fix missing host/port for vllm (axolotl-ai-cloud#2543)

* fix missing host/port for vllm

* set tensor parallel size so it doesn't always default to cli override

Author: winglian

src/axolotl/cli/args.py

@@ -39,16 +39,16 @@ class TrainerCliArgs:
 class VllmServeCliArgs:
     """Dataclass with CLI arguments for `axolotl vllm-serve` command."""
 
-    tensor_parallel_size: int = field(
-        default=1,
+    tensor_parallel_size: Optional[int] = field(
+        default=None,
         metadata={"help": "Number of tensor parallel workers to use."},
     )
-    host: str = field(
-        default="0.0.0.0",  # nosec B104
+    host: Optional[str] = field(
+        default=None,  # nosec B104
         metadata={"help": "Host address to run the server on."},
     )
-    port: int = field(
-        default=8000,
+    port: Optional[int] = field(
+        default=None,
         metadata={"help": "Port to run the server on."},
     )
     gpu_memory_utilization: Optional[float] = field(

src/axolotl/core/trainers/grpo/__init__.py

@@ -40,8 +40,8 @@ def set_training_args_kwargs(cls, cfg):
 
         if trl.use_vllm:
             grpo_args_kwargs["use_vllm"] = trl.use_vllm
-            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host
-            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port
+            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host
+            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port
             if trl.vllm_server_timeout:
                 grpo_args_kwargs["vllm_server_timeout"] = trl.vllm_server_timeout
             if trl.vllm_guided_decoding_regex:

src/axolotl/utils/schemas/vllm.py

@@ -36,3 +36,11 @@ class VllmConfig(BaseModel):
         default=None,
         json_schema_extra={"description": "Enable prefix caching for VLLM"},
     )
+    host: str | None = Field(
+        default="0.0.0.0",  # nosec B104
+        json_schema_extra={"description": "Host for the vLLM server to start on"},
+    )
+    port: int | None = Field(
+        default=8000,
+        json_schema_extra={"description": "Port of the vLLM server to start on"},
+    )