sglang 0.3.2__tar.gz → 0.3.3.post1__tar.gz

{sglang-0.3.2/sglang.egg-info → sglang-0.3.3.post1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.3.2
+Version: 0.3.3.post1
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -233,6 +233,7 @@ Requires-Dist: uvloop; extra == "srt"
 Requires-Dist: zmq; extra == "srt"
 Requires-Dist: vllm==0.5.5; extra == "srt"
 Requires-Dist: outlines>=0.0.44; extra == "srt"
+Requires-Dist: modelscope; extra == "srt"
 Provides-Extra: openai
 Requires-Dist: openai>=1.0; extra == "openai"
 Requires-Dist: tiktoken; extra == "openai"
@@ -256,8 +257,8 @@ Provides-Extra: dev
 Requires-Dist: sglang[all]; extra == "dev"
 Requires-Dist: sglang[test]; extra == "dev"
 
-<div align="center">
-<img src="https://raw.githubusercontent.com/sgl-project/sglang/main/assets/logo.png" alt="logo" width="400"></img>
+<div align="center"  id="sglangtop">
+<img src="https://raw.githubusercontent.com/sgl-project/sglang/main/assets/logo.png" alt="logo" width="400" margin="10px"></img>
 
 [![PyPI](https://img.shields.io/pypi/v/sglang)](https://pypi.org/project/sglang)
 ![PyPI - Downloads](https://img.shields.io/pypi/dm/sglang)
@@ -269,16 +270,10 @@ Requires-Dist: sglang[test]; extra == "dev"
 
 --------------------------------------------------------------------------------
 
-| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Paper**](https://arxiv.org/abs/2312.07104) | [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2ngly9muu-t37XiH87qvD~6rVBTkTEHw) | [**Join Weekly Development Meeting**](https://calendar.app.google/v2Tw3kuHkKYyp8VV7) |
+| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Paper**](https://arxiv.org/abs/2312.07104) | [**Slides**](https://github.com/sgl-project/sgl-learning-materials/blob/main/slides/amd_dev_day_v2.pptx) | [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2ngly9muu-t37XiH87qvD~6rVBTkTEHw) | [**Join Bi-Weekly Development Meeting (Oct. 19)**](https://calendar.app.google/GYW7S8QGoanCuaxW6) |
 
-SGLang is a fast serving framework for large language models and vision language models.
-It makes your interaction with models faster and more controllable by co-designing the backend runtime and frontend language.
-The core features include:
-
-- **Fast Backend Runtime**: Provides efficient serving with RadixAttention for prefix caching, jump-forward constrained decoding, continuous batching, token attention (paged attention), tensor parallelism, FlashInfer kernels, chunked prefill, and quantization (INT4/FP8/AWQ/GPTQ).
-- **Flexible Frontend Language**: Offers an intuitive interface for programming LLM applications, including chained generation calls, advanced prompting, control flow, multi-modal inputs, parallelism, and external interactions.
-- **Extensive Model Support**: Supports a wide range of generative models (Llama 3, Gemma 2, Mistral, QWen, DeepSeek, LLaVA, etc.) and embedding models (e5-mistral), with easy extensibility for integrating new models.
-- **Active Community**: SGLang is open-source and backed by an active community with industry adoption.
+## Upcoming Events
+- [Oct. 16, 2024] Online meetup for efficient LLM deployment and serving, co-hosted by SGLang, FlashInfer, and MLC LLM! Fill out the [Google form](https://forms.gle/B3YeedLxmrrhL1NM8) to receive the invite link.
 
 ## News
 - [2024/09] 🔥 SGLang v0.3 Release: 7x Faster DeepSeek MLA, 1.5x Faster torch.compile, Multi-Image/Video LLaVA-OneVision ([blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/)).
@@ -294,6 +289,16 @@ The core features include:
 
 </details>
 
+## About
+SGLang is a fast serving framework for large language models and vision language models.
+It makes your interaction with models faster and more controllable by co-designing the backend runtime and frontend language.
+The core features include:
+
+- **Fast Backend Runtime**: Provides efficient serving with RadixAttention for prefix caching, jump-forward constrained decoding, continuous batching, token attention (paged attention), tensor parallelism, FlashInfer kernels, chunked prefill, and quantization (INT4/FP8/AWQ/GPTQ).
+- **Flexible Frontend Language**: Offers an intuitive interface for programming LLM applications, including chained generation calls, advanced prompting, control flow, multi-modal inputs, parallelism, and external interactions.
+- **Extensive Model Support**: Supports a wide range of generative models (Llama, Gemma, Mistral, QWen, DeepSeek, LLaVA, etc.) and embedding models (e5-mistral), with easy extensibility for integrating new models.
+- **Active Community**: SGLang is open-source and backed by an active community with industry adoption.
+
 ## Contents
 - [Install](#install)
 - [Backend: SGLang Runtime (SRT)](#backend-sglang-runtime-srt)
@@ -318,7 +323,7 @@ pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/
 ### Method 2: From source
 ```
 # Use the last release branch
-git clone -b v0.3.2 https://github.com/sgl-project/sglang.git
+git clone -b v0.3.3.post1 https://github.com/sgl-project/sglang.git
 cd sglang
 
 pip install --upgrade pip
@@ -339,7 +344,7 @@ docker run --gpus all \
     --env "HF_TOKEN=<secret>" \
     --ipc=host \
     lmsysorg/sglang:latest \
-    python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --host 0.0.0.0 --port 30000
+    python3 -m sglang.launch_server --model-path meta-llama/Llama-3.1-8B-Instruct --host 0.0.0.0 --port 30000
 ```
 
 ### Method 4: Using docker compose
@@ -379,7 +384,7 @@ resources:
 run: |
   conda deactivate
   python3 -m sglang.launch_server \
-    --model-path meta-llama/Meta-Llama-3.1-8B-Instruct \
+    --model-path meta-llama/Llama-3.1-8B-Instruct \
     --host 0.0.0.0 \
     --port 30000
 ```
@@ -421,7 +426,8 @@ curl http://localhost:30000/generate \
     }
   }'
 ```
-Learn more about the argument format [here](docs/en/sampling_params.md).
+
+Learn more about the argument specification, streaming, and multi-modal support [here](docs/en/sampling_params.md).
 
 ### OpenAI Compatible API
 In addition, the server supports OpenAI-compatible APIs.
@@ -460,7 +466,7 @@ response = client.embeddings.create(
 print(response)
 ```
 
-It supports streaming, vision, and most features of the Chat/Completions/Models/Batch endpoints specified by the [OpenAI API Reference](https://platform.openai.com/docs/api-reference/).
+It supports streaming, vision, and almost all features of the Chat/Completions/Models/Batch endpoints specified by the [OpenAI API Reference](https://platform.openai.com/docs/api-reference/).
 
 ### Additional Server Arguments
 - To enable multi-GPU tensor parallelism, add `--tp 2`. If it reports the error "peer access is not supported between these two devices", add `--enable-p2p-check` to the server launch command.
@@ -481,10 +487,11 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --chunked-prefill-size 4096
 ```
 - To enable torch.compile acceleration, add `--enable-torch-compile`. It accelerates small models on small batch sizes.
+- To enable torchao quantization, add `--torchao-config int4wo-128`. It supports various quantization strategies.
 - To enable fp8 weight quantization, add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.
 - To enable fp8 kv cache quantization, add `--kv-cache-dtype fp8_e5m2`.
 - If the model does not have a chat template in the Hugging Face tokenizer, you can specify a [custom chat template](docs/en/custom_chat_template.md).
-- To run tensor parallelism on multiple nodes, add `--nnodes 2`. If you have two nodes with two GPUs on each node and want to run TP=4, let `sgl-dev-0` be the hostname of the first node and `50000` be an available port.
+- To run tensor parallelism on multiple nodes, add `--nnodes 2`. If you have two nodes with two GPUs on each node and want to run TP=4, let `sgl-dev-0` be the hostname of the first node and `50000` be an available port, you can use the following commands. If you meet deadlock, please try to add `--disable-cuda-graph`
 ```
 # Node 0
 python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --tp 4 --nccl-init sgl-dev-0:50000 --nnodes 2 --node-rank 0
@@ -499,9 +506,9 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 - Llama / Llama 2 / Llama 3 / Llama 3.1
 - Mistral / Mixtral / Mistral NeMo
 - Gemma / Gemma 2
-- OLMoE
 - Qwen / Qwen 2 / Qwen 2 MoE
 - DeepSeek / DeepSeek 2
+- OLMoE
 - [LLaVA-OneVision](https://llava-vl.github.io/blog/2024-08-05-llava-onevision/)
   - `python3 -m sglang.launch_server --model-path lmms-lab/llava-onevision-qwen2-7b-ov --port=30000 --chat-template=chatml-llava`
   - `python3 -m sglang.launch_server --model-path lmms-lab/llava-onevision-qwen2-72b-ov --port=30000 --tp-size=8 --chat-template=chatml-llava`
@@ -523,7 +530,6 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 - XVERSE / XVERSE MoE
 - SmolLM
 
-
 **Embedding Models**
 
 - e5-mistral
@@ -544,6 +550,17 @@ Launch [Qwen2-7B-Instruct](https://www.modelscope.cn/models/qwen/qwen2-7b-instru
 ```
 SGLANG_USE_MODELSCOPE=true python -m sglang.launch_server --model-path qwen/Qwen2-7B-Instruct --port 30000
 ```
+
+Or start it by docker.
+```bash
+docker run --gpus all \
+    -p 30000:30000 \
+    -v ~/.cache/modelscope:/root/.cache/modelscope \
+    --env "SGLANG_USE_MODELSCOPE=true" \
+    --ipc=host \
+    lmsysorg/sglang:latest \
+    python3 -m sglang.launch_server --model-path Qwen/Qwen2.5-7B-Instruct --host 0.0.0.0 --port 30000
+```
   
 </details>
 
@@ -582,7 +599,7 @@ GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/
 The frontend language can be used with local models or API models. It is an alternative to the OpenAI API. You may found it easier to use for complex prompting workflow.
 
 ### Quick Start
-The example below shows how to use sglang to answer a mulit-turn question.
+The example below shows how to use sglang to answer a multi-turn question.
 
 #### Using Local Models
 First, launch a server with
@@ -830,3 +847,11 @@ Learn more at this [blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/).
 ## Citation And Acknowledgment
 Please cite our paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.
 We also learned from the design and reused code from the following projects: [Guidance](https://github.com/guidance-ai/guidance), [vLLM](https://github.com/vllm-project/vllm), [LightLLM](https://github.com/ModelTC/lightllm), [FlashInfer](https://github.com/flashinfer-ai/flashinfer), [Outlines](https://github.com/outlines-dev/outlines), and [LMQL](https://github.com/eth-sri/lmql).
+
+
+
+<p align="center">
+  <a href="#sglangtop" target="_blank">
+  <bold>Back To Top </bold>
+  </a>
+</p>

{sglang-0.3.2 → sglang-0.3.3.post1}/README.md

@@ -1,5 +1,5 @@
-<div align="center">
-<img src="https://raw.githubusercontent.com/sgl-project/sglang/main/assets/logo.png" alt="logo" width="400"></img>
+<div align="center"  id="sglangtop">
+<img src="https://raw.githubusercontent.com/sgl-project/sglang/main/assets/logo.png" alt="logo" width="400" margin="10px"></img>
 
 [![PyPI](https://img.shields.io/pypi/v/sglang)](https://pypi.org/project/sglang)
 ![PyPI - Downloads](https://img.shields.io/pypi/dm/sglang)
@@ -11,16 +11,10 @@
 
 --------------------------------------------------------------------------------
 
-| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Paper**](https://arxiv.org/abs/2312.07104) | [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2ngly9muu-t37XiH87qvD~6rVBTkTEHw) | [**Join Weekly Development Meeting**](https://calendar.app.google/v2Tw3kuHkKYyp8VV7) |
+| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Paper**](https://arxiv.org/abs/2312.07104) | [**Slides**](https://github.com/sgl-project/sgl-learning-materials/blob/main/slides/amd_dev_day_v2.pptx) | [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2ngly9muu-t37XiH87qvD~6rVBTkTEHw) | [**Join Bi-Weekly Development Meeting (Oct. 19)**](https://calendar.app.google/GYW7S8QGoanCuaxW6) |
 
-SGLang is a fast serving framework for large language models and vision language models.
-It makes your interaction with models faster and more controllable by co-designing the backend runtime and frontend language.
-The core features include:
-
-- **Fast Backend Runtime**: Provides efficient serving with RadixAttention for prefix caching, jump-forward constrained decoding, continuous batching, token attention (paged attention), tensor parallelism, FlashInfer kernels, chunked prefill, and quantization (INT4/FP8/AWQ/GPTQ).
-- **Flexible Frontend Language**: Offers an intuitive interface for programming LLM applications, including chained generation calls, advanced prompting, control flow, multi-modal inputs, parallelism, and external interactions.
-- **Extensive Model Support**: Supports a wide range of generative models (Llama 3, Gemma 2, Mistral, QWen, DeepSeek, LLaVA, etc.) and embedding models (e5-mistral), with easy extensibility for integrating new models.
-- **Active Community**: SGLang is open-source and backed by an active community with industry adoption.
+## Upcoming Events
+- [Oct. 16, 2024] Online meetup for efficient LLM deployment and serving, co-hosted by SGLang, FlashInfer, and MLC LLM! Fill out the [Google form](https://forms.gle/B3YeedLxmrrhL1NM8) to receive the invite link.
 
 ## News
 - [2024/09] 🔥 SGLang v0.3 Release: 7x Faster DeepSeek MLA, 1.5x Faster torch.compile, Multi-Image/Video LLaVA-OneVision ([blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/)).
@@ -36,6 +30,16 @@ The core features include:
 
 </details>
 
+## About
+SGLang is a fast serving framework for large language models and vision language models.
+It makes your interaction with models faster and more controllable by co-designing the backend runtime and frontend language.
+The core features include:
+
+- **Fast Backend Runtime**: Provides efficient serving with RadixAttention for prefix caching, jump-forward constrained decoding, continuous batching, token attention (paged attention), tensor parallelism, FlashInfer kernels, chunked prefill, and quantization (INT4/FP8/AWQ/GPTQ).
+- **Flexible Frontend Language**: Offers an intuitive interface for programming LLM applications, including chained generation calls, advanced prompting, control flow, multi-modal inputs, parallelism, and external interactions.
+- **Extensive Model Support**: Supports a wide range of generative models (Llama, Gemma, Mistral, QWen, DeepSeek, LLaVA, etc.) and embedding models (e5-mistral), with easy extensibility for integrating new models.
+- **Active Community**: SGLang is open-source and backed by an active community with industry adoption.
+
 ## Contents
 - [Install](#install)
 - [Backend: SGLang Runtime (SRT)](#backend-sglang-runtime-srt)
@@ -60,7 +64,7 @@ pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/
 ### Method 2: From source
 ```
 # Use the last release branch
-git clone -b v0.3.2 https://github.com/sgl-project/sglang.git
+git clone -b v0.3.3.post1 https://github.com/sgl-project/sglang.git
 cd sglang
 
 pip install --upgrade pip
@@ -81,7 +85,7 @@ docker run --gpus all \
     --env "HF_TOKEN=<secret>" \
     --ipc=host \
     lmsysorg/sglang:latest \
-    python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --host 0.0.0.0 --port 30000
+    python3 -m sglang.launch_server --model-path meta-llama/Llama-3.1-8B-Instruct --host 0.0.0.0 --port 30000
 ```
 
 ### Method 4: Using docker compose
@@ -121,7 +125,7 @@ resources:
 run: |
   conda deactivate
   python3 -m sglang.launch_server \
-    --model-path meta-llama/Meta-Llama-3.1-8B-Instruct \
+    --model-path meta-llama/Llama-3.1-8B-Instruct \
     --host 0.0.0.0 \
     --port 30000
 ```
@@ -163,7 +167,8 @@ curl http://localhost:30000/generate \
     }
   }'
 ```
-Learn more about the argument format [here](docs/en/sampling_params.md).
+
+Learn more about the argument specification, streaming, and multi-modal support [here](docs/en/sampling_params.md).
 
 ### OpenAI Compatible API
 In addition, the server supports OpenAI-compatible APIs.
@@ -202,7 +207,7 @@ response = client.embeddings.create(
 print(response)
 ```
 
-It supports streaming, vision, and most features of the Chat/Completions/Models/Batch endpoints specified by the [OpenAI API Reference](https://platform.openai.com/docs/api-reference/).
+It supports streaming, vision, and almost all features of the Chat/Completions/Models/Batch endpoints specified by the [OpenAI API Reference](https://platform.openai.com/docs/api-reference/).
 
 ### Additional Server Arguments
 - To enable multi-GPU tensor parallelism, add `--tp 2`. If it reports the error "peer access is not supported between these two devices", add `--enable-p2p-check` to the server launch command.
@@ -223,10 +228,11 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --chunked-prefill-size 4096
 ```
 - To enable torch.compile acceleration, add `--enable-torch-compile`. It accelerates small models on small batch sizes.
+- To enable torchao quantization, add `--torchao-config int4wo-128`. It supports various quantization strategies.
 - To enable fp8 weight quantization, add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.
 - To enable fp8 kv cache quantization, add `--kv-cache-dtype fp8_e5m2`.
 - If the model does not have a chat template in the Hugging Face tokenizer, you can specify a [custom chat template](docs/en/custom_chat_template.md).
-- To run tensor parallelism on multiple nodes, add `--nnodes 2`. If you have two nodes with two GPUs on each node and want to run TP=4, let `sgl-dev-0` be the hostname of the first node and `50000` be an available port.
+- To run tensor parallelism on multiple nodes, add `--nnodes 2`. If you have two nodes with two GPUs on each node and want to run TP=4, let `sgl-dev-0` be the hostname of the first node and `50000` be an available port, you can use the following commands. If you meet deadlock, please try to add `--disable-cuda-graph`
 ```
 # Node 0
 python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --tp 4 --nccl-init sgl-dev-0:50000 --nnodes 2 --node-rank 0
@@ -241,9 +247,9 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 - Llama / Llama 2 / Llama 3 / Llama 3.1
 - Mistral / Mixtral / Mistral NeMo
 - Gemma / Gemma 2
-- OLMoE
 - Qwen / Qwen 2 / Qwen 2 MoE
 - DeepSeek / DeepSeek 2
+- OLMoE
 - [LLaVA-OneVision](https://llava-vl.github.io/blog/2024-08-05-llava-onevision/)
   - `python3 -m sglang.launch_server --model-path lmms-lab/llava-onevision-qwen2-7b-ov --port=30000 --chat-template=chatml-llava`
   - `python3 -m sglang.launch_server --model-path lmms-lab/llava-onevision-qwen2-72b-ov --port=30000 --tp-size=8 --chat-template=chatml-llava`
@@ -265,7 +271,6 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 - XVERSE / XVERSE MoE
 - SmolLM
 
-
 **Embedding Models**
 
 - e5-mistral
@@ -286,6 +291,17 @@ Launch [Qwen2-7B-Instruct](https://www.modelscope.cn/models/qwen/qwen2-7b-instru
 ```
 SGLANG_USE_MODELSCOPE=true python -m sglang.launch_server --model-path qwen/Qwen2-7B-Instruct --port 30000
 ```
+
+Or start it by docker.
+```bash
+docker run --gpus all \
+    -p 30000:30000 \
+    -v ~/.cache/modelscope:/root/.cache/modelscope \
+    --env "SGLANG_USE_MODELSCOPE=true" \
+    --ipc=host \
+    lmsysorg/sglang:latest \
+    python3 -m sglang.launch_server --model-path Qwen/Qwen2.5-7B-Instruct --host 0.0.0.0 --port 30000
+```
   
 </details>
 
@@ -324,7 +340,7 @@ GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/
 The frontend language can be used with local models or API models. It is an alternative to the OpenAI API. You may found it easier to use for complex prompting workflow.
 
 ### Quick Start
-The example below shows how to use sglang to answer a mulit-turn question.
+The example below shows how to use sglang to answer a multi-turn question.
 
 #### Using Local Models
 First, launch a server with
@@ -572,3 +588,11 @@ Learn more at this [blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/).
 ## Citation And Acknowledgment
 Please cite our paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.
 We also learned from the design and reused code from the following projects: [Guidance](https://github.com/guidance-ai/guidance), [vLLM](https://github.com/vllm-project/vllm), [LightLLM](https://github.com/ModelTC/lightllm), [FlashInfer](https://github.com/flashinfer-ai/flashinfer), [Outlines](https://github.com/outlines-dev/outlines), and [LMQL](https://github.com/eth-sri/lmql).
+
+
+
+<p align="center">
+  <a href="#sglangtop" target="_blank">
+  <bold>Back To Top </bold>
+  </a>
+</p>

{sglang-0.3.2 → sglang-0.3.3.post1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sglang"
-version = "0.3.2"
+version = "0.3.3.post1"
 description = "SGLang is yet another fast serving framework for large language models and vision language models."
 readme = "README.md"
 requires-python = ">=3.8"
@@ -23,7 +23,7 @@ dependencies = [
 srt = ["aiohttp", "decord", "fastapi", "hf_transfer", "huggingface_hub", "interegular",
        "packaging", "pillow", "psutil", "pydantic", "python-multipart",
        "torch", "torchao", "uvicorn", "uvloop", "zmq",
-       "vllm==0.5.5", "outlines>=0.0.44"]
+       "vllm==0.5.5", "outlines>=0.0.44", "modelscope"]
 openai = ["openai>=1.0", "tiktoken"]
 anthropic = ["anthropic>=0.20.0"]
 litellm = ["litellm>=1.0.0"]

{sglang-0.3.2 → sglang-0.3.3.post1}/sglang/__init__.py

@@ -1,6 +1,7 @@
 # SGL API Components
 
 from sglang.api import (
+    Engine,
     Runtime,
     assistant,
     assistant_begin,
@@ -31,6 +32,7 @@ from sglang.lang.choices import (
 # SGLang DSL APIs
 __all__ = [
     "Runtime",
+    "Engine",
     "assistant",
     "assistant_begin",
     "assistant_end",

{sglang-0.3.2 → sglang-0.3.3.post1}/sglang/api.py

@@ -33,13 +33,23 @@ def function(
 
 
 def Runtime(*args, **kwargs):
-    # Avoid importing unnecessary dependency
     os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+
+    # Avoid importing unnecessary dependency
     from sglang.srt.server import Runtime
 
     return Runtime(*args, **kwargs)
 
 
+def Engine(*args, **kwargs):
+    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+
+    # Avoid importing unnecessary dependency
+    from sglang.srt.server import Engine
+
+    return Engine(*args, **kwargs)
+
+
 def set_default_backend(backend: BaseBackend):
     global_config.default_backend = backend
 
@@ -48,6 +58,10 @@ def flush_cache(backend: Optional[BaseBackend] = None):
     backend = backend or global_config.default_backend
     if backend is None:
         return False
+
+    # If backend is Runtime
+    if hasattr(backend, "endpoint"):
+        backend = backend.endpoint
     return backend.flush_cache()
 
 
@@ -55,12 +69,17 @@ def get_server_args(backend: Optional[BaseBackend] = None):
     backend = backend or global_config.default_backend
     if backend is None:
         return None
+
+    # If backend is Runtime
+    if hasattr(backend, "endpoint"):
+        backend = backend.endpoint
     return backend.get_server_args()
 
 
 def gen(
     name: Optional[str] = None,
     max_tokens: Optional[int] = None,
+    min_tokens: Optional[int] = None,
     stop: Optional[Union[str, List[str]]] = None,
     stop_token_ids: Optional[List[int]] = None,
     temperature: Optional[float] = None,
@@ -100,6 +119,7 @@ def gen(
     return SglGen(
         name,
         max_tokens,
+        min_tokens,
         stop,
         stop_token_ids,
         temperature,
@@ -139,6 +159,7 @@ def gen_int(
     return SglGen(
         name,
         max_tokens,
+        None,
         stop,
         stop_token_ids,
         temperature,
@@ -177,6 +198,7 @@ def gen_string(
     return SglGen(
         name,
         max_tokens,
+        None,
         stop,
         stop_token_ids,
         temperature,

{sglang-0.3.2 → sglang-0.3.3.post1}/sglang/bench_latency.py

@@ -47,6 +47,7 @@ I'm going to the park
 import argparse
 import dataclasses
 import itertools
+import json
 import logging
 import multiprocessing
 import os
@@ -62,10 +63,11 @@ import torch.distributed as dist
 from sglang.srt.configs.model_config import ModelConfig
 from sglang.srt.hf_transformers_utils import get_tokenizer
 from sglang.srt.managers.schedule_batch import Req, ScheduleBatch
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_executor.model_runner import ModelRunner
 from sglang.srt.sampling.sampling_params import SamplingParams
 from sglang.srt.server import _set_envs_and_config
-from sglang.srt.server_args import ServerArgs
+from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.utils import (
     configure_logger,
     kill_child_process,
@@ -121,7 +123,7 @@ class BenchArgs:
         )
 
 
-def load_model(server_args, tp_rank):
+def load_model(server_args, port_args, tp_rank):
     suppress_other_loggers()
     rank_print = print if tp_rank == 0 else lambda *args, **kwargs: None
 
@@ -129,6 +131,7 @@ def load_model(server_args, tp_rank):
         server_args.model_path,
         server_args.trust_remote_code,
         context_length=server_args.context_length,
+        model_override_args=json.loads(server_args.json_model_override_args),
     )
     model_runner = ModelRunner(
         model_config=model_config,
@@ -136,7 +139,7 @@ def load_model(server_args, tp_rank):
         gpu_id=tp_rank,
         tp_rank=tp_rank,
         tp_size=server_args.tp_size,
-        nccl_port=28888,
+        nccl_port=port_args.nccl_port,
         server_args=server_args,
     )
     rank_print(f"max_total_num_tokens={model_runner.max_total_num_tokens}")
@@ -167,9 +170,13 @@ def prepare_inputs_for_correctness_test(bench_args, tokenizer):
         assert len(input_ids[i]) > bench_args.cut_len
 
         tmp_input_ids = input_ids[i][: bench_args.cut_len]
-        req = Req(rid=i, origin_input_text=prompts[i], origin_input_ids=tmp_input_ids)
+        req = Req(
+            rid=i,
+            origin_input_text=prompts[i],
+            origin_input_ids=tmp_input_ids,
+            sampling_params=sampling_params,
+        )
         req.prefix_indices = []
-        req.sampling_params = sampling_params
         req.fill_ids = req.origin_input_ids
         req.extend_input_len = len(req.fill_ids) - len(req.prefix_indices)
         reqs.append(req)
@@ -199,9 +206,13 @@ def prepare_synthetic_inputs_for_latency_test(batch_size, input_len):
 
     reqs = []
     for i in range(len(input_ids)):
-        req = Req(rid=i, origin_input_text="", origin_input_ids=list(input_ids[i]))
+        req = Req(
+            rid=i,
+            origin_input_text="",
+            origin_input_ids=list(input_ids[i]),
+            sampling_params=sampling_params,
+        )
         req.prefix_indices = []
-        req.sampling_params = sampling_params
         req.fill_ids = req.origin_input_ids
         req.extend_input_len = len(req.fill_ids) - len(req.prefix_indices)
         reqs.append(req)
@@ -209,6 +220,7 @@ def prepare_synthetic_inputs_for_latency_test(batch_size, input_len):
     return reqs
 
 
+@torch.inference_mode()
 def extend(reqs, model_runner):
     batch = ScheduleBatch.init_new(
         reqs=reqs,
@@ -217,28 +229,33 @@ def extend(reqs, model_runner):
         tree_cache=None,
     )
     batch.prepare_for_extend(model_runner.model_config.vocab_size)
-    logits_output = model_runner.forward(batch)
-    next_token_ids = model_runner.sample(logits_output, batch).tolist()
+    model_worker_batch = batch.get_model_worker_batch()
+    forward_batch = ForwardBatch.init_new(model_worker_batch, model_runner)
+    logits_output = model_runner.forward(forward_batch)
+    next_token_ids = model_runner.sample(logits_output, forward_batch).tolist()
     return next_token_ids, logits_output.next_token_logits, batch
 
 
+@torch.inference_mode()
 def decode(input_token_ids, batch, model_runner):
     batch.prepare_for_decode(input_token_ids)
-    logits_output = model_runner.forward(batch)
-    next_token_ids = model_runner.sample(logits_output, batch).tolist()
+    model_worker_batch = batch.get_model_worker_batch()
+    forward_batch = ForwardBatch.init_new(model_worker_batch, model_runner)
+    logits_output = model_runner.forward(forward_batch)
+    next_token_ids = model_runner.sample(logits_output, forward_batch).tolist()
     return next_token_ids, logits_output.next_token_logits
 
 
-@torch.inference_mode()
 def correctness_test(
     server_args,
+    port_args,
     bench_args,
     tp_rank,
 ):
     rank_print = print if tp_rank == 0 else lambda *args, **kwargs: None
 
     # Load the model
-    model_runner, tokenizer = load_model(server_args, tp_rank)
+    model_runner, tokenizer = load_model(server_args, port_args, tp_rank)
 
     # Prepare inputs
     input_ids, reqs = prepare_inputs_for_correctness_test(bench_args, tokenizer)
@@ -271,7 +288,6 @@ def correctness_test(
         rank_print(tokenizer.decode(output_ids[i]), "\n")
 
 
-@torch.inference_mode()
 def latency_test_run_once(
     run_name, model_runner, rank_print, reqs, batch_size, input_len, output_len
 ):
@@ -324,13 +340,16 @@ def latency_test_run_once(
             rank_print(
                 f"Decode.  latency: {latency:6.5f} s, throughput: {throughput:9.2f} token/s"
             )
-    med_decode_latency = np.median(decode_latencies)
-    med_decode_throughput = batch_size / med_decode_latency
-    rank_print(
-        f"Decode.  median latency: {med_decode_latency:6.5f} s, median throughput: {med_decode_throughput:9.2f} token/s"
-    )
-    measurement_results["median_decode_latency"] = med_decode_latency
-    measurement_results["median_decode_throughput"] = med_decode_throughput
+
+    # record decode timing from 2nd output
+    if output_len > 1:
+        med_decode_latency = np.median(decode_latencies)
+        med_decode_throughput = batch_size / med_decode_latency
+        rank_print(
+            f"Decode.  median latency: {med_decode_latency:6.5f} s, median throughput: {med_decode_throughput:9.2f} token/s"
+        )
+        measurement_results["median_decode_latency"] = med_decode_latency
+        measurement_results["median_decode_throughput"] = med_decode_throughput
 
     throughput = (input_len + output_len) * batch_size / tot_latency
     rank_print(
@@ -343,15 +362,15 @@ def latency_test_run_once(
 
 def latency_test(
     server_args,
+    port_args,
     bench_args,
     tp_rank,
 ):
     configure_logger(server_args, prefix=f" TP{tp_rank}")
-    _set_envs_and_config(server_args)
     rank_print = print if tp_rank == 0 else lambda *args, **kwargs: None
 
     # Load the model
-    model_runner, tokenizer = load_model(server_args, tp_rank)
+    model_runner, tokenizer = load_model(server_args, port_args, tp_rank)
 
     # Prepare inputs for warm up
     reqs = prepare_synthetic_inputs_for_latency_test(
@@ -367,7 +386,7 @@ def latency_test(
         reqs,
         bench_args.batch_size[0],
         bench_args.input_len[0],
-        4,  # shorter decoding to speed up the warmup
+        8,  # shorter decoding to speed up the warmup
     )
     rank_print("Benchmark ...")
 
@@ -453,6 +472,7 @@ def plot_latency_test(
 
 
 def main(server_args, bench_args):
+    _set_envs_and_config(server_args)
 
     if server_args.model_path:
         if bench_args.correctness_test:
@@ -468,8 +488,10 @@ def main(server_args, bench_args):
             "provide --result-filename for plotting the results"
         )
 
+    port_args = PortArgs.init_new(server_args)
+
     if server_args.tp_size == 1:
-        work_func(server_args, bench_args, 0)
+        work_func(server_args, port_args, bench_args, 0)
     else:
         workers = []
         for tp_rank in range(server_args.tp_size):
@@ -477,6 +499,7 @@ def main(server_args, bench_args):
                 target=work_func,
                 args=(
                     server_args,
+                    port_args,
                     bench_args,
                     tp_rank,
                 ),
@@ -503,8 +526,6 @@ if __name__ == "__main__":
         format="%(message)s",
     )
 
-    multiprocessing.set_start_method("spawn", force=True)
-
     try:
         main(server_args, bench_args)
     except Exception as e: