chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[diffusion] doc: minor update docs (#13177)

Browse Source
This commit is contained in:
Mick
2025-11-21 14:35:29 +08:00
committed by GitHub
parent 8c212a2029
commit eda2f70033
23 changed files with 38 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/multimodal_gen/README.md
View File

@@ -2,9 +2,9 @@
<img src=https://github.com/lm-sys/lm-sys.github.io/releases/download/test/sgl-diffusion-logo.png width="80%"/>
</div>
**sgl-diffusion is an inference framework for accelerated image/video generation.**
**SGLang diffusion is an inference framework for accelerated image/video generation.**
SGLang-Diffusion features an end-to-end unified pipeline for accelerating diffusion models. It is designed to be modular and extensible, allowing users to easily add new models and optimizations.
SGLang diffusion features an end-to-end unified pipeline for accelerating diffusion models. It is designed to be modular and extensible, allowing users to easily add new models and optimizations.
## Key Features

2
python/sglang/multimodal_gen/configs/models/base.py
View File

@@ -46,7 +46,7 @@ class ModelConfig:
# Diffuser/Transformer parameters
arch_config: ArchConfig = field(default_factory=ArchConfig)
# sgl-diffusion-specific parameters here
# sglang-diffusion-specific parameters here
# i.e. STA, quantization, teacache
def __getattr__(self, name):

2
python/sglang/multimodal_gen/configs/models/dits/base.py
View File

@@ -43,7 +43,7 @@ class DiTArchConfig(ArchConfig):
class DiTConfig(ModelConfig):
arch_config: DiTArchConfig = field(default_factory=DiTArchConfig)
# sgl-diffusionDiT-specific parameters
# sglang-diffusion DiT-specific parameters
prefix: str = ""
quant_config: QuantizationConfig | None = None

2
python/sglang/multimodal_gen/configs/models/vaes/base.py
View File

@@ -26,7 +26,7 @@ class VAEArchConfig(ArchConfig):
class VAEConfig(ModelConfig):
arch_config: VAEArchConfig = field(default_factory=VAEArchConfig)
# sgl-diffusionVAE-specific parameters
# sglang-diffusion VAE-specific parameters
load_encoder: bool = True
load_decoder: bool = True

2
python/sglang/multimodal_gen/csrc/attn/vmoba_attn/README.md
View File

@@ -1,4 +1,4 @@
# Attention Kernel Used in sgl-diffusion
# Attention Kernel Used in SGLang diffusion
## VMoBA: Mixture-of-Block Attention for Video Diffusion Models (VMoBA)

10
python/sglang/multimodal_gen/docs/cli.md
View File

@@ -1,11 +1,11 @@
# sgl-diffusion CLI Inference
# SGLang diffusion CLI Inference
The sgl-diffusion CLI provides a quick way to access the sgl-diffusion inference pipeline for image and video generation.
The SGLang-diffusion CLI provides a quick way to access the inference pipeline for image and video generation.
## Prerequisites
- A working sgl-diffusion installation and the `sgl-diffusion` CLI available in `$PATH`.
- Python 3.10+ if you plan to use the OpenAI Python SDK.
- A working SGLang diffusion installation and the `sglang` CLI available in `$PATH`.
- Python 3.11+ if you plan to use the OpenAI Python SDK.
## Supported Arguments
@@ -127,7 +127,7 @@ sglang generate --help
## Serve
Launch the sgl-diffusion HTTP server and interact with it using the OpenAI SDK and curl. The server implements an OpenAI-compatible subset for Videos under the `/v1/videos` namespace.
Launch the SGLang diffusion HTTP server and interact with it using the OpenAI SDK and curl. The server implements an OpenAI-compatible subset for Videos under the `/v1/videos` namespace.
### Start the server

4
python/sglang/multimodal_gen/docs/install.md
View File

@@ -1,6 +1,6 @@
# Install sgl-diffusion
# Install SGLang-diffusion
You can install sgl-diffusion using one of the methods below.
You can install sglang-diffusion using one of the methods below.
This page primarily applies to common NVIDIA GPU platforms.

2
python/sglang/multimodal_gen/docs/support_matrix.md
View File

@@ -10,7 +10,7 @@ The symbols used have the following meanings:
## Models x Optimization
The `HuggingFace Model ID` can be passed directly to `from_pretrained()` methods, and sgl-diffusion will use the optimal
The `HuggingFace Model ID` can be passed directly to `from_pretrained()` methods, and sglang-diffusion will use the optimal
default parameters when initializing and generating videos.
### Video Generation Models

9
python/sglang/multimodal_gen/docs/support_new_models.md
View File

@@ -1,10 +1,12 @@
# How to Support New Diffusion Models
This document explains how to add support for new diffusion models in SGLang Diffusion.
This document explains how to add support for new diffusion models in SGLang diffusion.
## Architecture Overview
SGLang Diffusion is engineered for both performance and flexibility, built upon a modular pipeline architecture. This design allows developers to easily construct complex, customized pipelines for various diffusion models by combining and reusing different components.
SGLang diffusion is engineered for both performance and flexibility, built upon a modular pipeline architecture. This
design allows developers to easily construct complex, customized pipelines for various diffusion models by combining and
reusing different components.
At its core, the architecture revolves around two key concepts, as highlighted in our [blog post](https://lmsys.org/blog/2025-11-07-sglang-diffusion/#architecture):
@@ -101,4 +103,5 @@ To illustrate the process, let's look at how `Qwen-Image-Edit` is implemented. T
5. **Register Configs**:
- Register the configs in the central registry ([`registry.py`](https://github.com/sgl-project/sglang/blob/main/python/sglang/multimodal_gen/registry.py)) via `_register_configs` to enable automatic loading and instantiation for the model. Modules are automatically loaded and injected based on the config and repository structure.
By following this modular pattern of defining configurations and composing pipelines, you can integrate new diffusion models into SGLang with clarity and ease.
By following this pattern of defining configurations and composing pipelines, you can integrate new diffusion models
into SGLang with ease.

2
python/sglang/multimodal_gen/envs.py
View File

@@ -163,7 +163,7 @@ def maybe_convert_int(value: str | None) -> int | None:
environment_variables: dict[str, Callable[[], Any]] = {
# ================== Installation Time Env Vars ==================
# Target device of sgl-diffusion, supporting [cuda (by default),
# Target device of sglang-diffusion, supporting [cuda (by default),
# rocm, neuron, cpu, openvino]
"SGLANG_DIFFUSION_TARGET_DEVICE": lambda: os.getenv(
"SGLANG_DIFFUSION_TARGET_DEVICE", "cuda"

2
python/sglang/multimodal_gen/runtime/distributed/device_communicators/pynccl.py
View File

@@ -74,7 +74,7 @@ class PyNcclCommunicator:
self.available = True
self.disabled = False
logger.info("sgl-diffusion is using nccl==%s", self.nccl.ncclGetVersion())
logger.info("sglang-diffusion is using nccl==%s", self.nccl.ncclGetVersion())
if self.rank == 0:
# get the unique id from NCCL

2
python/sglang/multimodal_gen/runtime/distributed/parallel_state.py
View File

@@ -13,7 +13,7 @@
# Copyright 2023 The vLLM team.
# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
"""sgl-diffusion distributed state.
"""sglang-diffusion distributed state.
It takes over the control of the distributed environment from PyTorch.
The typical workflow is:

2
python/sglang/multimodal_gen/runtime/entrypoints/cli/generate.py
View File

@@ -62,7 +62,7 @@ def generate_cmd(args: argparse.Namespace):
class GenerateSubcommand(CLISubcommand):
"""The `generate` subcommand for the sgl-diffusion CLI"""
"""The `generate` subcommand for the sglang-diffusion CLI"""
def __init__(self) -> None:
self.name = "generate"

2
python/sglang/multimodal_gen/runtime/entrypoints/cli/main.py
View File

@@ -21,7 +21,7 @@ def cmd_init() -> list[CLISubcommand]:
def main() -> None:
parser = FlexibleArgumentParser(description="sgl-diffusion CLI")
parser = FlexibleArgumentParser(description="sglang-diffusion CLI")
parser.add_argument("-v", "--version", action="version", version="0.1.0")
subparsers = parser.add_subparsers(required=False, dest="subparser")

2
python/sglang/multimodal_gen/runtime/entrypoints/cli/serve.py
View File

@@ -35,7 +35,7 @@ def execute_serve_cmd(args: argparse.Namespace, unknown_args: list[str] | None =
class ServeSubcommand(CLISubcommand):
"""The `serve` subcommand for the sgl-diffusion CLI"""
"""The `serve` subcommand for the sglang-diffusion CLI"""
def __init__(self) -> None:
self.name = "serve"

2
python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py
View File

@@ -2,7 +2,7 @@
# SPDX-License-Identifier: Apache-2.0
"""
DiffGenerator module for sgl-diffusion.
DiffGenerator module for sglang-diffusion.
This module provides a consolidated interface for generating videos using
diffusion models.

2
python/sglang/multimodal_gen/runtime/entrypoints/utils.py
View File

@@ -2,7 +2,7 @@
# SPDX-License-Identifier: Apache-2.0
"""
DiffGenerator module for sgl-diffusion.
DiffGenerator module for sglang-diffusion.
This module provides a consolidated interface for generating videos using
diffusion models.

4
python/sglang/multimodal_gen/runtime/launch_server.py
View File

@@ -66,7 +66,7 @@ def launch_server(server_args: ServerArgs, launch_http_server: bool = True):
task_pipes_to_slaves_w,
result_pipes_from_slaves_r,
),
name=f"sgl-diffusionWorker-{i}",
name=f"sglang-diffusionWorker-{i}",
daemon=True,
)
else: # Slave workers
@@ -83,7 +83,7 @@ def launch_server(server_args: ServerArgs, launch_http_server: bool = True):
task_pipes_to_slaves_r[i - 1],
result_pipes_from_slaves_w[i - 1],
),
name=f"sgl-diffusionWorker-{i}",
name=f"sglang-diffusionWorker-{i}",
daemon=True,
)
scheduler_pipe_readers.append(reader)

4
python/sglang/multimodal_gen/runtime/layers/attention/selector.py
View File

@@ -41,7 +41,7 @@ def backend_name_to_enum(backend_name: str) -> AttentionBackendEnum | None:
def get_env_variable_attn_backend() -> AttentionBackendEnum | None:
"""
Get the backend override specified by the sgl-diffusion attention
Get the backend override specified by the sglang-diffusion attention
backend environment variable, if one is specified.
Returns:
@@ -169,7 +169,7 @@ def global_force_attn_backend_context_manager(
attn_backend: AttentionBackendEnum,
) -> Generator[None, None, None]:
"""
Globally force a sgl-diffusion attention backend override within a
Globally force a sglang-diffusion attention backend override within a
context manager, reverting the global attention backend
override to its prior state upon exiting the context
manager.

2
python/sglang/multimodal_gen/runtime/loader/fsdp_load.py
View File

@@ -4,7 +4,7 @@
# Adapted from torchtune
# Copyright 2024 The TorchTune Authors.
# Copyright 2025 The sgl-diffusion Authors.
# Copyright 2025 The sglang-diffusion Authors.
import contextlib
from collections.abc import Callable, Generator

2
python/sglang/multimodal_gen/runtime/models/vaes/hunyuanvae.py
View File

@@ -3,7 +3,7 @@
# SPDX-License-Identifier: Apache-2.0
# Adapted from diffusers
# Copyright 2024 The Hunyuan Team, The HuggingFace Team and The sgl-diffusion Team. All rights reserved.
# Copyright 2024 The Hunyuan Team, The HuggingFace Team and The sglang-diffusion Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.

4
python/sglang/multimodal_gen/runtime/server_args.py
View File

@@ -2,7 +2,7 @@
# SPDX-License-Identifier: Apache-2.0
# Inspired by SGLang: https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/server_args.py
"""The arguments of sgl-diffusion Inference."""
"""The arguments of sglang-diffusion Inference."""
import argparse
import dataclasses
import inspect
@@ -393,7 +393,7 @@ class ServerArgs:
type=str,
choices=ExecutionMode.choices(),
default=ServerArgs.mode.value,
help="The mode to run sgl-diffusion",
help="The mode to run SGLang-diffusion",
)
# Workload type

4
python/sglang/multimodal_gen/utils.py
View File

@@ -450,8 +450,8 @@ def import_pynvml():
status without initializing CUDA context in the current process.
Historically, there are two packages that provide pynvml:
- `nvidia-ml-py` (https://pypi.org/project/nvidia-ml-py/): The official
wrapper. It is a dependency of sgl-diffusion, and is installed when users
install sgl-diffusion. It provides a Python module named `pynvml`.
wrapper. It is a dependency of sglang-diffusion, and is installed when users
install sglang-diffusion. It provides a Python module named `pynvml`.
- `pynvml` (https://pypi.org/project/pynvml/): An unofficial wrapper.
Prior to version 12.0, it also provides a Python module `pynvml`,
and therefore conflicts with the official one which is a standalone Python file.