[Misc] Clean Up `EngineArgs.create_engine_config` (#13734)

Signed-off-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com>

vllm/config.py

@@ -1124,6 +1124,10 @@ class CacheConfig:
         return {key: str(value) for key, value in self.__dict__.items()}
 
     def _verify_args(self) -> None:
+        if self.cpu_offload_gb < 0:
+            raise ValueError("CPU offload space must be non-negative"
+                             f", but got {self.cpu_offload_gb}")
+
         if self.gpu_memory_utilization > 1.0:
             raise ValueError(
                 "GPU memory utilization must be less than 1.0. Got "

vllm/engine/arg_utils.py

@@ -1062,6 +1062,17 @@ class EngineArgs:
         return engine_args
 
     def create_model_config(self) -> ModelConfig:
+        # gguf file needs a specific model loader and doesn't use hf_repo
+        if check_gguf_file(self.model):
+            self.quantization = self.load_format = "gguf"
+
+        # NOTE: This is to allow model loading from S3 in CI
+        if (not isinstance(self, AsyncEngineArgs) and envs.VLLM_CI_USE_S3
+                and self.model in MODELS_ON_S3
+                and self.load_format == LoadFormat.AUTO):  # noqa: E501
+            self.model = f"{MODEL_WEIGHTS_S3_BUCKET}/{self.model}"
+            self.load_format = LoadFormat.RUNAI_STREAMER
+
         return ModelConfig(
             model=self.model,
             task=self.task,
@@ -1101,26 +1112,6 @@ class EngineArgs:
         )
 
     def create_load_config(self) -> LoadConfig:
-        return LoadConfig(
-            load_format=self.load_format,
-            download_dir=self.download_dir,
-            model_loader_extra_config=self.model_loader_extra_config,
-            ignore_patterns=self.ignore_patterns,
-        )
-
-    def create_engine_config(self,
-                             usage_context: Optional[UsageContext] = None
-                             ) -> VllmConfig:
-        from vllm.platforms import current_platform
-        current_platform.pre_register_and_update()
-
-        if envs.VLLM_USE_V1:
-            self._override_v1_engine_args(usage_context)
-
-        # gguf file needs a specific model loader and doesn't use hf_repo
-        if check_gguf_file(self.model):
-            self.quantization = self.load_format = "gguf"
-
         # bitsandbytes quantization needs a specific model loader
         # so we make sure the quant method and the load format are consistent
         if (self.quantization == "bitsandbytes" or
@@ -1137,19 +1128,23 @@ class EngineArgs:
                 "BitsAndBytes load format and QLoRA adapter only support "
                 f"'bitsandbytes' quantization, but got {self.quantization}")
 
-        assert self.cpu_offload_gb >= 0, (
-            "CPU offload space must be non-negative"
-            f", but got {self.cpu_offload_gb}")
+        return LoadConfig(
+            load_format=self.load_format,
+            download_dir=self.download_dir,
+            model_loader_extra_config=self.model_loader_extra_config,
+            ignore_patterns=self.ignore_patterns,
+        )
 
-        device_config = DeviceConfig(device=self.device)
+    def create_engine_config(self,
+                             usage_context: Optional[UsageContext] = None
+                             ) -> VllmConfig:
+        from vllm.platforms import current_platform
+        current_platform.pre_register_and_update()
 
-        # NOTE: This is to allow model loading from S3 in CI
-        if (not isinstance(self, AsyncEngineArgs) and envs.VLLM_CI_USE_S3
-                and self.model in MODELS_ON_S3
-                and self.load_format == LoadFormat.AUTO):  # noqa: E501
-            self.model = f"{MODEL_WEIGHTS_S3_BUCKET}/{self.model}"
-            self.load_format = LoadFormat.RUNAI_STREAMER
+        if envs.VLLM_USE_V1:
+            self._override_v1_engine_args(usage_context)
 
+        device_config = DeviceConfig(device=self.device)
         model_config = self.create_model_config()
 
         if (model_config.is_multimodal_model and not envs.VLLM_USE_V1
@@ -1281,16 +1276,6 @@ class EngineArgs:
             if speculative_config is None \
             else speculative_config.num_lookahead_slots
 
-        if not self.use_v2_block_manager:
-            logger.warning(
-                "[DEPRECATED] Block manager v1 has been removed, "
-                "and setting --use-v2-block-manager to True or False has "
-                "no effect on vLLM behavior. Please remove "
-                "--use-v2-block-manager in your engine argument. "
-                "If your use case is not supported by "
-                "SelfAttnBlockSpaceManager (i.e. block manager v2),"
-                " please file an issue with detailed information.")
-
         scheduler_config = SchedulerConfig(
             runner_type=model_config.runner_type,
             max_num_batched_tokens=self.max_num_batched_tokens,