sglang 0.4.1.post6__tar.gz → 0.4.2__tar.gz

{sglang-0.4.1.post6/sglang.egg-info → sglang-0.4.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: sglang
-Version: 0.4.1.post6
+Version: 0.4.2
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -236,13 +236,13 @@ Requires-Dist: pyzmq>=25.1.2; extra == "runtime-common"
 Requires-Dist: torchao>=0.7.0; extra == "runtime-common"
 Requires-Dist: uvicorn; extra == "runtime-common"
 Requires-Dist: uvloop; extra == "runtime-common"
-Requires-Dist: xgrammar>=0.1.6; extra == "runtime-common"
+Requires-Dist: xgrammar>=0.1.10; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"
 Requires-Dist: cuda-python; extra == "srt"
-Requires-Dist: sgl-kernel>=0.0.2.post12; extra == "srt"
+Requires-Dist: sgl-kernel>=0.0.3; extra == "srt"
 Requires-Dist: torch; extra == "srt"
-Requires-Dist: vllm<=0.6.4.post1,>=0.6.3.post1; extra == "srt"
+Requires-Dist: vllm==0.6.4.post1; extra == "srt"
 Requires-Dist: flashinfer==0.1.6; extra == "srt"
 Provides-Extra: srt-hip
 Requires-Dist: sglang[runtime_common]; extra == "srt-hip"
@@ -252,6 +252,9 @@ Provides-Extra: srt-xpu
 Requires-Dist: sglang[runtime_common]; extra == "srt-xpu"
 Provides-Extra: srt-hpu
 Requires-Dist: sglang[runtime_common]; extra == "srt-hpu"
+Provides-Extra: srt-cpu
+Requires-Dist: sglang[runtime_common]; extra == "srt-cpu"
+Requires-Dist: torch; extra == "srt-cpu"
 Provides-Extra: openai
 Requires-Dist: openai>=1.0; extra == "openai"
 Requires-Dist: tiktoken; extra == "openai"
@@ -288,6 +291,11 @@ Requires-Dist: sglang[srt_hpu]; extra == "all-hpu"
 Requires-Dist: sglang[openai]; extra == "all-hpu"
 Requires-Dist: sglang[anthropic]; extra == "all-hpu"
 Requires-Dist: sglang[litellm]; extra == "all-hpu"
+Provides-Extra: all-cpu
+Requires-Dist: sglang[srt_cpu]; extra == "all-cpu"
+Requires-Dist: sglang[openai]; extra == "all-cpu"
+Requires-Dist: sglang[anthropic]; extra == "all-cpu"
+Requires-Dist: sglang[litellm]; extra == "all-cpu"
 Provides-Extra: dev
 Requires-Dist: sglang[all]; extra == "dev"
 Requires-Dist: sglang[test]; extra == "dev"
@@ -300,6 +308,9 @@ Requires-Dist: sglang[test]; extra == "dev-xpu"
 Provides-Extra: dev-hpu
 Requires-Dist: sglang[all_hpu]; extra == "dev-hpu"
 Requires-Dist: sglang[test]; extra == "dev-hpu"
+Provides-Extra: dev-cpu
+Requires-Dist: sglang[all_cpu]; extra == "dev-cpu"
+Requires-Dist: sglang[test]; extra == "dev-cpu"
 
 <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>
@@ -322,16 +333,16 @@ Requires-Dist: sglang[test]; extra == "dev-hpu"
 | [**Slides**](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#slides) |
 
 ## News
-- [2024/12] 🔥 SGLang v0.4: Zero-Overhead Batch Scheduler, Cache-Aware Load Balancer, Faster Structured Outputs ([blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)).
-- [2024/10] 🔥 The First SGLang Online Meetup ([slides](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#the-first-sglang-online-meetup)).
-- [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/)).
-- [2024/07] Faster Llama3 Serving with SGLang Runtime (vs. TensorRT-LLM, vLLM) ([blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/)).
+- [2025/01] 🔥 SGLang provides day one support for DeepSeek V3/R1 models on NVIDIA and AMD GPUs with DeekSeek-specific optimizations. ([instructions](https://github.com/sgl-project/sglang/tree/main/benchmark/deepseek_v3), [AMD blog](https://www.amd.com/en/developer/resources/technical-articles/amd-instinct-gpus-power-deepseek-v3-revolutionizing-ai-development-with-sglang.html))
+- [2024/12] 🔥 v0.4 Release: Zero-Overhead Batch Scheduler, Cache-Aware Load Balancer, Faster Structured Outputs ([blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)).
+- [2024/09] 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/)).
+- [2024/07] v0.2 Release: Faster Llama3 Serving with SGLang Runtime (vs. TensorRT-LLM, vLLM) ([blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/)).
 
 <details>
 <summary>More</summary>
 
+- [2024/10] The First SGLang Online Meetup ([slides](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#the-first-sglang-online-meetup)).
 - [2024/02] SGLang enables **3x faster JSON decoding** with compressed finite state machine ([blog](https://lmsys.org/blog/2024-02-05-compressed-fsm/)).
-- [2024/04] SGLang is used by the official **LLaVA-NeXT (video)** release ([blog](https://llava-vl.github.io/blog/2024-04-30-llava-next-video/)).
 - [2024/01] SGLang provides up to **5x faster inference** with RadixAttention ([blog](https://lmsys.org/blog/2024-01-17-sglang/)).
 - [2024/01] SGLang powers the serving of the official **LLaVA v1.6** release demo ([usage](https://github.com/haotian-liu/LLaVA?tab=readme-ov-file#demo)).
 
@@ -361,7 +372,7 @@ Learn more in the release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-s
 [Development Roadmap (2024 Q4)](https://github.com/sgl-project/sglang/issues/1487)
 
 ## Adoption and Sponsorship
-The project is supported by (alphabetically): AMD, Baseten, DataCrunch, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, LMSYS.org, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, UCLA, xAI, 01.AI.
+The project is supported by (alphabetically): AMD, Baseten, Cursor, DataCrunch, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, LMSYS.org, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, UCLA, xAI, 01.AI.
 
 ## Acknowledgment and Citation
 We learned 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). Please cite the paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.

{sglang-0.4.1.post6 → sglang-0.4.2}/README.md

@@ -19,16 +19,16 @@
 | [**Slides**](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#slides) |
 
 ## News
-- [2024/12] 🔥 SGLang v0.4: Zero-Overhead Batch Scheduler, Cache-Aware Load Balancer, Faster Structured Outputs ([blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)).
-- [2024/10] 🔥 The First SGLang Online Meetup ([slides](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#the-first-sglang-online-meetup)).
-- [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/)).
-- [2024/07] Faster Llama3 Serving with SGLang Runtime (vs. TensorRT-LLM, vLLM) ([blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/)).
+- [2025/01] 🔥 SGLang provides day one support for DeepSeek V3/R1 models on NVIDIA and AMD GPUs with DeekSeek-specific optimizations. ([instructions](https://github.com/sgl-project/sglang/tree/main/benchmark/deepseek_v3), [AMD blog](https://www.amd.com/en/developer/resources/technical-articles/amd-instinct-gpus-power-deepseek-v3-revolutionizing-ai-development-with-sglang.html))
+- [2024/12] 🔥 v0.4 Release: Zero-Overhead Batch Scheduler, Cache-Aware Load Balancer, Faster Structured Outputs ([blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)).
+- [2024/09] 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/)).
+- [2024/07] v0.2 Release: Faster Llama3 Serving with SGLang Runtime (vs. TensorRT-LLM, vLLM) ([blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/)).
 
 <details>
 <summary>More</summary>
 
+- [2024/10] The First SGLang Online Meetup ([slides](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#the-first-sglang-online-meetup)).
 - [2024/02] SGLang enables **3x faster JSON decoding** with compressed finite state machine ([blog](https://lmsys.org/blog/2024-02-05-compressed-fsm/)).
-- [2024/04] SGLang is used by the official **LLaVA-NeXT (video)** release ([blog](https://llava-vl.github.io/blog/2024-04-30-llava-next-video/)).
 - [2024/01] SGLang provides up to **5x faster inference** with RadixAttention ([blog](https://lmsys.org/blog/2024-01-17-sglang/)).
 - [2024/01] SGLang powers the serving of the official **LLaVA v1.6** release demo ([usage](https://github.com/haotian-liu/LLaVA?tab=readme-ov-file#demo)).
 
@@ -58,7 +58,7 @@ Learn more in the release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-s
 [Development Roadmap (2024 Q4)](https://github.com/sgl-project/sglang/issues/1487)
 
 ## Adoption and Sponsorship
-The project is supported by (alphabetically): AMD, Baseten, DataCrunch, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, LMSYS.org, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, UCLA, xAI, 01.AI.
+The project is supported by (alphabetically): AMD, Baseten, Cursor, DataCrunch, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, LMSYS.org, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, UCLA, xAI, 01.AI.
 
 ## Acknowledgment and Citation
 We learned 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). Please cite the paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.

{sglang-0.4.1.post6 → sglang-0.4.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sglang"
-version = "0.4.1.post6"
+version = "0.4.2"
 description = "SGLang is yet another fast serving framework for large language models and vision language models."
 readme = "README.md"
 requires-python = ">=3.8"
@@ -23,11 +23,11 @@ runtime_common = [
     "packaging", "pillow", "prometheus-client>=0.20.0",
     "psutil", "pydantic", "python-multipart",
     "pyzmq>=25.1.2", "torchao>=0.7.0", "uvicorn", "uvloop",
-    "xgrammar>=0.1.6"
+    "xgrammar>=0.1.10"
 ]
 srt = [
     "sglang[runtime_common]", "cuda-python",
-    "sgl-kernel>=0.0.2.post12", "torch", "vllm>=0.6.3.post1,<=0.6.4.post1",
+    "sgl-kernel>=0.0.3", "torch", "vllm==0.6.4.post1",
     "flashinfer==0.1.6"
 ]
 
@@ -40,6 +40,10 @@ srt_xpu = ["sglang[runtime_common]"]
 #For Intel Gaudi(device : hpu) follow the installation guide
 #https://docs.vllm.ai/en/latest/getting_started/gaudi-installation.html
 srt_hpu = ["sglang[runtime_common]"]
+# CPU: currently, there are no pre-built vllm wheels for CPU.
+# To install vllm for CPU, please follow the instruction here:
+# https://docs.vllm.ai/en/latest/getting_started/installation/cpu/index.html
+srt_cpu = ["sglang[runtime_common]", "torch"]
 
 openai = ["openai>=1.0", "tiktoken"]
 anthropic = ["anthropic>=0.20.0"]
@@ -57,11 +61,13 @@ all = ["sglang[srt]", "sglang[openai]", "sglang[anthropic]", "sglang[litellm]"]
 all_hip = ["sglang[srt_hip]", "sglang[openai]", "sglang[anthropic]", "sglang[litellm]"]
 all_xpu = ["sglang[srt_xpu]", "sglang[openai]", "sglang[anthropic]", "sglang[litellm]"]
 all_hpu = ["sglang[srt_hpu]", "sglang[openai]", "sglang[anthropic]", "sglang[litellm]"]
+all_cpu = ["sglang[srt_cpu]", "sglang[openai]", "sglang[anthropic]", "sglang[litellm]"]
 
 dev = ["sglang[all]", "sglang[test]"]
 dev_hip = ["sglang[all_hip]", "sglang[test]"]
 dev_xpu = ["sglang[all_xpu]", "sglang[test]"]
 dev_hpu = ["sglang[all_hpu]", "sglang[test]"]
+dev_cpu = ["sglang[all_cpu]", "sglang[test]"]
 
 [project.urls]
 "Homepage" = "https://github.com/sgl-project/sglang"

{sglang-0.4.1.post6 → sglang-0.4.2}/sglang/__init__.py

@@ -1,5 +1,6 @@
-# SGL API Components
+# SGLang public APIs
 
+# Frontend Language APIs
 from sglang.api import (
     Engine,
     Runtime,
@@ -23,16 +24,26 @@ from sglang.api import (
     user_end,
     video,
 )
+from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
 from sglang.lang.choices import (
     greedy_token_selection,
     token_length_normalized,
     unconditional_likelihood_normalized,
 )
+from sglang.utils import LazyImport
+
+Anthropic = LazyImport("sglang.lang.backend.anthropic", "Anthropic")
+LiteLLM = LazyImport("sglang.lang.backend.litellm", "LiteLLM")
+OpenAI = LazyImport("sglang.lang.backend.openai", "OpenAI")
+VertexAI = LazyImport("sglang.lang.backend.vertexai", "VertexAI")
+
+# Other configs
+from sglang.global_config import global_config
+from sglang.version import __version__
 
-# SGLang DSL APIs
 __all__ = [
-    "Runtime",
     "Engine",
+    "Runtime",
     "assistant",
     "assistant_begin",
     "assistant_end",
@@ -52,27 +63,14 @@ __all__ = [
     "user_begin",
     "user_end",
     "video",
+    "RuntimeEndpoint",
     "greedy_token_selection",
     "token_length_normalized",
     "unconditional_likelihood_normalized",
+    "Anthropic",
+    "LiteLLM",
+    "OpenAI",
+    "VertexAI",
+    "global_config",
+    "__version__",
 ]
-
-# Global Configurations
-from sglang.global_config import global_config
-
-__all__ += ["global_config"]
-
-from sglang.version import __version__
-
-__all__ += ["__version__"]
-
-# SGLang Backends
-from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
-from sglang.utils import LazyImport
-
-Anthropic = LazyImport("sglang.lang.backend.anthropic", "Anthropic")
-LiteLLM = LazyImport("sglang.lang.backend.litellm", "LiteLLM")
-OpenAI = LazyImport("sglang.lang.backend.openai", "OpenAI")
-VertexAI = LazyImport("sglang.lang.backend.vertexai", "VertexAI")
-
-__all__ += ["Anthropic", "LiteLLM", "OpenAI", "VertexAI", "RuntimeEndpoint"]

{sglang-0.4.1.post6 → sglang-0.4.2}/sglang/api.py

@@ -1,6 +1,5 @@
 """Public APIs of the language."""
 
-import os
 import re
 from typing import Callable, List, Optional, Union
 
@@ -33,19 +32,15 @@ def function(
 
 
 def Runtime(*args, **kwargs):
-    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
-
     # Avoid importing unnecessary dependency
-    from sglang.srt.server import Runtime
+    from sglang.lang.backend.runtime_endpoint 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
+    from sglang.srt.entrypoints.engine import Engine
 
     return Engine(*args, **kwargs)
 

{sglang-0.4.1.post6 → sglang-0.4.2}/sglang/bench_offline_throughput.py

@@ -27,7 +27,8 @@ from sglang.bench_serving import (
     sample_random_requests,
     set_ulimit,
 )
-from sglang.srt.server import Engine, Runtime
+from sglang.lang.backend.runtime_endpoint import Runtime
+from sglang.srt.entrypoints.engine import Engine
 from sglang.srt.server_args import ServerArgs
 
 
@@ -39,20 +40,22 @@ class BenchArgs:
     dataset_path: str = ""
     num_prompts: int = 1000
     sharegpt_output_len: Optional[int] = None
+    sharegpt_context_len: Optional[int] = None
     random_input_len: int = 1024
     random_output_len: int = 1024
     random_range_ratio: float = 0.0
-    gen_num_groups: int = 64
-    gen_prompts_per_group: int = 16
-    gen_system_prompt_len: int = 2048
-    gen_question_len: int = 128
-    gen_output_len: int = 256
+    gsp_num_groups: int = 64
+    gsp_prompts_per_group: int = 16
+    gsp_system_prompt_len: int = 2048
+    gsp_question_len: int = 128
+    gsp_output_len: int = 256
+    seed: int = 1
     disable_ignore_eos: bool = False
     extra_request_body: Optional[str] = None
-    seed: int = 1
+    apply_chat_template: bool = False
+    profile: bool = False
     skip_warmup: bool = False
     do_not_exit: bool = False
-    profile: bool = False
 
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser):
@@ -82,6 +85,12 @@ class BenchArgs:
             default=BenchArgs.sharegpt_output_len,
             help="Output length for each request. Overrides the output length from the ShareGPT dataset.",
         )
+        parser.add_argument(
+            "--sharegpt-context-len",
+            type=int,
+            default=BenchArgs.sharegpt_context_len,
+            help="The context length of the model for the ShareGPT dataset. Requests longer than the context length will be dropped.",
+        )
         parser.add_argument(
             "--random-input-len",
             type=int,
@@ -102,51 +111,62 @@ class BenchArgs:
             "used only for random dataset.",
         )
         parser.add_argument(
-            "--gen-num-groups",
+            "--gsp-num-groups",
             type=int,
-            default=BenchArgs.gen_num_groups,
+            default=BenchArgs.gsp_num_groups,
             help="Number of groups with shared prefix, used"
             "only for generate-shared-prefix",
         )
         parser.add_argument(
-            "--gen-prompts-per-group",
+            "--gsp-prompts-per-group",
             type=int,
-            default=BenchArgs.gen_prompts_per_group,
+            default=BenchArgs.gsp_prompts_per_group,
             help="Number of prompts per group of shared prefix, used"
             "only for generate-shared-prefix",
         )
         parser.add_argument(
-            "--gen-system-prompt-len",
+            "--gsp-system-prompt-len",
             type=int,
-            default=BenchArgs.gen_system_prompt_len,
+            default=BenchArgs.gsp_system_prompt_len,
             help="System prompt length, used" "only for generate-shared-prefix",
         )
         parser.add_argument(
-            "--gen-question-len",
+            "--gsp-question-len",
             type=int,
-            default=BenchArgs.gen_question_len,
+            default=BenchArgs.gsp_question_len,
             help="Question length, used" "only for generate-shared-prefix",
         )
         parser.add_argument(
-            "--gen-output-len",
+            "--gsp-output-len",
             type=int,
-            default=BenchArgs.gen_output_len,
+            default=BenchArgs.gsp_output_len,
             help="Target length in tokens for outputs in generated-shared-prefix dataset",
         )
+        parser.add_argument("--seed", type=int, default=1, help="The random seed.")
         parser.add_argument(
             "--disable-ignore-eos",
-            type=bool,
-            default=BenchArgs.disable_ignore_eos,
+            action="store_true",
             help="Disable ignore EOS token",
         )
         parser.add_argument(
             "--extra-request-body",
             metavar='{"key1": "value1", "key2": "value2"}',
             type=str,
+            default=BenchArgs.extra_request_body,
             help="Append given JSON object to the request payload. You can use this to specify"
             "additional generate params like sampling params.",
         )
-        parser.add_argument("--seed", type=int, default=1, help="The random seed.")
+        parser.add_argument(
+            "--apply-chat-template",
+            action="store_true",
+            help="Apply chat template",
+        )
+        parser.add_argument(
+            "--profile",
+            action="store_true",
+            help="Use Torch Profiler. The endpoint must be launched with "
+            "SGLANG_TORCH_PROFILER_DIR to enable profiler.",
+        )
         parser.add_argument(
             "--skip-warmup",
             action="store_true",
@@ -157,12 +177,6 @@ class BenchArgs:
             action="store_true",
             help="Do not exit the program. This is useful for nsys profile with --duration and --delay.",
         )
-        parser.add_argument(
-            "--profile",
-            action="store_true",
-            help="Use Torch Profiler. The endpoint must be launched with "
-            "SGLANG_TORCH_PROFILER_DIR to enable profiler.",
-        )
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):

{sglang-0.4.1.post6 → sglang-0.4.2}/sglang/bench_one_batch.py

@@ -9,7 +9,8 @@ It accepts server arguments (the same as launch_server.py) and benchmark argumen
 python -m sglang.bench_one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --load-format dummy
 ## sweep through multiple data points and store (append) the results in a jsonl file:
 python -m sglang.bench_one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch 1 12 14 --input-len 256 512 --output-len 32 256 --run-name test_run
-
+## run with profiling:
+python -m sglang.bench_one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch 1 12 14 --input-len 256 512 --profile
 # Usage (correctness test):
 python -m sglang.bench_one_batch --model-path TinyLlama/TinyLlama-1.1B-Chat-v0.4 --correct
 
@@ -56,15 +57,21 @@ import torch
 import torch.distributed as dist
 
 from sglang.srt.configs.model_config import ModelConfig
+from sglang.srt.entrypoints.engine import _set_envs_and_config
 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 PortArgs, ServerArgs
 from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
-from sglang.srt.utils import configure_logger, kill_process_tree, suppress_other_loggers
+from sglang.srt.utils import (
+    configure_logger,
+    get_bool_env_var,
+    kill_process_tree,
+    set_gpu_proc_affinity,
+    suppress_other_loggers,
+)
 
 
 @dataclasses.dataclass
@@ -77,6 +84,8 @@ class BenchArgs:
     correctness_test: bool = False
     # This is only used for correctness test
     cut_len: int = 4
+    profile: bool = False
+    profile_filename_prefix: str = "profile"
 
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser):
@@ -95,6 +104,16 @@ class BenchArgs:
         )
         parser.add_argument("--correctness-test", action="store_true")
         parser.add_argument("--cut-len", type=int, default=BenchArgs.cut_len)
+        parser.add_argument(
+            "--profile", action="store_true", help="Use Torch Profiler."
+        )
+        parser.add_argument(
+            "--profile-filename-prefix",
+            type=str,
+            default=BenchArgs.profile_filename_prefix,
+            help="Prefix of the profiling file names. The full profiling result file(s) be "
+            '"[profile_filename_prefix]_batch[batch_size]_input[input_len]_output[output_len].trace.json.gz"',
+        )
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):
@@ -216,6 +235,7 @@ def extend(reqs, model_runner):
         model_config=model_runner.model_config,
         enable_overlap=False,
         spec_algorithm=SpeculativeAlgorithm.NONE,
+        enable_custom_logit_processor=False,
     )
     batch.prepare_for_extend()
     model_worker_batch = batch.get_model_worker_batch()
@@ -286,7 +306,16 @@ def synchronize(device):
 
 
 def latency_test_run_once(
-    run_name, model_runner, rank_print, reqs, batch_size, input_len, output_len, device
+    run_name,
+    model_runner,
+    rank_print,
+    reqs,
+    batch_size,
+    input_len,
+    output_len,
+    device,
+    profile,
+    profile_filename_prefix,
 ):
     max_batch_size = model_runner.max_total_num_tokens // (input_len + output_len)
     if batch_size > max_batch_size:
@@ -308,6 +337,17 @@ def latency_test_run_once(
 
     tot_latency = 0
 
+    profiler = None
+    if profile:
+        profiler = torch.profiler.profile(
+            activities=[
+                torch.profiler.ProfilerActivity.CPU,
+                torch.profiler.ProfilerActivity.CUDA,
+            ],
+            with_stack=True,
+        )
+        profiler.start()
+
     # Prefill
     synchronize(device)
     tic = time.time()
@@ -338,6 +378,14 @@ def latency_test_run_once(
                 f"Decode.  latency: {latency:6.5f} s, throughput: {throughput:9.2f} token/s"
             )
 
+    if profile:
+        profiler.stop()
+        profile_filename = f"{profile_filename_prefix}_batch{batch_size}_input{input_len}_output{output_len}.trace.json.gz"
+        parent_dir = os.path.dirname(os.path.abspath(profile_filename))
+        os.makedirs(parent_dir, exist_ok=True)
+        profiler.export_chrome_trace(profile_filename)
+        rank_print(f"torch profiler chrome trace saved to {profile_filename}")
+
     # Record decode timing from 2nd output
     if output_len > 1:
         med_decode_latency = np.median(decode_latencies)
@@ -363,6 +411,10 @@ def latency_test(
     bench_args,
     tp_rank,
 ):
+    # Set CPU affinity
+    if get_bool_env_var("SGLANG_SET_CPU_AFFINITY"):
+        set_gpu_proc_affinity(server_args.tp_size, server_args.nnodes, tp_rank)
+
     # Configure the logger
     configure_logger(server_args, prefix=f" TP{tp_rank}")
     rank_print = print if tp_rank == 0 else lambda *args, **kwargs: None
@@ -386,6 +438,8 @@ def latency_test(
         bench_args.input_len[0],
         8,  # shorter decoding to speed up the warmup
         server_args.device,
+        profile=False,
+        profile_filename_prefix="",  # not used
     )
 
     rank_print("Benchmark ...")
@@ -405,6 +459,8 @@ def latency_test(
             il,
             ol,
             server_args.device,
+            bench_args.profile if tp_rank == 0 else None,
+            bench_args.profile_filename_prefix,
         )
         if ret is not None:
             result_list.append(ret)

{sglang-0.4.1.post6 → sglang-0.4.2}/sglang/bench_one_batch_server.py

@@ -22,7 +22,7 @@ from typing import Tuple
 import numpy as np
 import requests
 
-from sglang.srt.server import launch_server
+from sglang.srt.entrypoints.http_server import launch_server
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import kill_process_tree