fix multimodal gen issues (#12765)

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

python/sglang/multimodal_gen/README.md

@@ -24,7 +24,7 @@ sgl-diffusion has the following features:
 ## Getting Started
 
 ```bash
-uv pip install sglang[.diffusion] --prerelease=allow
+uv pip install 'sglang[diffusion]' --prerelease=allow
 ```
 
 For more information, check the [docs](https://github.com/sgl-project/sglang/tree/main/python/sglang/multimodal_gen/docs/install.md).

python/sglang/multimodal_gen/docs/cli.md

@@ -143,7 +143,7 @@ SERVER_ARGS=(
   --ring-degree=2
 )
 
-sglang serve"${SERVER_ARGS[@]}"
+sglang serve "${SERVER_ARGS[@]}"
 ```
 
 - **--model-path**: Which model to load. The example uses `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`.

python/sglang/multimodal_gen/docs/install.md

@@ -11,7 +11,7 @@ It is recommended to use uv for a faster installation:
 ```bash
 pip install --upgrade pip
 pip install uv
-uv pip install sglang[.diffusion] --prerelease=allow
+uv pip install 'sglang[diffusion]' --prerelease=allow
 ```
 
 ## Method 2: From source
@@ -29,13 +29,9 @@ pip install -e "python/.[diffusion]"
 uv pip install --prerelease=allow  -e "python/.[diffusion]"
 ```
 
-**Quick fixes for common problems:**
-
-- If you want to develop sgl-diffusion, it is recommended to use Docker. The Docker image is `lmsysorg/sgl-diffusion:latest`.
-
 ## Method 3: Using Docker
 
-The Docker images are available on Docker Hub at [lmsysorg/sgl-diffusion](), built from the [Dockerfile](https://github.com/sgl-project/sgl-diffusion/tree/main/docker).
+The Docker images are available on Docker Hub at [lmsysorg/sglang](), built from the [Dockerfile](https://github.com/sgl-project/sglang/tree/main/docker).
 Replace `<secret>` below with your HuggingFace Hub [token](https://huggingface.co/docs/hub/en/security-tokens).
 
 ```bash
@@ -45,7 +41,7 @@ docker run --gpus all \
     -v ~/.cache/huggingface:/root/.cache/huggingface \
     --env "HF_TOKEN=<secret>" \
     --ipc=host \
-    lmsysorg/sglang:diffusion \
+    lmsysorg/sglang:dev \
     sglang generate --model-path black-forest-labs/FLUX.1-dev \
     --prompt "A logo With Bold Large text: SGL Diffusion" \
     --save-output

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

@@ -69,8 +69,8 @@ def _build_sampling_params_from_request(
         save_output=True,
     )
     sampling_params = sampling_params.from_user_sampling_params(user_params)
+    sampling_params.set_output_file_name()
     sampling_params.log(server_args)
-    sampling_params.set_output_file_ext()
     return sampling_params
 
 

python/sglang/multimodal_gen/runtime/server_args.py

@@ -777,13 +777,6 @@ class ServerArgs:
                 )
             self.sp_degree = self.ulysses_degree = self.ring_degree = 1
 
-        if (
-            self.ring_degree is not None
-            and self.ring_degree > 1
-            and self.attention_backend != "fa3"
-        ):
-            raise ValueError("Ring Attention is only supported for fa3 backend for now")
-
         if self.sp_degree == -1:
             # assume we leave all remaining gpus to sp
             num_gpus_per_group = self.dp_size * self.tp_size
@@ -815,6 +808,17 @@ class ServerArgs:
                 f"Ring degree not set, " f"using default value {self.ring_degree}"
             )
 
+        if self.ring_degree > 1:
+            if self.attention_backend != None and self.attention_backend != "fa3":
+                raise ValueError(
+                    "Ring Attention is only supported for fa3 backend for now"
+                )
+            else:
+                self.attention_backend = "fa3"
+                logger.info(
+                    "Ring Attention is currently only supported for fa3, attention_backend has been automatically set to fa3"
+                )
+
         if self.sp_degree == -1:
             self.sp_degree = self.ring_degree * self.ulysses_degree
             logger.info(

python/sglang/multimodal_gen/test/cli/test_serve.py

@@ -3,9 +3,13 @@
 import asyncio
 import base64
 import subprocess
+import tempfile
 import time
 import unittest
+import uuid
+from contextlib import contextmanager
 from pathlib import Path
+from urllib.request import urlopen
 
 from openai import OpenAI
 
@@ -13,6 +17,20 @@ from sglang.multimodal_gen.runtime.utils.common import kill_process_tree
 from sglang.multimodal_gen.test.test_utils import is_mp4, is_png, wait_for_port
 
 
+@contextmanager
+def downloaded_temp_file(url: str, prefix: str = "i2v_input_", suffix: str = ".jpg"):
+    tmp_path = Path(tempfile.gettempdir()) / f"{prefix}{uuid.uuid4().hex}{suffix}"
+    with urlopen(url) as resp:
+        tmp_path.write_bytes(resp.read())
+    try:
+        yield tmp_path
+    finally:
+        try:
+            tmp_path.unlink(missing_ok=True)
+        except Exception:
+            pass
+
+
 def wait_for_video_completion(client, video_id, timeout=300, check_interval=3):
     start = time.time()
     video = client.videos.retrieve(video_id)
@@ -27,7 +45,7 @@ def wait_for_video_completion(client, video_id, timeout=300, check_interval=3):
 
 class TestVideoHttpServer(unittest.TestCase):
     model_name = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-    timeout = 120
+    timeout = 500
     extra_args = []
 
     def _create_wait_and_download(
@@ -77,7 +95,7 @@ class TestVideoHttpServer(unittest.TestCase):
             api_key="sk-proj-1234567890", base_url="http://localhost:30010/v1"
         )
         content = self._create_wait_and_download(
-            client, "A calico cat playing a piano on stage", "832x480"
+            client, "A plane is taking off.", "832x480"
         )
         self.assertTrue(is_mp4(content))
 
@@ -97,7 +115,7 @@ class TestVideoHttpServer(unittest.TestCase):
         async def send_concurrent_requests():
             tasks = [
                 generate_and_check_video(
-                    "A dog playing a piano on stage",
+                    "A ship is beside the port.",
                     "832x480",
                 )
                 for _ in range(num_requests)
@@ -107,14 +125,6 @@ class TestVideoHttpServer(unittest.TestCase):
         asyncio.run(send_concurrent_requests())
 
 
-class TestFastWan2_1HttpServer(TestVideoHttpServer):
-    model_name = "FastVideo/FastWan2.1-T2V-1.3B-Diffusers"
-
-
-class TestFastWan2_2HttpServer(TestVideoHttpServer):
-    model_name = "FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers"
-
-
 class TestImage2VideoHttpServer(unittest.TestCase):
     model_name = "Wan-AI/Wan2.2-I2V-A14B-Diffusers"
     timeout = 1200
@@ -124,15 +134,17 @@ class TestImage2VideoHttpServer(unittest.TestCase):
         self, client: OpenAI, prompt: str, size: str
     ) -> bytes:
 
-        image_path = "https://github.com/Wan-Video/Wan2.2/blob/990af50de458c19590c245151197326e208d7191/examples/i2v_input.JPG?raw=true"
+        image_url = "https://github.com/Wan-Video/Wan2.2/blob/990af50de458c19590c245151197326e208d7191/examples/i2v_input.JPG?raw=true"
-        image_path = Path(image_path)
+        with downloaded_temp_file(
-        video = client.videos.create(
+            image_url, prefix="i2v_input_", suffix=".jpg"
-            prompt=prompt,
+        ) as tmp_path:
-            input_reference=image_path,
+            video = client.videos.create(
-            size=size,
+                prompt=prompt,
-            seconds=10,
+                input_reference=tmp_path,
-            extra_body={"fps": 16, "num_frames": 125},
+                size=size,
-        )
+                seconds=10,
+                extra_body={"fps": 16, "num_frames": 125},
+            )
         # TODO: Some combinations of num_frames and fps may cause errors and need further investigation.
         video_id = video.id
         self.assertEqual(video.status, "queued")
@@ -149,18 +161,20 @@ class TestImage2VideoHttpServer(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.base_command = [
-            "sgl-diffusion",
+            "sglang",
             "serve",
             "--model-path",
             f"{cls.model_name}",
+            "--num-gpus",
+            "4",
+            "--ulysses-degree",
+            "4",
             "--port",
             "30010",
         ]
 
         process = subprocess.Popen(
             cls.base_command + cls.extra_args,
-            # stdout=subprocess.PIPE,
-            # stderr=subprocess.PIPE,
             text=True,
             bufsize=1,
         )
@@ -176,7 +190,7 @@ class TestImage2VideoHttpServer(unittest.TestCase):
             api_key="sk-proj-1234567890", base_url="http://localhost:30010/v1"
         )
         content = self._create_wait_and_download(
-            client, "A girl is fighting a monster.", "832x480"
+            client, "A cat surfing on the sea.", "832x480"
         )
         self.assertTrue(is_mp4(content))
 
@@ -196,7 +210,7 @@ class TestImage2VideoHttpServer(unittest.TestCase):
         async def send_concurrent_requests():
             tasks = [
                 generate_and_check_video(
-                    "A dog playing a piano on stage",
+                    "A cat surfing on the sea.",
                     "832x480",
                 )
                 for _ in range(num_requests)