support --no-enable-chunked-prefill of v1

vllm/config.py

@@ -418,6 +418,9 @@ class ModelConfig:
     - "transformers" will use the Transformers model implementation."""
     override_attention_dtype: Optional[str] = None
     """Override dtype for attention"""
+    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:
         """
@@ -448,6 +451,7 @@ 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

@@ -1004,6 +1004,7 @@ class EngineArgs:
             enable_sleep_mode=self.enable_sleep_mode,
             model_impl=self.model_impl,
             override_attention_dtype=self.override_attention_dtype,
+            enable_chunked_prefill=self.enable_chunked_prefill,
         )
 
     def create_load_config(self) -> LoadConfig:
@@ -1593,6 +1594,9 @@ 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:
@@ -1607,6 +1611,10 @@ 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
                 logger.info("(%s) chunked prefill by default", action)