[diffusion] api: add sampling parameters and model info endpoint to OpenAI API (#15071)

Co-authored-by: niehen6174 <niehen.6174@gmail.com>
Co-authored-by: Mick <mickjagger19@icloud.com>
Co-authored-by: niehen6174 <nihen6174@gmail.com>

python/sglang/multimodal_gen/docs/openai_api.md

@@ -25,6 +25,33 @@ sglang serve "${SERVER_ARGS[@]}"
 - **--model-path**: Path to the model or model ID.
 - **--port**: HTTP port to listen on (default: `30000`).
 
+#### Get Model Information
+
+**Endpoint:** `GET /models`
+
+Returns information about the model served by this server, including model path, task type, pipeline configuration, and precision settings.
+
+**Curl Example:**
+
+```bash
+curl -sS -X GET "http://localhost:30010/models"
+```
+
+**Response Example:**
+
+```json
+{
+  "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+  "task_type": "T2V",
+  "workload_type": "serving",
+  "pipeline_name": "wan_pipeline",
+  "pipeline_class": "WanPipeline",
+  "num_gpus": 4,
+  "dit_precision": "bf16",
+  "vae_precision": "fp16"
+}
+```
+
 ---
 
 ## Endpoints

python/sglang/multimodal_gen/runtime/entrypoints/http_server.py

@@ -3,7 +3,7 @@
 import asyncio
 from contextlib import asynccontextmanager
 
-from fastapi import APIRouter, FastAPI
+from fastapi import APIRouter, FastAPI, Request
 
 from sglang.multimodal_gen.runtime.entrypoints.openai import image_api, video_api
 from sglang.multimodal_gen.runtime.server_args import ServerArgs
@@ -40,6 +40,30 @@ async def health():
     return {"status": "ok"}
 
 
+@health_router.get("/models")
+async def get_models(request: Request):
+    """Get information about the model served by this server."""
+    from sglang.multimodal_gen.registry import get_model_info
+
+    server_args: ServerArgs = request.app.state.server_args
+    model_info = get_model_info(server_args.model_path)
+
+    response = {
+        "model_path": server_args.model_path,
+        "num_gpus": server_args.num_gpus,
+        "task_type": server_args.pipeline_config.task_type.name,
+        "dit_precision": server_args.pipeline_config.dit_precision,
+        "vae_precision": server_args.pipeline_config.vae_precision,
+        "workload_type": server_args.workload_type.value,
+    }
+
+    if model_info:
+        response["pipeline_name"] = model_info.pipeline_cls.pipeline_name
+        response["pipeline_class"] = model_info.pipeline_cls.__name__
+
+    return response
+
+
 @health_router.get("/health_generate")
 async def health_generate():
     # TODO : health generate endpoint

python/sglang/multimodal_gen/runtime/entrypoints/openai/image_api.py

@@ -55,6 +55,10 @@ def _build_sampling_params_from_request(
     image_path: Optional[str] = None,
     seed: Optional[int] = None,
     generator_device: Optional[str] = None,
+    negative_prompt: Optional[str] = None,
+    guidance_scale: Optional[float] = None,
+    num_inference_steps: Optional[int] = None,
+    enable_teacache: Optional[bool] = None,
 ) -> SamplingParams:
     if size is None:
         width, height = None, None
@@ -77,6 +81,10 @@ def _build_sampling_params_from_request(
         output_file_name=f"{request_id}.{ext}",
         seed=seed,
         generator_device=generator_device,
+        guidance_scale=guidance_scale,
+        num_inference_steps=num_inference_steps,
+        enable_teacache=enable_teacache,
+        **({"negative_prompt": negative_prompt} if negative_prompt is not None else {}),
     )
     return sampling_params
 
@@ -118,6 +126,10 @@ async def generations(
         background=request.background,
         seed=request.seed,
         generator_device=request.generator_device,
+        negative_prompt=request.negative_prompt,
+        guidance_scale=request.guidance_scale,
+        num_inference_steps=request.num_inference_steps,
+        enable_teacache=request.enable_teacache,
     )
     batch = prepare_request(
         server_args=get_global_server_args(),
@@ -169,6 +181,10 @@ async def edits(
     seed: Optional[int] = Form(1024),
     generator_device: Optional[str] = Form("cuda"),
     user: Optional[str] = Form(None),
+    negative_prompt: Optional[str] = Form(None),
+    guidance_scale: Optional[float] = Form(None),
+    num_inference_steps: Optional[int] = Form(None),
+    enable_teacache: Optional[bool] = Form(False),
 ):
     request_id = generate_request_id()
     # Resolve images from either `image` or `image[]` (OpenAI SDK sends `image[]` when list is provided)
@@ -198,6 +214,10 @@ async def edits(
         image_path=input_paths,
         seed=seed,
         generator_device=generator_device,
+        negative_prompt=negative_prompt,
+        guidance_scale=guidance_scale,
+        num_inference_steps=num_inference_steps,
+        enable_teacache=enable_teacache,
     )
     batch = _build_req_from_sampling(sampling)
 

python/sglang/multimodal_gen/runtime/entrypoints/openai/protocol.py

@@ -29,6 +29,10 @@ class ImageGenerationsRequest(BaseModel):
     seed: Optional[int] = 1024
     generator_device: Optional[str] = "cuda"
     user: Optional[str] = None
+    negative_prompt: Optional[str] = None
+    guidance_scale: Optional[float] = None
+    num_inference_steps: Optional[int] = None
+    enable_teacache: Optional[bool] = False
 
 
 # Video API protocol models
@@ -62,6 +66,7 @@ class VideoGenerationsRequest(BaseModel):
     guidance_scale: Optional[float] = None
     guidance_scale_2: Optional[float] = None
     negative_prompt: Optional[str] = None
+    enable_teacache: Optional[bool] = False
 
 
 class VideoListResponse(BaseModel):

python/sglang/multimodal_gen/runtime/entrypoints/openai/video_api.py

@@ -82,6 +82,8 @@ def _build_sampling_params_from_request(
         sampling_kwargs["guidance_scale_2"] = request.guidance_scale_2
     if request.negative_prompt is not None:
         sampling_kwargs["negative_prompt"] = request.negative_prompt
+    if request.enable_teacache is not None:
+        sampling_kwargs["enable_teacache"] = request.enable_teacache
     sampling_params = SamplingParams.from_user_sampling_params_args(
         model_path=server_args.model_path,
         server_args=server_args,
@@ -139,6 +141,12 @@ async def create_video(
     size: Optional[str] = Form(None),
     fps: Optional[int] = Form(None),
     num_frames: Optional[int] = Form(None),
+    seed: Optional[int] = Form(1024),
+    generator_device: Optional[str] = Form("cuda"),
+    negative_prompt: Optional[str] = Form(None),
+    guidance_scale: Optional[float] = Form(None),
+    num_inference_steps: Optional[int] = Form(None),
+    enable_teacache: Optional[bool] = Form(False),
     extra_body: Optional[str] = Form(None),
 ):
     content_type = request.headers.get("content-type", "").lower()
@@ -180,6 +188,12 @@ async def create_video(
             size=size,
             fps=fps_val,
             num_frames=num_frames_val,
+            seed=seed,
+            generator_device=generator_device,
+            negative_prompt=negative_prompt,
+            guidance_scale=guidance_scale,
+            num_inference_steps=num_inference_steps,
+            enable_teacache=enable_teacache,
         )
     else:
         try: