[diffusion] fix: make input/output file save paths configurable and disableable (#19580)

Co-authored-by: Mick <mickjagger19@icloud.com>

.gitignore

@@ -248,6 +248,7 @@ lmms-eval
 **/.serena/
 ctags/
 outputs/
+inputs/
 
 # Eval Cache
 .longbench_cache/
@@ -272,6 +273,7 @@ sgl-kernel/csrc/**/*_musa/
 *.mudmp
 
 # Others
+# diffusion 3D outputs
 *.glb
 *.ply
 *.npz

python/sglang/multimodal_gen/configs/pipeline_configs/base.py

@@ -539,7 +539,7 @@ class PipelineConfig:
         cls, kwargs: dict[str, Any], config_cli_prefix: str = ""
     ) -> "PipelineConfig":
         """
-        Load PipelineConfig from kwargs Dictionary.
+        Load PipelineConfig from kwargs Dictionary, as part of the ServerArg initialization process
         kwargs: dictionary of kwargs
         config_cli_prefix: prefix of CLI arguments for this PipelineConfig instance
         """
@@ -583,7 +583,11 @@ class PipelineConfig:
                     f"using {pipeline_config_cls.__name__} directly without model_index.json"
                 )
             else:
-                model_info = get_model_info(model_path, backend=kwargs.get("backend"))
+                model_info = get_model_info(
+                    model_path,
+                    backend=kwargs.get("backend"),
+                    model_id=kwargs.get("model_id"),
+                )
                 if model_info is None:
                     from sglang.multimodal_gen.registry import (
                         _PIPELINE_CONFIG_REGISTRY,
@@ -599,7 +603,11 @@ class PipelineConfig:
                     )
                 pipeline_config_cls = model_info.pipeline_config_cls
         else:
-            model_info = get_model_info(model_path, backend=kwargs.get("backend"))
+            model_info = get_model_info(
+                model_path,
+                backend=kwargs.get("backend"),
+                model_id=kwargs.get("model_id"),
+            )
             if model_info is None:
                 raise ValueError(
                     f"Could not get model info for '{model_path}'. "

python/sglang/multimodal_gen/configs/sample/sampling_params.py

@@ -14,13 +14,16 @@ import unicodedata
 import uuid
 from dataclasses import dataclass
 from enum import Enum, auto
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
-from sglang.multimodal_gen.utils import StoreBoolean
+from sglang.multimodal_gen.utils import StoreBoolean, expand_path_fields
 
 logger = init_logger(__name__)
 
+if TYPE_CHECKING:
+    from sglang.multimodal_gen.runtime.server_args import ServerArgs
+
 
 def _json_safe(obj: Any):
     """
@@ -354,6 +357,8 @@ class SamplingParams:
         """
         final adjustment, called after merged with user params
         """
+        expand_path_fields(self)
+
         # TODO: SamplingParams should not rely on ServerArgs
         pipeline_config = server_args.pipeline_config
         if not isinstance(self.prompt, str):
@@ -374,11 +379,14 @@ class SamplingParams:
 
         self.data_type = server_args.pipeline_config.task_type.data_type()
 
-        if self.output_path is None and server_args.output_path is not None:
+        if self.output_path is None:
+            if server_args.output_path is not None:
                 self.output_path = server_args.output_path
                 logger.debug(
                     f"Overriding output_path with server configuration: {self.output_path}"
                 )
+            else:
+                self.save_output = False
 
         # Process negative prompt
         if self.negative_prompt is not None and not self.negative_prompt.isspace():
@@ -504,15 +512,18 @@ class SamplingParams:
         from sglang.multimodal_gen.registry import get_model_info
 
         backend = kwargs.pop("backend", None)
-        model_info = get_model_info(model_path, backend=backend)
+        model_id = kwargs.pop("model_id", None)
+        model_info = get_model_info(model_path, backend=backend, model_id=model_id)
         sampling_params: SamplingParams = model_info.sampling_param_cls(**kwargs)
         return sampling_params
 
     @staticmethod
-    def from_user_sampling_params_args(model_path: str, server_args, *args, **kwargs):
+    def from_user_sampling_params_args(
+        model_path: str, server_args: "ServerArgs", *args, **kwargs
+    ):
         try:
             sampling_params = SamplingParams.from_pretrained(
-                model_path, backend=server_args.backend
+                model_path, backend=server_args.backend, model_id=server_args.model_id
             )
         except (AttributeError, ValueError) as e:
             # Handle safetensors files or other cases where model_index.json is not available
@@ -890,6 +901,8 @@ class SamplingParams:
         return {attr: getattr(args, attr) for attr in attrs if hasattr(args, attr)}
 
     def output_file_path(self):
+        if self.output_path is None:
+            return None
         return os.path.join(self.output_path, self.output_file_name)
 
     def _merge_with_user_params(self, user_params: "SamplingParams"):

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

@@ -1,6 +1,7 @@
 # Copied and adapted from: https://github.com/hao-ai-lab/FastVideo
 
 import base64
+import contextlib
 import os
 import time
 from typing import List, Optional
@@ -23,6 +24,7 @@ from sglang.multimodal_gen.runtime.entrypoints.openai.utils import (
     merge_image_input_list,
     process_generation_batch,
     save_image_to_path,
+    temp_dir_if_disabled,
 )
 from sglang.multimodal_gen.runtime.entrypoints.utils import prepare_request
 from sglang.multimodal_gen.runtime.pipelines_core.schedule_batch import OutputBatch
@@ -53,11 +55,13 @@ def _build_image_response_kwargs(
     b64_list: list[str] | None = None,
     cloud_url: str | None = None,
     fallback_url: str | None = None,
+    is_persistent: bool = True,
 ) -> dict:
     """Build ImageResponse data list.
 
     For b64_json: uses pre-read b64_list (call _read_b64_for_paths first).
     For url: uses cloud_url or fallback_url.
+    file_path is omitted when is_persistent=False to avoid exposing stale temp paths.
     """
     ret = None
     if resp_format == "b64_json":
@@ -67,7 +71,7 @@ def _build_image_response_kwargs(
             ImageResponseData(
                 b64_json=b64,
                 revised_prompt=prompt,
-                file_path=os.path.abspath(path),
+                file_path=os.path.abspath(path) if is_persistent else None,
             )
             for b64, path in zip(b64_list, save_file_path_list)
         ]
@@ -84,7 +88,11 @@ def _build_image_response_kwargs(
                 ImageResponseData(
                     url=url,
                     revised_prompt=prompt,
-                    file_path=os.path.abspath(save_file_path_list[0]),
+                    file_path=(
+                        os.path.abspath(save_file_path_list[0])
+                        if is_persistent
+                        else None
+                    ),
                 )
             ],
         }
@@ -103,13 +111,17 @@ async def generations(
     request: ImageGenerationsRequest,
 ):
     request_id = generate_request_id()
+    server_args = get_global_server_args()
     ext = choose_output_image_ext(request.output_format, request.background)
+
+    with temp_dir_if_disabled(server_args.output_path) as output_dir:
         sampling = build_sampling_params(
             request_id,
             prompt=request.prompt,
             size=request.size,
             num_outputs_per_prompt=max(1, min(int(request.n or 1), 10)),
             output_file_name=f"{request_id}.{ext}",
+            output_path=output_dir,
             seed=request.seed,
             generator_device=request.generator_device,
             num_inference_steps=request.num_inference_steps,
@@ -121,7 +133,7 @@ async def generations(
             output_quality=request.output_quality,
         )
         batch = prepare_request(
-        server_args=get_global_server_args(),
+            server_args=server_args,
             sampling_params=sampling,
         )
         # Add diffusers_kwargs if provided
@@ -136,17 +148,20 @@ async def generations(
 
         # read b64 before cloud upload may delete the local file
         b64_list = (
-        _read_b64_for_paths(save_file_path_list) if resp_format == "b64_json" else None
+            _read_b64_for_paths(save_file_path_list)
+            if resp_format == "b64_json"
+            else None
         )
 
         cloud_url = await cloud_storage.upload_and_cleanup(save_file_path)
 
+        is_persistent = server_args.output_path is not None
         await IMAGE_STORE.upsert(
             request_id,
             {
                 "id": request_id,
                 "created_at": int(time.time()),
-            "file_path": None if cloud_url else save_file_path,
+                "file_path": None if cloud_url or not is_persistent else save_file_path,
                 "url": cloud_url,
             },
         )
@@ -159,8 +174,8 @@ async def generations(
             result,
             b64_list=b64_list,
             cloud_url=cloud_url,
-        # Provide a local fallback URL when cloud upload is unavailable, mirroring /v1/images/edits.
+            fallback_url=f"/v1/images/{request_id}/content" if is_persistent else None,
-        fallback_url=f"/v1/images/{request_id}/content",
+            is_persistent=is_persistent,
         )
 
     return ImageResponse(**response_kwargs)
@@ -193,6 +208,7 @@ async def edits(
     num_frames: int = Form(1),
 ):
     request_id = generate_request_id()
+    server_args = get_global_server_args()
     # Resolve images from either `image` or `image[]` (OpenAI SDK sends `image[]` when list is provided)
     images = image or image_array
     urls = url or url_array
@@ -202,23 +218,27 @@ async def edits(
             status_code=422, detail="Field 'image' or 'url' is required"
         )
 
-    # Save all input images; additional images beyond the first are saved for potential future use
-    uploads_dir = os.path.join("inputs", "uploads")
-    os.makedirs(uploads_dir, exist_ok=True)
-
     image_list = merge_image_input_list(images, urls)
 
+    with contextlib.ExitStack() as stack:
+        uploads_dir = stack.enter_context(
+            temp_dir_if_disabled(server_args.input_save_path)
+        )
+        output_dir = stack.enter_context(temp_dir_if_disabled(server_args.output_path))
+
         input_paths = []
         try:
             for idx, img in enumerate(image_list):
                 filename = img.filename if hasattr(img, "filename") else f"image_{idx}"
                 input_path = await save_image_to_path(
-                img, os.path.join(uploads_dir, f"{request_id}_{idx}_{filename}")
+                    img,
+                    os.path.join(uploads_dir, f"{request_id}_{idx}_{filename}"),
                 )
                 input_paths.append(input_path)
         except Exception as e:
             raise HTTPException(
-            status_code=400, detail=f"Failed to process image source: {str(e)}"
+                status_code=400,
+                detail=f"Failed to process image source: {str(e)}",
             )
 
         ext = choose_output_image_ext(output_format, background)
@@ -228,6 +248,7 @@ async def edits(
             size=size,
             num_outputs_per_prompt=max(1, min(int(n or 1), 10)),
             output_file_name=f"{request_id}.{ext}",
+            output_path=output_dir,
             image_path=input_paths,
             seed=seed,
             generator_device=generator_device,
@@ -241,7 +262,7 @@ async def edits(
             output_quality=output_quality,
         )
         batch = prepare_request(
-        server_args=get_global_server_args(),
+            server_args=server_args,
             sampling_params=sampling,
         )
         save_file_path_list, result = await process_generation_batch(
@@ -252,19 +273,23 @@ async def edits(
 
         # read b64 before cloud upload may delete the local file
         b64_list = (
-        _read_b64_for_paths(save_file_path_list) if resp_format == "b64_json" else None
+            _read_b64_for_paths(save_file_path_list)
+            if resp_format == "b64_json"
+            else None
         )
 
         cloud_url = await cloud_storage.upload_and_cleanup(save_file_path)
 
+        is_persistent = server_args.output_path is not None
+        is_input_persistent = server_args.input_save_path is not None
         await IMAGE_STORE.upsert(
             request_id,
             {
                 "id": request_id,
                 "created_at": int(time.time()),
-            "file_path": None if cloud_url else save_file_path,
+                "file_path": None if cloud_url or not is_persistent else save_file_path,
                 "url": cloud_url,
-            "input_image_paths": input_paths,
+                "input_image_paths": input_paths if is_input_persistent else None,
                 "num_input_images": len(input_paths),
             },
         )
@@ -277,7 +302,8 @@ async def edits(
             result,
             b64_list=b64_list,
             cloud_url=cloud_url,
-        fallback_url=f"/v1/images/{request_id}/content",
+            fallback_url=f"/v1/images/{request_id}/content" if is_persistent else None,
+            is_persistent=is_persistent,
         )
 
     return ImageResponse(**response_kwargs)
@@ -298,7 +324,12 @@ async def download_image_content(
         )
 
     file_path = item.get("file_path")
-    if not file_path or not os.path.exists(file_path):
+    if not file_path:
+        raise HTTPException(
+            status_code=404,
+            detail="Image was not persisted on disk (output_path is disabled). Use b64_json response_format or configure cloud storage.",
+        )
+    if not os.path.exists(file_path):
         raise HTTPException(status_code=404, detail="Image is still being generated")
 
     ext = os.path.splitext(file_path)[1].lower()

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

@@ -2,8 +2,11 @@
 import base64
 import os
 import re
+import shutil
+import tempfile
 import time
-from typing import Any, List, Optional, Union
+from contextlib import contextmanager
+from typing import Any, Generator, List, Optional, Union
 
 import httpx
 from fastapi import UploadFile
@@ -47,6 +50,23 @@ DEFAULT_FPS = 24
 DEFAULT_VIDEO_SECONDS = 4
 
 
+@contextmanager
+def temp_dir_if_disabled(
+    configured_path: str | None,
+) -> Generator[str, None, None]:
+    """Yield *configured_path* when it is set, otherwise create a temporary
+    directory that is automatically removed when the context exits."""
+    if configured_path is not None:
+        os.makedirs(configured_path, exist_ok=True)
+        yield configured_path
+    else:
+        tmp = tempfile.mkdtemp(prefix="sglang_")
+        try:
+            yield tmp
+        finally:
+            shutil.rmtree(tmp, ignore_errors=True)
+
+
 def _parse_size(size: str) -> tuple[int, int] | tuple[None, None]:
     try:
         parts = size.lower().replace(" ", "").split("x")

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

@@ -3,6 +3,8 @@
 import asyncio
 import json
 import os
+import shutil
+import tempfile
 import time
 from typing import Any, Dict, Optional
 
@@ -99,14 +101,15 @@ def _video_job_from_sampling(
     }
 
 
-async def _save_first_input_image(image_sources, request_id: str) -> str | None:
+async def _save_first_input_image(
+    image_sources, request_id: str, uploads_dir: str
+) -> str | None:
     """Save the first input image from a list of sources and return its path."""
     image_list = merge_image_input_list(image_sources)
     if not image_list:
         return None
     image = image_list[0]
 
-    uploads_dir = os.path.join("inputs", "uploads")
     os.makedirs(uploads_dir, exist_ok=True)
 
     filename = image.filename if hasattr(image, "filename") else "url_image"
@@ -114,7 +117,13 @@ async def _save_first_input_image(image_sources, request_id: str) -> str | None:
     return await save_image_to_path(image, target_path)
 
 
-async def _dispatch_job_async(job_id: str, batch: Req) -> None:
+async def _dispatch_job_async(
+    job_id: str,
+    batch: Req,
+    *,
+    temp_dirs: list[str] | None = None,
+    output_persistent: bool = True,
+) -> None:
     from sglang.multimodal_gen.runtime.scheduler_client import async_scheduler_client
 
     try:
@@ -125,12 +134,15 @@ async def _dispatch_job_async(job_id: str, batch: Req) -> None:
 
         cloud_url = await cloud_storage.upload_and_cleanup(save_file_path)
 
+        persistent_path = (
+            save_file_path if not cloud_url and output_persistent else None
+        )
         update_fields = {
             "status": "completed",
             "progress": 100,
             "completed_at": int(time.time()),
             "url": cloud_url,
-            "file_path": save_file_path if not cloud_url else None,
+            "file_path": persistent_path,
         }
         update_fields = add_common_data_to_response(
             update_fields, request_id=job_id, result=result
@@ -141,6 +153,9 @@ async def _dispatch_job_async(job_id: str, batch: Req) -> None:
         await VIDEO_STORE.update_fields(
             job_id, {"status": "failed", "error": {"message": str(e)}}
         )
+    finally:
+        for td in temp_dirs or []:
+            shutil.rmtree(td, ignore_errors=True)
 
 
 # TODO: support image to video generation
@@ -176,6 +191,22 @@ async def create_video(
     server_args = get_global_server_args()
     task_type = server_args.pipeline_config.task_type
 
+    # Resolve input upload directory (may be a temp dir when saving is disabled)
+    temp_dirs: list[str] = []
+    if server_args.input_save_path is not None:
+        uploads_dir = server_args.input_save_path
+        os.makedirs(uploads_dir, exist_ok=True)
+    else:
+        uploads_dir = tempfile.mkdtemp(prefix="sglang_input_")
+        temp_dirs.append(uploads_dir)
+
+    # Resolve output directory
+    effective_output_path = server_args.output_path
+    output_persistent = True
+    if "multipart/form-data" not in content_type:
+        # JSON body may carry a per-request output_path; checked after parsing below
+        pass
+
     if "multipart/form-data" in content_type:
         if not prompt:
             raise HTTPException(status_code=400, detail="prompt is required")
@@ -187,7 +218,9 @@ async def create_video(
                 detail="input_reference or reference_url is required for image-to-video generation",
             )
         try:
-            input_path = await _save_first_input_image(image_sources, request_id)
+            input_path = await _save_first_input_image(
+                image_sources, request_id, uploads_dir
+            )
         except Exception as e:
             raise HTTPException(
                 status_code=400, detail=f"Failed to process image source: {str(e)}"
@@ -258,7 +291,7 @@ async def create_video(
             if payload.get("reference_url"):
                 try:
                     input_path = await _save_first_input_image(
-                        payload.get("reference_url"), request_id
+                        payload.get("reference_url"), request_id, uploads_dir
                     )
                 except Exception as e:
                     raise HTTPException(
@@ -270,6 +303,17 @@ async def create_video(
         except Exception as e:
             raise HTTPException(status_code=400, detail=f"Invalid request body: {e}")
 
+    # Resolve per-request output_path override
+    effective_output_path = req.output_path or server_args.output_path
+    if effective_output_path is None:
+        output_tmp = tempfile.mkdtemp(prefix="sglang_output_")
+        temp_dirs.append(output_tmp)
+        effective_output_path = output_tmp
+        output_persistent = False
+
+    # Inject resolved output_path so _build_video_sampling_params picks it up
+    req.output_path = effective_output_path
+
     logger.debug(f"Server received from create_video endpoint: req={req}")
 
     try:
@@ -289,7 +333,14 @@ async def create_video(
     if req.diffusers_kwargs:
         batch.extra["diffusers_kwargs"] = req.diffusers_kwargs
     # Enqueue the job asynchronously and return immediately
-    asyncio.create_task(_dispatch_job_async(request_id, batch))
+    asyncio.create_task(
+        _dispatch_job_async(
+            request_id,
+            batch,
+            temp_dirs=temp_dirs or None,
+            output_persistent=output_persistent,
+        )
+    )
     return VideoResponse(**job)
 
 

python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py

@@ -247,11 +247,9 @@ class Req:
             base, ext = os.path.splitext(output_file_name)
             output_file_name = f"{base}_{output_idx}{ext}"
 
-        return (
+        if self.output_path is None or not output_file_name:
-            os.path.join(self.output_path, output_file_name)
+            return None
-            if output_file_name
+        return os.path.join(self.output_path, output_file_name)
-            else None
-        )
 
     def set_as_warmup(self):
         self.is_warmup = True

python/sglang/multimodal_gen/runtime/server_args.py

@@ -39,7 +39,11 @@ from sglang.multimodal_gen.runtime.utils.logging_utils import (
     configure_logger,
     init_logger,
 )
-from sglang.multimodal_gen.utils import FlexibleArgumentParser, StoreBoolean
+from sglang.multimodal_gen.utils import (
+    FlexibleArgumentParser,
+    StoreBoolean,
+    expand_path_fields,
+)
 
 logger = init_logger(__name__)
 
@@ -293,6 +297,7 @@ class ServerArgs:
     scheduler_port: int = 5555
 
     output_path: str | None = "outputs/"
+    input_save_path: str | None = "inputs/uploads"
 
     # Prompt text file for batch processing
     prompt_file_path: str | None = None
@@ -332,7 +337,8 @@ class ServerArgs:
         return self.host is None or self.port is None
 
     def _adjust_path(self):
-        self.model_path = os.path.expanduser(self.model_path)
+        expand_path_fields(self)
+        self._adjust_save_paths()
 
     def _adjust_parameters(self):
         """set defaults and normalize values."""
@@ -354,6 +360,13 @@ class ServerArgs:
         self._validate_parallelism()
         self._validate_cfg_parallel()
 
+    def _adjust_save_paths(self):
+        """Normalize empty-string save paths to None (disabled)."""
+        if self.output_path is not None and self.output_path.strip() == "":
+            self.output_path = None
+        if self.input_save_path is not None and self.input_save_path.strip() == "":
+            self.input_save_path = None
+
     def _adjust_quant_config(self):
         """validate and adjust"""
 
@@ -838,7 +851,13 @@ class ServerArgs:
             "--output-path",
             type=str,
             default=ServerArgs.output_path,
-            help="Directory path to save generated images/videos",
+            help='Directory path to save generated images/videos. Set to "" to disable persistent saving.',
+        )
+        parser.add_argument(
+            "--input-save-path",
+            type=str,
+            default=ServerArgs.input_save_path,
+            help='Directory path to save uploaded input images/videos. Set to "" to disable persistent saving.',
         )
 
         # LoRA

python/sglang/multimodal_gen/utils.py

@@ -34,6 +34,24 @@ logger = init_logger(__name__)
 
 T = TypeVar("T")
 
+
+def expand_path_fields(obj) -> None:
+    """In-place expanduser on all dataclass fields whose name ends with '_path' or '_paths'."""
+    eu = os.path.expanduser
+    for f in fields(obj):
+        v = getattr(obj, f.name)
+        if f.name.endswith("_path") and isinstance(v, str):
+            setattr(obj, f.name, eu(v))
+        elif f.name.endswith("_path") and isinstance(v, list):
+            setattr(obj, f.name, [eu(x) if isinstance(x, str) else x for x in v])
+        elif f.name.endswith("_paths") and isinstance(v, dict):
+            setattr(
+                obj,
+                f.name,
+                {k: eu(p) if isinstance(p, str) else p for k, p in v.items()},
+            )
+
+
 # TODO(will): used to convert server_args.precision to torch.dtype. Find a
 # cleaner way to do this.
 PRECISION_TO_TYPE = {