support --no-enable-prefix-caching

vllm/config/__init__.py

@@ -470,6 +470,9 @@ class ModelConfig:
     logits_processors: Optional[list[Union[str, type[LogitsProcessor]]]] = None
     """One or more logits processors' fully-qualified class names or class
     definitions"""
+    enable_chunked_prefill: Optional[bool] = None
+    """If True, prefill requests can be chunked based
+    on the remaining max_num_batched_tokens."""
 
     def compute_hash(self) -> str:
         """
@@ -500,6 +503,7 @@ class ModelConfig:
         factors.append(self.rope_theta)
         # hf_config can control how the model looks!
         factors.append(self.hf_config.to_json_string())
+        factors.append(self.enable_chunked_prefill)
         str_factors = str(factors)
         assert_hashable(str_factors)
         return hashlib.sha256(str(factors).encode()).hexdigest()

vllm/engine/arg_utils.py

@@ -1590,6 +1590,9 @@ class EngineArgs:
         # For pooling tasks the default is False
         if model_config.runner_type != "pooling":
             self.enable_chunked_prefill = True
+            if model_config.enable_chunked_prefill is not None and \
+                model_config.enable_chunked_prefill is False:
+                self.enable_chunked_prefill = False
             if self.enable_prefix_caching is None:
                 self.enable_prefix_caching = True
         else:
@@ -1602,6 +1605,10 @@ class EngineArgs:
 
             action = "Enabling" if \
                 incremental_prefill_supported else "Disabling"
+                
+            if model_config.enable_chunked_prefill is not None and \
+                model_config.enable_chunked_prefill is False:
+                self.enable_chunked_prefill = False
 
             if self.enable_chunked_prefill is None:
                 self.enable_chunked_prefill = incremental_prefill_supported