device = vllm_config.device_config.device if vllm_config.device_config else None
 
     def get_inputs(self):

vllm_ascend/worker/model_runner_v1.py

@@ -1165,7 +1165,8 @@ class NPUModelRunner(GPUModelRunner):
                                                maybe_padded_num_tokens)
                 else:
                     update_attn_params(self.update_stream, forward_context,
-                                       maybe_padded_num_tokens)
+                                       maybe_padded_num_tokens,
+                                       self.vllm_config)
 
         if get_forward_context().sp_enabled and not isinstance(
                 hidden_states, IntermediateTensors):
@@ -1957,7 +1958,7 @@ class NPUModelRunner(GPUModelRunner):
                                                positions.shape[0])
                 else:
                     update_attn_params(self.update_stream, forward_context,
-                                       num_tokens)
+                                       num_tokens, self.vllm_config)
 
         if self.drafter and self.drafter.name == SpecDcodeType.EAGLE3:
             hidden_states, _ = hidden_states