fix

Signed-off-by: wang.yuqi <[email protected]>

Author: noooop

tests/models/language/pooling/test_extract_hidden_states.py

@@ -30,6 +30,7 @@ def test_embed_models(hf_runner, vllm_runner, model: str):
 
         for n, output in zip(n_prompt_tokens, pooling_outputs):
             assert len(output.prompt_token_ids) == n
+            assert len(output.outputs.data) == n
             assert output.num_cached_tokens == 0
 
         # test enable_prefix_caching plus all pooling
@@ -40,4 +41,5 @@ def test_embed_models(hf_runner, vllm_runner, model: str):
 
         for n, output in zip(n_prompt_tokens, pooling_outputs):
             assert len(output.prompt_token_ids) == n
+            assert len(output.outputs.data) == n
             assert output.num_cached_tokens == 0

vllm/pooling_params.py

@@ -57,7 +57,7 @@ class PoolingParams(
     ## Internal use only
     task: PoolingTask | None = None
     requires_token_ids: bool = False
-    not_using_prefix_caching: bool = None
+    disable_prefix_caching: bool = None
     extra_kwargs: dict[str, Any] | None = None
     output_kind: RequestOutputKind = RequestOutputKind.FINAL_ONLY
 
@@ -123,14 +123,14 @@ def _merge_default_parameters(
             if getattr(self, k, None) is None:
                 setattr(self, k, getattr(pooler_config, k))
 
-        if self.not_using_prefix_caching is None:
-            # If enable_prefix_caching is enabled,
+        if self.disable_prefix_caching is None:
+            # If prefix caching is enabled,
             # the output of all pooling may less than n_prompt_tokens,
-            # we need to not using prefix_caching at this request.
+            # we need to disable prefix_caching at this request.
             if self.task in ["token_embed", "token_classify"]:
-                self.not_using_prefix_caching = True
+                self.disable_prefix_caching = True
             else:
-                self.not_using_prefix_caching = False
+                self.disable_prefix_caching = False
 
         self._verify_step_pooling(pooler_config, valid_parameters)
 

vllm/sampling_params.py

@@ -252,7 +252,7 @@ class SamplingParams(
     generated token can complete the sequence."""
     _bad_words_token_ids: list[list[int]] | None = None
 
-    not_using_prefix_caching: bool = None
+    disable_prefix_caching: bool = None
 
     @staticmethod
     def from_optional(
@@ -414,11 +414,11 @@ def __post_init__(self) -> None:
             self.structured_outputs = self.guided_decoding
             self.guided_decoding = None
 
-        if self.not_using_prefix_caching is None:
-            # If enable_prefix_caching is enabled,
+        if self.disable_prefix_caching is None:
+            # If prefix caching is enabled,
             # the output of prompt logprobs may less than n_prompt_tokens,
-            # we need to not using prefix_caching at this request.
-            self.not_using_prefix_caching = not self.prompt_logprobs
+            # we need to disable prefix_caching at this request.
+            self.disable_prefix_caching = not self.prompt_logprobs
 
     def _verify_args(self) -> None:
         if not isinstance(self.n, int):

vllm/v1/core/kv_cache_manager.py

@@ -193,11 +193,11 @@ def get_computed_blocks(self, request: Request) -> tuple[KVCacheBlocks, int]:
             not self.enable_caching
             or (
                 request.sampling_params is not None
-                and request.sampling_params.not_using_prefix_caching
+                and request.sampling_params.disable_prefix_caching
             )
             or (
                 request.pooling_params is not None
-                and request.pooling_params.not_using_prefix_caching
+                and request.pooling_params.disable_prefix_caching
             )
         ):
             return self.empty_kv_cache_blocks, 0