sglang 0.2.9.post1__tar.gz → 0.2.11__tar.gz

{sglang-0.2.9.post1/sglang.egg-info → sglang-0.2.11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.2.9.post1
+Version: 0.2.11
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -224,13 +224,13 @@ Requires-Dist: packaging; extra == "srt"
 Requires-Dist: pillow; extra == "srt"
 Requires-Dist: psutil; extra == "srt"
 Requires-Dist: pydantic; extra == "srt"
+Requires-Dist: python-multipart; extra == "srt"
 Requires-Dist: torch; extra == "srt"
 Requires-Dist: uvicorn; extra == "srt"
 Requires-Dist: uvloop; extra == "srt"
 Requires-Dist: zmq; extra == "srt"
-Requires-Dist: vllm==0.5.3.post1; extra == "srt"
+Requires-Dist: vllm==0.5.4; extra == "srt"
 Requires-Dist: outlines>=0.0.44; extra == "srt"
-Requires-Dist: python-multipart; extra == "srt"
 Provides-Extra: openai
 Requires-Dist: openai>=1.0; extra == "openai"
 Requires-Dist: tiktoken; extra == "openai"
@@ -238,11 +238,18 @@ Provides-Extra: anthropic
 Requires-Dist: anthropic>=0.20.0; extra == "anthropic"
 Provides-Extra: litellm
 Requires-Dist: litellm>=1.0.0; extra == "litellm"
+Provides-Extra: test
+Requires-Dist: jsonlines; extra == "test"
+Requires-Dist: matplotlib; extra == "test"
+Requires-Dist: pandas; extra == "test"
 Provides-Extra: all
 Requires-Dist: sglang[srt]; extra == "all"
 Requires-Dist: sglang[openai]; extra == "all"
 Requires-Dist: sglang[anthropic]; extra == "all"
 Requires-Dist: sglang[litellm]; extra == "all"
+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>
@@ -295,20 +302,20 @@ pip install --upgrade pip
 pip install "sglang[all]"
 
 # Install FlashInfer CUDA kernels
-pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
+pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/
 ```
 
 ### Method 2: From source
 ```
-# Use the stable v0.2.9.post1 branch
-git clone -b v0.2.9.post1 https://github.com/sgl-project/sglang.git
+# Use the last release branch
+git clone -b v0.2.11 https://github.com/sgl-project/sglang.git
 cd sglang
 
 pip install --upgrade pip
 pip install -e "python[all]"
 
 # Install FlashInfer CUDA kernels
-pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
+pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/
 ```
 
 ### Method 3: Using docker
@@ -382,7 +389,7 @@ response = client.chat.completions.create(
 print(response)
 ```
 
-It supports streaming, vision, and most features of the Chat/Completions/Models endpoints specified by the [OpenAI API Reference](https://platform.openai.com/docs/api-reference/).
+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/).
 
 ### Additional Server Arguments
 - Add `--tp 2` to enable tensor parallelism. If it indicates `peer access is not supported between these two devices`, add `--enable-p2p-check` option.
@@ -393,10 +400,14 @@ 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 --port 30000 --dp 2 --tp 2
 ```
-- If you see out-of-memory errors during serving, please try to reduce the memory usage of the KV cache pool by setting a smaller value of `--mem-fraction-static`. The default value is `0.9`
+- If you see out-of-memory errors during serving, please try to reduce the memory usage of the KV cache pool by setting a smaller value of `--mem-fraction-static`. The default value is `0.9`.
 ```
 python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --mem-fraction-static 0.7
 ```
+- If you see out-of-memory errors during prefill for long prompts on a model that supports long context, consider using chunked prefill.
+```
+python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --port 30000 --chunked-prefill-size 8192
+```
 - See [hyperparameter_tuning.md](docs/en/hyperparameter_tuning.md) on tuning hyperparameters for better performance.
 - Add `--nnodes 2` to run tensor parallelism on multiple nodes. 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.
 ```
@@ -410,22 +421,6 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 - To enable fp8 quantization, you can add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.
 - To enable experimental torch.compile support, you can add `--enable-torch-compile`. It accelerates small models on small batch sizes.
 
-### Run Llama 3.1 405B
-
-```bash
-## Run 405B (fp8) on a single node
-python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct-FP8 --tp 8
-
-## Run 405B (fp16) on two nodes
-# replace the `172.16.4.52:20000` with your own first node ip address and port, disable CUDA Graph temporarily
-
-# on the first node
-GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 0 --disable-cuda-graph --mem-frac 0.75
-
-# on the second
-GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 1 --disable-cuda-graph --mem-frac 0.75
-```
-
 ### Supported Models
 
 - Llama / Llama 2 / Llama 3 / Llama 3.1
@@ -451,9 +446,25 @@ GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/
 
 Instructions for supporting a new model are [here](https://github.com/sgl-project/sglang/blob/main/docs/en/model_support.md).
 
+### Run Llama 3.1 405B
+
+```bash
+## Run 405B (fp8) on a single node
+python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct-FP8 --tp 8
+
+## Run 405B (fp16) on two nodes
+# replace the `172.16.4.52:20000` with your own first node ip address and port, disable CUDA Graph temporarily
+
+# on the first node
+GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 0 --disable-cuda-graph --mem-frac 0.75
+
+# on the second
+GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 1 --disable-cuda-graph --mem-frac 0.75
+```
+
 ### Benchmark Performance
 
-- Benchmark a single static batch by running the following command without launching a server. The arguments are the same as those for `launch_server.py`. This is not a dynamic batching server, so it may run out of memory for a batch size that can run successfully with a real server. This is because a real server will truncate the prefill into several batches/chunks, while this unit test does not do this.
+- Benchmark a single static batch by running the following command without launching a server. The arguments are the same as for `launch_server.py`. Note that this is not a dynamic batching server, so it may run out of memory for a batch size that a real server can handle. A real server truncates the prefill into several batches, while this unit test does not. For accurate large batch testing, consider using `sglang.bench_serving`.
   ```
   python -m sglang.bench_latency --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch 32 --input-len 256 --output-len 32
   ```

{sglang-0.2.9.post1 → sglang-0.2.11}/README.md

@@ -49,20 +49,20 @@ pip install --upgrade pip
 pip install "sglang[all]"
 
 # Install FlashInfer CUDA kernels
-pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
+pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/
 ```
 
 ### Method 2: From source
 ```
-# Use the stable v0.2.9.post1 branch
-git clone -b v0.2.9.post1 https://github.com/sgl-project/sglang.git
+# Use the last release branch
+git clone -b v0.2.11 https://github.com/sgl-project/sglang.git
 cd sglang
 
 pip install --upgrade pip
 pip install -e "python[all]"
 
 # Install FlashInfer CUDA kernels
-pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
+pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/
 ```
 
 ### Method 3: Using docker
@@ -136,7 +136,7 @@ response = client.chat.completions.create(
 print(response)
 ```
 
-It supports streaming, vision, and most features of the Chat/Completions/Models endpoints specified by the [OpenAI API Reference](https://platform.openai.com/docs/api-reference/).
+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/).
 
 ### Additional Server Arguments
 - Add `--tp 2` to enable tensor parallelism. If it indicates `peer access is not supported between these two devices`, add `--enable-p2p-check` option.
@@ -147,10 +147,14 @@ 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 --port 30000 --dp 2 --tp 2
 ```
-- If you see out-of-memory errors during serving, please try to reduce the memory usage of the KV cache pool by setting a smaller value of `--mem-fraction-static`. The default value is `0.9`
+- If you see out-of-memory errors during serving, please try to reduce the memory usage of the KV cache pool by setting a smaller value of `--mem-fraction-static`. The default value is `0.9`.
 ```
 python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --mem-fraction-static 0.7
 ```
+- If you see out-of-memory errors during prefill for long prompts on a model that supports long context, consider using chunked prefill.
+```
+python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --port 30000 --chunked-prefill-size 8192
+```
 - See [hyperparameter_tuning.md](docs/en/hyperparameter_tuning.md) on tuning hyperparameters for better performance.
 - Add `--nnodes 2` to run tensor parallelism on multiple nodes. 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.
 ```
@@ -164,22 +168,6 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 - To enable fp8 quantization, you can add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.
 - To enable experimental torch.compile support, you can add `--enable-torch-compile`. It accelerates small models on small batch sizes.
 
-### Run Llama 3.1 405B
-
-```bash
-## Run 405B (fp8) on a single node
-python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct-FP8 --tp 8
-
-## Run 405B (fp16) on two nodes
-# replace the `172.16.4.52:20000` with your own first node ip address and port, disable CUDA Graph temporarily
-
-# on the first node
-GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 0 --disable-cuda-graph --mem-frac 0.75
-
-# on the second
-GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 1 --disable-cuda-graph --mem-frac 0.75
-```
-
 ### Supported Models
 
 - Llama / Llama 2 / Llama 3 / Llama 3.1
@@ -205,9 +193,25 @@ GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/
 
 Instructions for supporting a new model are [here](https://github.com/sgl-project/sglang/blob/main/docs/en/model_support.md).
 
+### Run Llama 3.1 405B
+
+```bash
+## Run 405B (fp8) on a single node
+python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct-FP8 --tp 8
+
+## Run 405B (fp16) on two nodes
+# replace the `172.16.4.52:20000` with your own first node ip address and port, disable CUDA Graph temporarily
+
+# on the first node
+GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 0 --disable-cuda-graph --mem-frac 0.75
+
+# on the second
+GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 1 --disable-cuda-graph --mem-frac 0.75
+```
+
 ### Benchmark Performance
 
-- Benchmark a single static batch by running the following command without launching a server. The arguments are the same as those for `launch_server.py`. This is not a dynamic batching server, so it may run out of memory for a batch size that can run successfully with a real server. This is because a real server will truncate the prefill into several batches/chunks, while this unit test does not do this.
+- Benchmark a single static batch by running the following command without launching a server. The arguments are the same as for `launch_server.py`. Note that this is not a dynamic batching server, so it may run out of memory for a batch size that a real server can handle. A real server truncates the prefill into several batches, while this unit test does not. For accurate large batch testing, consider using `sglang.bench_serving`.
   ```
   python -m sglang.bench_latency --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch 32 --input-len 256 --output-len 32
   ```

{sglang-0.2.9.post1 → sglang-0.2.11}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sglang"
-version = "0.2.9.post1"
+version = "0.2.11"
 description = "SGLang is yet another fast serving framework for large language models and vision language models."
 readme = "README.md"
 requires-python = ">=3.8"
@@ -20,12 +20,16 @@ dependencies = [
 ]
 
 [project.optional-dependencies]
-srt = ["aiohttp", "fastapi", "hf_transfer", "huggingface_hub", "interegular", "packaging", "pillow",
-       "psutil", "pydantic", "torch", "uvicorn", "uvloop", "zmq", "vllm==0.5.3.post1", "outlines>=0.0.44", "python-multipart"]
+srt = ["aiohttp", "fastapi", "hf_transfer", "huggingface_hub", "interegular",
+       "packaging", "pillow", "psutil", "pydantic", "python-multipart",
+       "torch", "uvicorn", "uvloop", "zmq",
+       "vllm==0.5.4", "outlines>=0.0.44"]
 openai = ["openai>=1.0", "tiktoken"]
 anthropic = ["anthropic>=0.20.0"]
 litellm = ["litellm>=1.0.0"]
+test = ["jsonlines", "matplotlib", "pandas"]
 all = ["sglang[srt]", "sglang[openai]", "sglang[anthropic]", "sglang[litellm]"]
+dev = ["sglang[all]", "sglang[test]"]
 
 [project.urls]
 "Homepage" = "https://github.com/sgl-project/sglang"

{sglang-0.2.9.post1 → sglang-0.2.11}/sglang/__init__.py

@@ -22,6 +22,11 @@ from sglang.api import (
     user_end,
     video,
 )
+from sglang.lang.choices import (
+    greedy_token_selection,
+    token_length_normalized,
+    unconditional_likelihood_normalized,
+)
 
 # SGLang DSL APIs
 __all__ = [
@@ -45,6 +50,9 @@ __all__ = [
     "user_begin",
     "user_end",
     "video",
+    "greedy_token_selection",
+    "token_length_normalized",
+    "unconditional_likelihood_normalized",
 ]
 
 # Global Configurations

{sglang-0.2.9.post1 → sglang-0.2.11}/sglang/api.py

@@ -6,6 +6,7 @@ from typing import Callable, List, Optional, Union
 
 from sglang.global_config import global_config
 from sglang.lang.backend.base_backend import BaseBackend
+from sglang.lang.choices import ChoicesSamplingMethod, token_length_normalized
 from sglang.lang.ir import (
     SglExpr,
     SglExprList,
@@ -73,12 +74,18 @@ def gen(
     return_text_in_logprobs: Optional[bool] = None,
     dtype: Optional[type] = None,
     choices: Optional[List[str]] = None,
+    choices_method: Optional[ChoicesSamplingMethod] = None,
     regex: Optional[str] = None,
 ):
     """Call the model to generate. See the meaning of the arguments in docs/en/sampling_params.md"""
 
     if choices:
-        return SglSelect(name, choices, 0.0 if temperature is None else temperature)
+        return SglSelect(
+            name,
+            choices,
+            0.0 if temperature is None else temperature,
+            token_length_normalized if choices_method is None else choices_method,
+        )
 
     # check regex is valid
     if regex is not None:
@@ -186,9 +193,10 @@ def select(
     name: Optional[str] = None,
     choices: Optional[List[str]] = None,
     temperature: float = 0.0,
+    choices_method: ChoicesSamplingMethod = token_length_normalized,
 ):
     assert choices is not None
-    return SglSelect(name, choices, temperature)
+    return SglSelect(name, choices, temperature, choices_method)
 
 
 def _role_common(name: str, expr: Optional[SglExpr] = None):