Tiny refactor ModelConfig.from_server_args (sgl-project#5219)

Author: fzyzcjy

python/sglang/bench_one_batch.py

@@ -137,17 +137,7 @@ def load_model(server_args, port_args, tp_rank):
     suppress_other_loggers()
     rank_print = print if tp_rank == 0 else lambda *args, **kwargs: None
 
-    model_config = ModelConfig(
-        server_args.model_path,
-        trust_remote_code=server_args.trust_remote_code,
-        revision=server_args.revision,
-        context_length=server_args.context_length,
-        model_override_args=server_args.json_model_override_args,
-        is_embedding=server_args.is_embedding,
-        enable_multimodal=server_args.enable_multimodal,
-        dtype=server_args.dtype,
-        quantization=server_args.quantization,
-    )
+    model_config = ModelConfig.from_server_args(server_args)
     model_runner = ModelRunner(
         model_config=model_config,
         mem_fraction_static=server_args.mem_fraction_static,

python/sglang/srt/configs/model_config.py

@@ -24,6 +24,7 @@
 
 from sglang.srt.hf_transformers_utils import get_config, get_context_length
 from sglang.srt.layers.quantization import QUANTIZATION_METHODS
+from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import get_bool_env_var, is_hip
 
 logger = logging.getLogger(__name__)
@@ -210,6 +211,21 @@ def __init__(
         self.hf_eos_token_id = self.get_hf_eos_token_id()
         self.image_token_id = getattr(self.hf_config, "image_token_id", None)
 
+    @staticmethod
+    def from_server_args(server_args: ServerArgs, model_path: str = None, **kwargs):
+        return ModelConfig(
+            model_path=model_path or server_args.model_path,
+            trust_remote_code=server_args.trust_remote_code,
+            revision=server_args.revision,
+            context_length=server_args.context_length,
+            model_override_args=server_args.json_model_override_args,
+            is_embedding=server_args.is_embedding,
+            enable_multimodal=server_args.enable_multimodal,
+            dtype=server_args.dtype,
+            quantization=server_args.quantization,
+            **kwargs,
+        )
+
     # adapted from https://github.com/vllm-project/vllm/blob/main/vllm/config.py#L289
     def get_total_num_kv_heads(self) -> int:
         """Returns the total number of KV heads."""

python/sglang/srt/managers/scheduler.py

@@ -455,17 +455,7 @@ def __init__(
     def init_tokenizer(self):
         server_args = self.server_args
 
-        self.model_config = ModelConfig(
-            server_args.model_path,
-            trust_remote_code=server_args.trust_remote_code,
-            revision=server_args.revision,
-            context_length=server_args.context_length,
-            model_override_args=server_args.json_model_override_args,
-            is_embedding=server_args.is_embedding,
-            enable_multimodal=server_args.enable_multimodal,
-            dtype=server_args.dtype,
-            quantization=server_args.quantization,
-        )
+        self.model_config = ModelConfig.from_server_args(server_args)
         self.is_generation = self.model_config.is_generation
 
         if server_args.skip_tokenizer_init:

python/sglang/srt/managers/tokenizer_manager.py

@@ -165,17 +165,7 @@ def __init__(
         # Read model args
         self.model_path = server_args.model_path
         self.served_model_name = server_args.served_model_name
-        self.model_config = ModelConfig(
-            server_args.model_path,
-            trust_remote_code=server_args.trust_remote_code,
-            revision=server_args.revision,
-            context_length=server_args.context_length,
-            model_override_args=server_args.json_model_override_args,
-            is_embedding=server_args.is_embedding,
-            enable_multimodal=server_args.enable_multimodal,
-            dtype=server_args.dtype,
-            quantization=server_args.quantization,
-        )
+        self.model_config = ModelConfig.from_server_args(server_args)
 
         self.is_generation = self.model_config.is_generation
         self.is_image_gen = self.model_config.is_image_gen

python/sglang/srt/managers/tp_worker.py

@@ -65,20 +65,13 @@ def __init__(
         self.pp_rank = pp_rank
 
         # Init model and tokenizer
-        self.model_config = ModelConfig(
-            (
+        self.model_config = ModelConfig.from_server_args(
+            server_args,
+            model_path=(
                 server_args.model_path
                 if not is_draft_worker
                 else server_args.speculative_draft_model_path
             ),
-            trust_remote_code=server_args.trust_remote_code,
-            revision=server_args.revision,
-            context_length=server_args.context_length,
-            model_override_args=server_args.json_model_override_args,
-            is_embedding=server_args.is_embedding,
-            enable_multimodal=server_args.enable_multimodal,
-            dtype=server_args.dtype,
-            quantization=server_args.quantization,
             is_draft_model=is_draft_worker,
         )
 

test/srt/test_gptqmodel_dynamic.py

@@ -43,16 +43,7 @@ def check_quant_method(model_path: str, use_marlin_kernel: bool):
         pass
 
     server_args = ServerArgs(model_path=model_path, dtype=torch.float16)
-    model_config = ModelConfig(
-        server_args.model_path,
-        trust_remote_code=server_args.trust_remote_code,
-        revision=server_args.revision,
-        context_length=server_args.context_length,
-        model_override_args=server_args.json_model_override_args,
-        is_embedding=server_args.is_embedding,
-        dtype=server_args.dtype,
-        quantization=server_args.quantization,
-    )
+    model_config = ModelConfig.from_server_args(server_args)
 
     load_config = LoadConfig()
     device_config = DeviceConfig("cuda")