sglang 0.4.0__tar.gz → 0.4.0.post2__tar.gz

{sglang-0.4.0 → sglang-0.4.0.post2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.4.0
+Version: 0.4.0.post2
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -215,6 +215,7 @@ Requires-Dist: requests
 Requires-Dist: tqdm
 Requires-Dist: numpy
 Requires-Dist: IPython
+Requires-Dist: setproctitle
 Provides-Extra: runtime-common
 Requires-Dist: aiohttp; extra == "runtime-common"
 Requires-Dist: decord; extra == "runtime-common"
@@ -232,16 +233,17 @@ Requires-Dist: psutil; extra == "runtime-common"
 Requires-Dist: pydantic; extra == "runtime-common"
 Requires-Dist: python-multipart; extra == "runtime-common"
 Requires-Dist: pyzmq>=25.1.2; extra == "runtime-common"
-Requires-Dist: torchao; extra == "runtime-common"
+Requires-Dist: torchao>=0.7.0; extra == "runtime-common"
+Requires-Dist: gemlite; extra == "runtime-common"
 Requires-Dist: uvicorn; extra == "runtime-common"
 Requires-Dist: uvloop; extra == "runtime-common"
-Requires-Dist: xgrammar>=0.1.4; extra == "runtime-common"
+Requires-Dist: xgrammar>=0.1.6; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"
 Requires-Dist: torch; extra == "srt"
-Requires-Dist: vllm>=0.6.3.post1; extra == "srt"
+Requires-Dist: vllm<=0.6.4.post1,>=0.6.3.post1; extra == "srt"
 Requires-Dist: cuda-python; extra == "srt"
-Requires-Dist: flashinfer>=0.1.6; extra == "srt"
+Requires-Dist: flashinfer==0.1.6; extra == "srt"
 Provides-Extra: srt-hip
 Requires-Dist: sglang[runtime_common]; extra == "srt-hip"
 Requires-Dist: torch; extra == "srt-hip"
@@ -311,10 +313,14 @@ Requires-Dist: sglang[test]; extra == "dev-hpu"
 
 --------------------------------------------------------------------------------
 
-| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Documentation**](https://sgl-project.github.io/) | [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2tmmp6flg-89dOlJW2TjnBrTRk1I_~GA) |
-[**Join Bi-Weekly Development Meeting**](https://docs.google.com/document/d/1xEow4eIM152xNcRxqZz9VEcOiTQo8-CEuuQ5qTmkt-E/edit?usp=sharing) | [**Slides**](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#slides) |
+| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/)
+| [**Documentation**](https://sgl-project.github.io/)
+| [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2tmmp6flg-89dOlJW2TjnBrTRk1I_~GA)
+| [**Join Bi-Weekly Development Meeting**](https://docs.google.com/document/d/1xEow4eIM152xNcRxqZz9VEcOiTQo8-CEuuQ5qTmkt-E/edit?usp=sharing)
+| [**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/)).
@@ -346,13 +352,13 @@ The core features include:
 - [Frontend: Structured Generation Language (SGLang)](https://sgl-project.github.io/frontend/frontend.html)
 
 ## Benchmark And Performance
-Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/), [v0.3 blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/)
+Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/), [v0.3 blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/), [v0.4 blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)
 
 ## Roadmap
 [Development Roadmap (2024 Q4)](https://github.com/sgl-project/sglang/issues/1487)
 
 ## Adoption and Sponsorship
-The project is supported by (alphabetically): AMD, Baseten, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, NVIDIA, RunPod, Stanford, UC Berkeley, xAI and 01.AI.
+The project is supported by (alphabetically): AMD, Baseten, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, xAI and 01.AI.
 
 ## Acknowledgment and Citation
 We 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).

{sglang-0.4.0 → sglang-0.4.0.post2}/README.md

@@ -12,10 +12,14 @@
 
 --------------------------------------------------------------------------------
 
-| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Documentation**](https://sgl-project.github.io/) | [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2tmmp6flg-89dOlJW2TjnBrTRk1I_~GA) |
-[**Join Bi-Weekly Development Meeting**](https://docs.google.com/document/d/1xEow4eIM152xNcRxqZz9VEcOiTQo8-CEuuQ5qTmkt-E/edit?usp=sharing) | [**Slides**](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#slides) |
+| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/)
+| [**Documentation**](https://sgl-project.github.io/)
+| [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2tmmp6flg-89dOlJW2TjnBrTRk1I_~GA)
+| [**Join Bi-Weekly Development Meeting**](https://docs.google.com/document/d/1xEow4eIM152xNcRxqZz9VEcOiTQo8-CEuuQ5qTmkt-E/edit?usp=sharing)
+| [**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/)).
@@ -47,13 +51,13 @@ The core features include:
 - [Frontend: Structured Generation Language (SGLang)](https://sgl-project.github.io/frontend/frontend.html)
 
 ## Benchmark And Performance
-Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/), [v0.3 blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/)
+Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/), [v0.3 blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/), [v0.4 blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)
 
 ## Roadmap
 [Development Roadmap (2024 Q4)](https://github.com/sgl-project/sglang/issues/1487)
 
 ## Adoption and Sponsorship
-The project is supported by (alphabetically): AMD, Baseten, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, NVIDIA, RunPod, Stanford, UC Berkeley, xAI and 01.AI.
+The project is supported by (alphabetically): AMD, Baseten, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, xAI and 01.AI.
 
 ## Acknowledgment and Citation
 We 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).

{sglang-0.4.0 → sglang-0.4.0.post2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sglang"
-version = "0.4.0"
+version = "0.4.0.post2"
 description = "SGLang is yet another fast serving framework for large language models and vision language models."
 readme = "README.md"
 requires-python = ">=3.8"
@@ -13,7 +13,7 @@ classifiers = [
     "Programming Language :: Python :: 3",
     "License :: OSI Approved :: Apache Software License",
 ]
-dependencies = ["requests", "tqdm", "numpy", "IPython"]
+dependencies = ["requests", "tqdm", "numpy", "IPython", "setproctitle"]
 
 [project.optional-dependencies]
 runtime_common = ["aiohttp", "decord", "fastapi",
@@ -21,9 +21,9 @@ runtime_common = ["aiohttp", "decord", "fastapi",
     "orjson", "outlines>=0.0.44,<0.1.0",
     "packaging", "pillow", "prometheus-client>=0.20.0",
     "psutil", "pydantic", "python-multipart",
-    "pyzmq>=25.1.2", "torchao", "uvicorn", "uvloop",
-    "xgrammar>=0.1.4"]
-srt = ["sglang[runtime_common]", "torch", "vllm>=0.6.3.post1", "cuda-python", "flashinfer>=0.1.6"]
+    "pyzmq>=25.1.2", "torchao>=0.7.0", "gemlite", "uvicorn", "uvloop",
+    "xgrammar>=0.1.6"]
+srt = ["sglang[runtime_common]", "torch", "vllm>=0.6.3.post1,<=0.6.4.post1", "cuda-python", "flashinfer==0.1.6"]
 
 # HIP (Heterogeneous-computing Interface for Portability) for AMD
 # => base docker rocm/vllm-dev:20241022, not from public vllm whl
@@ -33,7 +33,7 @@ srt_hip = ["sglang[runtime_common]", "torch", "vllm==0.6.3.dev13"]
 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]"]
+srt_hpu = ["sglang[runtime_common]"]
 
 openai = ["openai>=1.0", "tiktoken"]
 anthropic = ["anthropic>=0.20.0"]
@@ -50,6 +50,7 @@ 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]"]
+
 dev = ["sglang[all]", "sglang[test]"]
 dev_hip = ["sglang[all_hip]", "sglang[test]"]
 dev_xpu = ["sglang[all_xpu]", "sglang[test]"]

{sglang-0.4.0 → sglang-0.4.0.post2}/sglang/__init__.py

@@ -66,7 +66,7 @@ from sglang.version import __version__
 
 __all__ += ["__version__"]
 
-# SGL Backends
+# SGLang Backends
 from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
 from sglang.utils import LazyImport
 

{sglang-0.4.0 → sglang-0.4.0.post2}/sglang/bench_offline_throughput.py

@@ -201,18 +201,17 @@ def throughput_test_once(
         for r in reqs
     ]
 
-    st = time.perf_counter()
     if profile:
         backend.start_profile()
 
+    st = time.perf_counter()
     gen_out = backend.generate(prompt=prompt, sampling_params=sampling_params)
+    latency = time.perf_counter() - st
 
     if profile:
         backend.stop_profile()
         monitor_trace_file(os.getenv("SGLANG_TORCH_PROFILER_DIR"))
 
-    latency = time.perf_counter() - st
-
     if backend_name == "runtime":
         gen_out = json.loads(gen_out)
 
@@ -285,7 +284,7 @@ def throughput_test(
     else:
         raise ValueError('Please set backend to either "engine" or "runtime"')
 
-    tokenizer_id = server_args.model_path
+    tokenizer_id = server_args.tokenizer_path or server_args.model_path
     tokenizer = get_tokenizer(tokenizer_id)
 
     # Set global environmnets
@@ -304,8 +303,8 @@ def throughput_test(
     warmup_requests = sample_random_requests(
         input_len=256,
         output_len=16,
-        num_prompts=16,
-        range_ratio=0.8,
+        num_prompts=min(bench_args.num_prompts, 16),
+        range_ratio=1.0,
         tokenizer=tokenizer,
         dataset_path=bench_args.dataset_path,
     )
@@ -321,6 +320,19 @@ def throughput_test(
             extra_request_body=extra_request_body,
             profile=False,
         )
+        time.sleep(0.5)
+
+    try:
+        import os
+        import pwd
+
+        from gemlite.core import GemLiteLinearTriton
+
+        GemLiteLinearTriton.cache_config(
+            f"/tmp/{pwd.getpwuid(os.getuid()).pw_gecos}_gemlite.json"
+        )
+    except ImportError:
+        pass
 
     logging.info("\nBenchmark...")
     result = throughput_test_once(

{sglang-0.4.0 → sglang-0.4.0.post2}/sglang/bench_one_batch.py

@@ -385,6 +385,19 @@ def latency_test(
         8,  # shorter decoding to speed up the warmup
         server_args.device,
     )
+
+    try:
+        import os
+        import pwd
+
+        from gemlite.core import GemLiteLinearTriton
+
+        GemLiteLinearTriton.cache_config(
+            f"/tmp/{pwd.getpwuid(os.getuid()).pw_gecos}_gemlite.json"
+        )
+    except ImportError:
+        pass
+
     rank_print("Benchmark ...")
 
     # Run the sweep

{sglang-0.4.0 → sglang-0.4.0.post2}/sglang/bench_serving.py

@@ -321,6 +321,8 @@ async def async_request_sglang_generate(
             },
             "stream": not args.disable_stream,
             "lora_path": request_func_input.lora_name,
+            "return_logprob": args.return_logprob,
+            "logprob_start_len": -1,
             **request_func_input.extra_request_body,
         }
         headers = {}
@@ -911,7 +913,7 @@ async def benchmark(
         prompt=test_prompt,
         api_url=api_url,
         prompt_len=test_prompt_len,
-        output_len=test_output_len,
+        output_len=min(test_output_len, 32),
         lora_name=lora_name,
         extra_request_body=extra_request_body,
     )
@@ -1413,6 +1415,11 @@ if __name__ == "__main__":
         action="store_true",
         help="Disable ignoring EOS.",
     )
+    parser.add_argument(
+        "--return-logprob",
+        action="store_true",
+        help="Return logprob.",
+    )
     parser.add_argument(
         "--extra-request-body",
         metavar='{"key1": "value1", "key2": "value2"}',

sglang-0.4.0.post2/sglang/check_env.py

@@ -0,0 +1,305 @@
+"""Check environment configurations and dependency versions."""
+
+import importlib
+import os
+import resource
+import subprocess
+import sys
+from collections import OrderedDict, defaultdict
+
+import torch
+
+from sglang.srt.utils import is_hip
+
+
+def is_cuda_v2():
+    return torch.version.cuda is not None
+
+
+# List of packages to check versions
+PACKAGE_LIST = [
+    "sglang",
+    "flashinfer",
+    "triton",
+    "transformers",
+    "torchao",
+    "numpy",
+    "aiohttp",
+    "fastapi",
+    "hf_transfer",
+    "huggingface_hub",
+    "interegular",
+    "modelscope",
+    "orjson",
+    "outlines",
+    "packaging",
+    "psutil",
+    "pydantic",
+    "multipart",
+    "zmq",
+    "torchao",
+    "uvicorn",
+    "uvloop",
+    "vllm",
+    "xgrammar",
+    "openai",
+    "tiktoken",
+    "anthropic",
+    "litellm",
+    "decord",
+]
+
+
+def get_package_versions(packages):
+    """
+    Get versions of specified packages.
+    """
+    versions = {}
+    for package in packages:
+        package_name = package.split("==")[0].split(">=")[0].split("<=")[0]
+        try:
+            module = importlib.import_module(package_name)
+            if hasattr(module, "__version__"):
+                versions[package_name] = module.__version__
+        except ModuleNotFoundError:
+            versions[package_name] = "Module Not Found"
+    return versions
+
+
+def get_cuda_info():
+    """
+    Get CUDA-related information if available.
+    """
+    if is_cuda_v2():
+        cuda_info = {"CUDA available": torch.cuda.is_available()}
+
+        if cuda_info["CUDA available"]:
+            cuda_info.update(_get_gpu_info())
+            cuda_info.update(_get_cuda_version_info())
+
+        return cuda_info
+    elif is_hip():
+        cuda_info = {"ROCM available": torch.cuda.is_available()}
+
+        if cuda_info["ROCM available"]:
+            cuda_info.update(_get_gpu_info())
+            cuda_info.update(_get_cuda_version_info())
+
+        return cuda_info
+
+
+def _get_gpu_info():
+    """
+    Get information about available GPUs.
+    """
+    devices = defaultdict(list)
+    capabilities = defaultdict(list)
+    for k in range(torch.cuda.device_count()):
+        devices[torch.cuda.get_device_name(k)].append(str(k))
+        capability = torch.cuda.get_device_capability(k)
+        capabilities[f"{capability[0]}.{capability[1]}"].append(str(k))
+
+    gpu_info = {}
+    for name, device_ids in devices.items():
+        gpu_info[f"GPU {','.join(device_ids)}"] = name
+
+    if len(capabilities) == 1:
+        # All GPUs have the same compute capability
+        cap, gpu_ids = list(capabilities.items())[0]
+        gpu_info[f"GPU {','.join(gpu_ids)} Compute Capability"] = cap
+    else:
+        # GPUs have different compute capabilities
+        for cap, gpu_ids in capabilities.items():
+            gpu_info[f"GPU {','.join(gpu_ids)} Compute Capability"] = cap
+
+    return gpu_info
+
+
+def _get_cuda_version_info():
+    """
+    Get CUDA version information.
+    """
+    if is_cuda_v2():
+        from torch.utils.cpp_extension import CUDA_HOME
+
+        cuda_info = {"CUDA_HOME": CUDA_HOME}
+
+        if CUDA_HOME and os.path.isdir(CUDA_HOME):
+            cuda_info.update(_get_nvcc_info())
+            cuda_info.update(_get_cuda_driver_version())
+
+        return cuda_info
+    elif is_hip():
+        from torch.utils.cpp_extension import ROCM_HOME as ROCM_HOME
+
+        cuda_info = {"ROCM_HOME": ROCM_HOME}
+
+        if ROCM_HOME and os.path.isdir(ROCM_HOME):
+            cuda_info.update(_get_nvcc_info())
+            cuda_info.update(_get_cuda_driver_version())
+
+        return cuda_info
+    else:
+        cuda_info = {"CUDA_HOME": ""}
+        return cuda_info
+
+
+def _get_nvcc_info():
+    """
+    Get NVCC version information.
+    """
+    if is_cuda_v2():
+        from torch.utils.cpp_extension import CUDA_HOME
+
+        try:
+            nvcc = os.path.join(CUDA_HOME, "bin/nvcc")
+            nvcc_output = (
+                subprocess.check_output(f'"{nvcc}" -V', shell=True)
+                .decode("utf-8")
+                .strip()
+            )
+            return {
+                "NVCC": nvcc_output[
+                    nvcc_output.rfind("Cuda compilation tools") : nvcc_output.rfind(
+                        "Build"
+                    )
+                ].strip()
+            }
+        except subprocess.SubprocessError:
+            return {"NVCC": "Not Available"}
+    elif is_hip():
+        from torch.utils.cpp_extension import ROCM_HOME
+
+        try:
+            hipcc = os.path.join(ROCM_HOME, "bin/hipcc")
+            hipcc_output = (
+                subprocess.check_output(f'"{hipcc}" --version', shell=True)
+                .decode("utf-8")
+                .strip()
+            )
+            return {
+                "HIPCC": hipcc_output[
+                    hipcc_output.rfind("HIP version") : hipcc_output.rfind("AMD clang")
+                ].strip()
+            }
+        except subprocess.SubprocessError:
+            return {"HIPCC": "Not Available"}
+    else:
+        return {"NVCC": "Not Available"}
+
+
+def _get_cuda_driver_version():
+    """
+    Get CUDA driver version.
+    """
+    versions = set()
+    if is_cuda_v2():
+        try:
+            output = subprocess.check_output(
+                [
+                    "nvidia-smi",
+                    "--query-gpu=driver_version",
+                    "--format=csv,noheader,nounits",
+                ]
+            )
+            versions = set(output.decode().strip().split("\n"))
+            if len(versions) == 1:
+                return {"CUDA Driver Version": versions.pop()}
+            else:
+                return {"CUDA Driver Versions": ", ".join(sorted(versions))}
+        except subprocess.SubprocessError:
+            return {"CUDA Driver Version": "Not Available"}
+    elif is_hip():
+        try:
+            output = subprocess.check_output(
+                [
+                    "rocm-smi",
+                    "--showdriverversion",
+                    "--csv",
+                ]
+            )
+            versions = set(output.decode().strip().split("\n"))
+            versions.discard("name, value")
+            ver = versions.pop()
+            ver = ver.replace('"Driver version", ', "").replace('"', "")
+
+            return {"ROCM Driver Version": ver}
+        except subprocess.SubprocessError:
+            return {"ROCM Driver Version": "Not Available"}
+    else:
+        return {"CUDA Driver Version": "Not Available"}
+
+
+def get_gpu_topology():
+    """
+    Get GPU topology information.
+    """
+    if is_cuda_v2():
+        try:
+            result = subprocess.run(
+                ["nvidia-smi", "topo", "-m"],
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                text=True,
+                check=True,
+            )
+            return "\n" + result.stdout if result.returncode == 0 else None
+        except subprocess.SubprocessError:
+            return None
+    elif is_hip():
+        try:
+            result = subprocess.run(
+                ["rocm-smi", "--showtopotype"],
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                text=True,
+                check=True,
+            )
+            return "\n" + result.stdout if result.returncode == 0 else None
+        except subprocess.SubprocessError:
+            return None
+    else:
+        return None
+
+
+def get_hypervisor_vendor():
+    try:
+        output = subprocess.check_output(["lscpu"], text=True)
+        for line in output.split("\n"):
+            if "Hypervisor vendor:" in line:
+                return line.split(":")[1].strip()
+        return None
+    except:
+        return None
+
+
+def check_env():
+    """
+    Check and print environment information.
+    """
+    env_info = OrderedDict()
+    env_info["Python"] = sys.version.replace("\n", "")
+    env_info.update(get_cuda_info())
+    env_info["PyTorch"] = torch.__version__
+    env_info.update(get_package_versions(PACKAGE_LIST))
+
+    gpu_topo = get_gpu_topology()
+    if gpu_topo:
+        if is_cuda_v2():
+            env_info["NVIDIA Topology"] = gpu_topo
+        elif is_hip():
+            env_info["AMD Topology"] = gpu_topo
+
+    hypervisor_vendor = get_hypervisor_vendor()
+    if hypervisor_vendor:
+        env_info["Hypervisor vendor"] = hypervisor_vendor
+
+    ulimit_soft, _ = resource.getrlimit(resource.RLIMIT_NOFILE)
+    env_info["ulimit soft"] = ulimit_soft
+
+    for k, v in env_info.items():
+        print(f"{k}: {v}")
+
+
+if __name__ == "__main__":
+    check_env()

{sglang-0.4.0 → sglang-0.4.0.post2}/sglang/lang/backend/runtime_endpoint.py

@@ -55,6 +55,7 @@ class RuntimeEndpoint(BaseBackend):
             self.base_url + "/flush_cache",
             api_key=self.api_key,
             verify=self.verify,
+            method="POST",
         )
         self._assert_success(res)
 

{sglang-0.4.0 → sglang-0.4.0.post2}/sglang/lang/chat_template.py

@@ -320,6 +320,28 @@ register_chat_template(
     )
 )
 
+register_chat_template(
+    ChatTemplate(
+        name="granite-3-instruct",
+        default_system_prompt=None,
+        role_prefix_and_suffix={
+            "system": (
+                "<|start_of_role|>system<|end_of_role|>",
+                "<|end_of_text|>",
+            ),
+            "user": (
+                "<|start_of_role|>user<|end_of_role|>",
+                "<|end_of_text|>",
+            ),
+            "assistant": (
+                "<|start_of_role|>assistant<|end_of_role|>",
+                "<|end_of_text|>",
+            ),
+        },
+        stop_str=("<|end_of_text|>",),
+    )
+)
+
 
 @register_chat_template_matching_function
 def match_dbrx(model_path: str):
@@ -402,6 +424,16 @@ def match_c4ai_command_r(model_path: str):
         return get_chat_template("c4ai-command-r")
 
 
+@register_chat_template_matching_function
+def match_granite_instruct(model_path: str):
+    model_path = model_path.lower()
+    # When future versions of Granite are released, this code may
+    # need to be updated. For now, assume that the Granite 3.0
+    # template works across the board.
+    if "granite" in model_path and "instruct" in model_path:
+        return get_chat_template("granite-3-instruct")
+
+
 if __name__ == "__main__":
     messages = [
         {"role": "system", "content": None},  # None means default