minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[diffusion] chore: refresh docs, retire stale knobs, and fix nightly attribution (#34663)

Browse Source
This commit is contained in:
Mick
2026-08-16 15:41:08 +08:00
committed by GitHub
parent a54de989c8
commit 2ee0d38a85
21 changed files with 370 additions and 130 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/cookbook/diffusion/FLUX/FLUX.mdx
+4 -6
View File
@@ -230,17 +230,16 @@ Test Environment:
```shell Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-video --dataset vbench --task t2v --num-prompts 1 --max-concurrency 1
--dataset vbench --task text-to-image --num-prompts 1 --max-concurrency 1
```
**Result**:
```text Output
================= Serving Benchmark Result =================
Backend: sglang-image
Model: black-forest-labs/FLUX.1-dev
Dataset: vbench
Task: t2v
Task: text-to-image
--------------------------------------------------
Benchmark duration (s): 50.97
Request rate: inf
@@ -317,17 +316,16 @@ Test Environment:
```shell Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-image --dataset vbench --task t2v --num-prompts 20 --max-concurrency 20
--dataset vbench --task text-to-image --num-prompts 20 --max-concurrency 20
```
**Result** :
```text Output
================= Serving Benchmark Result =================
Backend: sglang-image
Model: black-forest-labs/FLUX.1-dev
Dataset: vbench
Task: t2v
Task: text-to-image
--------------------------------------------------
Benchmark duration (s): 111.79
Request rate: inf
docs/cookbook/diffusion/LTX/LTX2 & LTX2.3.mdx
-2
View File
@@ -67,8 +67,6 @@ For two-stage pipelines, `--ltx2-two-stage-device-mode` controls transformer res
| `resident` | Best latency on high-VRAM GPUs because both DiTs can stay resident. |
| `original` | Closest to the original two-stage switching semantics. |
`snapshot` is kept only as a deprecated compatibility alias for `original` and may be removed after two release cycles; use `original` or `resident` in new configs.
Other deployment flags:
- `--lora-path`: Preload a community LoRA adapter.
docs/cookbook/diffusion/LingBot-Video/LingBot-Video-MoE.mdx
+147
View File
@@ -0,0 +1,147 @@
---
title: LingBot Video MoE
description: Serve the native LingBot Video MoE 30B-A3B text-to-video model with SGLang Diffusion.
metatags:
description: "Run robbyant/lingbot-video-moe-30b-a3b text-to-video generation with SGLang Diffusion."
---
import { DiffusionModelTags } from '/src/snippets/diffusion/model-tags.jsx';
<DiffusionModelTags tags={["video", "text-to-video", "mixture-of-experts"]} />
## 1. Model introduction
[LingBot Video MoE 30B-A3B](https://huggingface.co/robbyant/lingbot-video-moe-30b-a3b)
is a text-to-video mixture-of-experts model. SGLang Diffusion provides a native
pipeline for the public checkpoint:
| Model ID | Task | Default output |
| --- | --- | --- |
| `robbyant/lingbot-video-moe-30b-a3b` | Text to video | 480x480, 81 frames at 16 FPS |
The checkpoint expects a structured JSON caption rather than an unexpanded
natural-language prompt. The JSON is passed as the request's `prompt` string;
it is not an `extra_params` object.
## 2. Installation
Install SGLang with the diffusion dependencies:
```bash Command
uv pip install "sglang[diffusion]" --prerelease=allow
```
See the [SGLang Diffusion installation guide](/docs/sglang-diffusion/installation)
for platform-specific setup.
## 3. Serve LingBot Video MoE
Start the server with the Hugging Face model ID:
```bash Command
sglang serve \
--model-path robbyant/lingbot-video-moe-30b-a3b \
--port 30010
```
## 4. Generate a video
The following request uses the compact 17-frame, 12-step smoke-test profile.
Use the model defaults of 81 frames and 40 steps for the released generation
profile.
```python Python
import json
import time
from pathlib import Path
import requests
base_url = "http://127.0.0.1:30010"
prompt = json.dumps(
{
"comprehensive_description": {
"scene_content_description": (
"A small silver robot arm on a white table slowly reaches "
"toward a red cube. The background is a softly lit laboratory wall."
),
"camera_movement_description": (
"The camera is static at eye level in a medium shot."
),
},
"camera_info": {
"color": "Neutral",
"frame_size": "Medium",
"shot_type_angle": "Eye level",
"lens_size": "Medium",
"composition": "Center",
"lighting": "Soft light",
"lighting_type": "Artificial light",
},
"world_knowledge": [],
"prominent_elements": [
{
"name": "robot arm",
"description": "A small silver robot arm with a two-finger gripper.",
"actions": [
{
"timestamp": "[0.0s - 1.0s]",
"action": "reaches toward the red cube",
}
],
"location": "center of the frame",
"relative_size": "dominant",
"shape_and_color": "articulated silver metal arm",
"texture": "brushed metal",
"appearance_details": "two-finger gripper and visible joints",
"relationship": "reaching toward the red cube on the table",
"orientation": "upright, base on the table",
"pose": "reaching",
}
],
},
separators=(",", ":"),
)
response = requests.post(
f"{base_url}/v1/videos",
json={
"model": "robbyant/lingbot-video-moe-30b-a3b",
"prompt": prompt,
"size": "640x384",
"num_frames": 17,
"fps": 16,
"num_inference_steps": 12,
"guidance_scale": 6.0,
"flow_shift": 3.0,
"seed": 0,
},
timeout=60,
)
response.raise_for_status()
video_id = response.json()["id"]
while True:
job = requests.get(f"{base_url}/v1/videos/{video_id}", timeout=30).json()
if job["status"] == "completed":
break
if job["status"] == "failed":
raise RuntimeError(job.get("error") or "Video generation failed")
time.sleep(1)
video = requests.get(
f"{base_url}/v1/videos/{video_id}/content",
timeout=300,
)
video.raise_for_status()
Path("lingbot_video_moe.mp4").write_bytes(video.content)
```
## 5. Request constraints
- `num_frames` must be `1` or `4n+1`; examples include 17 and 81.
- Width and height must both be multiples of 16.
- The native defaults are `guidance_scale=6.0`, `flow_shift=3.0`,
`num_inference_steps=40`, and `fps=16`.
- Keep the prompt as serialized JSON. Raw free text is outside the
checkpoint's expected caption format.
docs/cookbook/diffusion/Qwen-Image/Qwen-Image-Edit.mdx
+4 -6
View File
@@ -220,17 +220,16 @@ sglang serve --model-path Qwen/Qwen-Image-Edit-2511 --port 30000
```shell Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-image --dataset vbench --task ti2i --num-prompts 1 --max-concurrency 1
--dataset vbench --task image-to-image --num-prompts 1 --max-concurrency 1
```
**Result**:
```text Output
================= Serving Benchmark Result =================
Backend: sglang-image
Model: Qwen/Qwen-Image-Edit-2511
Dataset: vbench
Task: ti2i
Task: image-to-image
--------------------------------------------------
Benchmark duration (s): 35.31
Request rate: inf
@@ -254,17 +253,16 @@ Peak Memory Median (MB): 47959.35
```shell Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-image --dataset vbench --task ti2i --num-prompts 20 --max-concurrency 20
--dataset vbench --task image-to-image --num-prompts 20 --max-concurrency 20
```
**Result**:
```text Output
================= Serving Benchmark Result =================
Backend: sglang-image
Model: Qwen/Qwen-Image-Edit-2511
Dataset: vbench
Task: ti2i
Task: image-to-image
--------------------------------------------------
Benchmark duration (s): 286.11
Request rate: inf
docs/cookbook/diffusion/Qwen-Image/Qwen-Image.mdx
+2 -2
View File
@@ -237,7 +237,7 @@ Test Environment:
```shell Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-image --dataset vbench --task text-to-image --num-prompts 1 --max-concurrency 1
--dataset vbench --task text-to-image --num-prompts 1 --max-concurrency 1
```
**Result**:
@@ -317,7 +317,7 @@ Test Environment:
```shell Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-image --dataset vbench --task text-to-image --num-prompts 20 --max-concurrency 20 --port 30000
--dataset vbench --task text-to-image --num-prompts 20 --max-concurrency 20 --port 30000
```
**Result**:
docs/cookbook/diffusion/SANA-Video/SANA-Video.mdx
+96
View File
@@ -0,0 +1,96 @@
---
title: SANA-Video
description: Serve the native SANA-Video 2B 480p text-to-video model with SGLang Diffusion.
metatags:
description: "Run Efficient-Large-Model/SANA-Video_2B_480p_diffusers text-to-video generation with SGLang Diffusion."
---
import { DiffusionModelTags } from '/src/snippets/diffusion/model-tags.jsx';
<DiffusionModelTags tags={["video", "text-to-video"]} />
## 1. Model introduction
[SANA-Video 2B 480p](https://huggingface.co/Efficient-Large-Model/SANA-Video_2B_480p_diffusers)
is a text-to-video model with a native SGLang Diffusion pipeline.
| Model ID | Task | Default output |
| --- | --- | --- |
| `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` | Text to video | 832x480, 81 frames at 16 FPS |
## 2. Installation
Install SGLang with the diffusion dependencies:
```bash Command
uv pip install "sglang[diffusion]" --prerelease=allow
```
See the [SGLang Diffusion installation guide](/docs/sglang-diffusion/installation)
for platform-specific setup.
## 3. Serve SANA-Video
```bash Command
sglang serve \
--model-path Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
--port 30010
```
## 4. Generate a video
The following request uses the compact 17-frame, 8-step profile covered by
server CI. Use the model defaults of 81 frames and 50 steps for the released
generation profile.
```python Python
import time
from pathlib import Path
import requests
base_url = "http://127.0.0.1:30010"
response = requests.post(
f"{base_url}/v1/videos",
json={
"model": "Efficient-Large-Model/SANA-Video_2B_480p_diffusers",
"prompt": (
"A red tram moves slowly through a sunlit city square while "
"pedestrians cross behind it. motion score: 30."
),
"size": "832x480",
"num_frames": 17,
"fps": 16,
"num_inference_steps": 8,
"guidance_scale": 6.0,
"seed": 42,
},
timeout=60,
)
response.raise_for_status()
video_id = response.json()["id"]
while True:
job = requests.get(f"{base_url}/v1/videos/{video_id}", timeout=30).json()
if job["status"] == "completed":
break
if job["status"] == "failed":
raise RuntimeError(job.get("error") or "Video generation failed")
time.sleep(1)
video = requests.get(
f"{base_url}/v1/videos/{video_id}/content",
timeout=300,
)
video.raise_for_status()
Path("sana_video.mp4").write_bytes(video.content)
```
## 5. Request constraints
- The default profile uses `832x480`, 81 frames, 50 inference steps, and 16 FPS.
- Frame counts are aligned to `4n+1`; for example, a request for 80 frames is
adjusted to 77.
- Use width and height values divisible by 16.
- The prompt supports an optional `motion score: N.` suffix to express the
desired amount of motion.
docs/cookbook/diffusion/Wan/Wan2.1.mdx
+2 -2
View File
@@ -200,7 +200,7 @@ You can use the built-in SGLang diffusion benchmark script to evaluate Wan2.1 pe
```bash Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-video --dataset vbench --task text-to-video --num-prompts 1 --max-concurrency 1
--dataset vbench --task text-to-video --num-prompts 1 --max-concurrency 1
```
**Result**:
@@ -304,7 +304,7 @@ You can use the built-in SGLang diffusion benchmark script to evaluate Wan2.1 pe
```bash Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-video --dataset vbench --task text-to-video --num-prompts 1 --max-concurrency 1
--dataset vbench --task text-to-video --num-prompts 1 --max-concurrency 1
```
**Result**:
docs/cookbook/diffusion/Wan/Wan2.2.mdx
+4 -6
View File
@@ -270,16 +270,15 @@ Test Environment:
**Benchmark Command**:
```shell Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-video --dataset vbench --task t2v --num-prompts 1 --max-concurrency 1
--dataset vbench --task text-to-video --num-prompts 1 --max-concurrency 1
```
**Result**:
```text Output
================= Serving Benchmark Result =================
Backend: sglang-video
Model: Wan-AI/Wan2.2-T2V-A14B-Diffusers
Dataset: vbench
Task: t2v
Task: text-to-video
--------------------------------------------------
Benchmark duration (s): 630.43
Request rate: inf
@@ -372,17 +371,16 @@ Test Environment:
```shell Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-video --dataset vbench --task t2v --num-prompts 20 --max-concurrency 20
--dataset vbench --task text-to-video --num-prompts 20 --max-concurrency 20
```
**Result**:
```text Output
================= Serving Benchmark Result =================
Backend: sglang-video
Model: Wan-AI/Wan2.2-T2V-A14B-Diffusers
Dataset: vbench
Task: t2v
Task: text-to-video
--------------------------------------------------
Benchmark duration (s): 5163.21
Request rate: inf
docs/cookbook/diffusion/Z-Image/Z-Image-Turbo.mdx
+2 -2
View File
@@ -225,7 +225,7 @@ Test Environment:
```shell Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-image --dataset vbench --task text-to-image --num-prompts 1 --max-concurrency 1
--dataset vbench --task text-to-image --num-prompts 1 --max-concurrency 1
```
**Result**:
@@ -305,7 +305,7 @@ Test Environment:
```shell Command
python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
--backend sglang-image --dataset vbench --task text-to-image --num-prompts 20 --max-concurrency 20
--dataset vbench --task text-to-image --num-prompts 20 --max-concurrency 20
```
**Result**:
docs/cookbook/diffusion/intro.mdx
+12
View File
@@ -80,6 +80,12 @@ Video models denoise a bounded latent video sequence for each request. Use these
href="/cookbook/diffusion/LTX/LTX2 & LTX2.3"
img="/cards/logos/ltx.svg"
/>
<Card
title="SANA-Video"
mode="card"
href="/cookbook/diffusion/SANA-Video/SANA-Video"
img="/cards/logos/sana.png"
/>
<Card
title="JoyAI-Echo"
mode="card"
@@ -98,6 +104,12 @@ Video models denoise a bounded latent video sequence for each request. Use these
href="/cookbook/diffusion/MiniMax/MiniMax-H3"
img="/cards/logos/minimax.png"
/>
<Card
title="LingBot Video MoE"
mode="card"
href="/cookbook/diffusion/LingBot-Video/LingBot-Video-MoE"
img="/cards/logos/inclusionai.png"
/>
</CardGroup>
## Realtime / World Models