sglang 0.4.4__py3-none-any.whl → 0.4.4.post2__py3-none-any.whl

sglang/test/test_block_fp8.py

@@ -11,6 +11,7 @@ from sglang.srt.layers.quantization.fp8_kernel import (
     static_quant_fp8,
     w8a8_block_fp8_matmul,
 )
+from sglang.test.test_utils import CustomTestCase
 
 _is_cuda = torch.cuda.is_available() and torch.version.cuda
 
@@ -44,7 +45,7 @@ def native_per_token_group_quant_fp8(
     return x_q, x_s
 
 
-class TestPerTokenGroupQuantFP8(unittest.TestCase):
+class TestPerTokenGroupQuantFP8(CustomTestCase):
     DTYPES = [torch.half, torch.bfloat16, torch.float32]
     NUM_TOKENS = [7, 83, 2048]
     D = [512, 4096, 5120, 13824]
@@ -111,7 +112,7 @@ def native_static_quant_fp8(x, x_s, dtype=torch.float8_e4m3fn):
     return x_q, x_s
 
 
-class TestStaticQuantFP8(unittest.TestCase):
+class TestStaticQuantFP8(CustomTestCase):
     DTYPES = [torch.half, torch.bfloat16, torch.float32]
     NUM_TOKENS = [7, 83, 2048]
     D = [512, 4096, 5120, 13824]
@@ -210,7 +211,7 @@ def native_w8a8_block_fp8_matmul(A, B, As, Bs, block_size, output_dtype=torch.fl
     return C
 
 
-class TestW8A8BlockFP8Matmul(unittest.TestCase):
+class TestW8A8BlockFP8Matmul(CustomTestCase):
 
     if not _is_cuda:
         OUT_DTYPES = [torch.float32, torch.half, torch.bfloat16]
@@ -331,7 +332,7 @@ def torch_w8a8_block_fp8_moe(a, w1, w2, w1_s, w2_s, score, topk, block_shape):
     ).sum(dim=1)
 
 
-class TestW8A8BlockFP8FusedMoE(unittest.TestCase):
+class TestW8A8BlockFP8FusedMoE(CustomTestCase):
     DTYPES = [torch.float32, torch.half, torch.bfloat16]
     M = [1, 33, 64, 222, 1024 * 128]
     N = [128, 1024, 2048]

sglang/test/test_block_fp8_ep.py

@@ -13,6 +13,7 @@ from sglang.srt.layers.moe.ep_moe.kernels import (
     silu_and_mul_triton_kernel,
 )
 from sglang.srt.layers.moe.topk import select_experts
+from sglang.test.test_utils import CustomTestCase
 
 
 # For test
@@ -232,7 +233,7 @@ def block_dequant(
     return x_dq_block
 
 
-class TestW8A8BlockFP8EPMoE(unittest.TestCase):
+class TestW8A8BlockFP8EPMoE(CustomTestCase):
     DTYPES = [torch.half, torch.bfloat16]
     M = [1, 222, 1024, 2048]
     N = [128, 1024, 2048]

sglang/test/test_dynamic_grad_mode.py

@@ -0,0 +1,58 @@
+import unittest
+
+import torch
+
+from sglang.srt.utils import DynamicGradMode
+from sglang.test.test_utils import CustomTestCase
+
+
+class TestDynamicGradMode(CustomTestCase):
+    def test_inference(self):
+        # Test inference_mode
+        DynamicGradMode.set_inference_mode(True)
+
+        @DynamicGradMode()
+        def create_tensor_x():
+            return torch.empty(0)
+
+        X = create_tensor_x()
+        self.assertTrue(not X.requires_grad and X.is_inference())
+
+    def test_no_grad(self):
+        # Test no_grad
+        DynamicGradMode.set_inference_mode(False)
+
+        @DynamicGradMode()
+        def create_tensor_y():
+            return torch.empty(0)
+
+        Y = create_tensor_y()
+        self.assertTrue(not Y.requires_grad and not Y.is_inference())
+
+    def test_nested_inference(self):
+        # Test no_grad nested inference_mode, inference_mode should has higher priority
+        DynamicGradMode.set_inference_mode(False)
+
+        @DynamicGradMode()
+        def create_tensor_z():
+            with torch.inference_mode():
+                return torch.empty(0)
+
+        Z = create_tensor_z()
+        self.assertTrue(not Z.requires_grad and Z.is_inference())
+
+    def test_nested_no_grad(self):
+        # Test inference_mode nested no_grad, inference_mode should has higher priority
+        DynamicGradMode.set_inference_mode(True)
+
+        @DynamicGradMode()
+        def create_tensor_w():
+            with torch.no_grad():
+                return torch.empty(0)
+
+        W = create_tensor_w()
+        self.assertTrue(not W.requires_grad and W.is_inference())
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)

sglang/test/test_layernorm.py

@@ -4,9 +4,10 @@ import unittest
 import torch
 
 from sglang.srt.layers.layernorm import GemmaRMSNorm, RMSNorm
+from sglang.test.test_utils import CustomTestCase
 
 
-class TestRMSNorm(unittest.TestCase):
+class TestRMSNorm(CustomTestCase):
     DTYPES = [torch.half, torch.bfloat16]
     NUM_TOKENS = [7, 83, 4096]
     HIDDEN_SIZES = [768, 769, 770, 771, 5120, 5124, 5125, 5126, 8192, 8199]
@@ -56,7 +57,7 @@ class TestRMSNorm(unittest.TestCase):
                 self._run_rms_norm_test(*params)
 
 
-class TestGemmaRMSNorm(unittest.TestCase):
+class TestGemmaRMSNorm(CustomTestCase):
     DTYPES = [torch.half, torch.bfloat16]
     NUM_TOKENS = [7, 83, 4096]
     HIDDEN_SIZES = [768, 769, 770, 771, 5120, 5124, 5125, 5126, 8192, 8199]

sglang/test/test_utils.py

@@ -1,15 +1,17 @@
 """Common utilities for testing and benchmarking"""
 
 import argparse
-import asyncio
 import copy
+import logging
 import os
 import random
 import subprocess
 import threading
 import time
+import traceback
 import unittest
 from concurrent.futures import ThreadPoolExecutor
+from dataclasses import dataclass
 from functools import partial
 from types import SimpleNamespace
 from typing import Callable, List, Optional, Tuple
@@ -32,6 +34,15 @@ DEFAULT_FP8_MODEL_NAME_FOR_ACCURACY_TEST = "neuralmagic/Meta-Llama-3-8B-Instruct
 DEFAULT_FP8_MODEL_NAME_FOR_DYNAMIC_QUANT_ACCURACY_TEST = (
     "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8-dynamic"
 )
+DEFAULT_FP8_MODEL_NAME_FOR_MODELOPT_QUANT_ACCURACY_TEST = (
+    "nvidia/Llama-3.1-8B-Instruct-FP8"
+)
+# TODO(yundai424): right now specifying to an older revision since the latest one
+#  carries kv cache quantization which doesn't work yet
+DEFAULT_FP8_MODEL_NAME_FOR_MODELOPT_QUANT_ACCURACY_TEST_REVISION = (
+    "13858565416dbdc0b4e7a4a677fadfbd5b9e5bb9"
+)
+
 DEFAULT_MODEL_NAME_FOR_TEST = "meta-llama/Llama-3.1-8B-Instruct"
 DEFAULT_SMALL_MODEL_NAME_FOR_TEST = "meta-llama/Llama-3.2-1B-Instruct"
 DEFAULT_MOE_MODEL_NAME_FOR_TEST = "mistralai/Mixtral-8x7B-Instruct-v0.1"
@@ -52,7 +63,6 @@ DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_QUANT_TP1 = "hugging-quants/Meta-Llama-3.1-8
 DEFAULT_SMALL_MODEL_NAME_FOR_TEST_QWEN = "Qwen/Qwen2.5-1.5B-Instruct"
 DEFAULT_SMALL_VLM_MODEL_NAME = "Qwen/Qwen2-VL-2B"
 
-
 DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST = "meta-llama/Llama-2-7b-chat-hf"
 DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST = "lmsys/sglang-EAGLE-llama2-chat-7B"
 
@@ -423,6 +433,11 @@ def popen_launch_server(
                     return process
             except requests.RequestException:
                 pass
+
+            return_code = process.poll()
+            if return_code is not None:
+                raise Exception(f"Server unexpectedly exits ({return_code=}).")
+
             time.sleep(10)
 
     kill_process_tree(process.pid)
@@ -453,7 +468,13 @@ def run_with_timeout(
     return ret_value[0]
 
 
-def run_unittest_files(files: List, timeout_per_file: float):
+@dataclass
+class TestFile:
+    name: str
+    estimated_time: float = 60
+
+
+def run_unittest_files(files: List[TestFile], timeout_per_file: float):
     tic = time.time()
     success = True
 
@@ -870,7 +891,6 @@ def run_mulit_request_test(
     enable_overlap=False,
     chunked_prefill_size=32,
 ):
-
     def workload_func(base_url, model):
         def run_one(_):
             prompt = """
@@ -905,6 +925,10 @@ def run_mulit_request_test(
 
 
 def write_github_step_summary(content):
+    if not os.environ.get("GITHUB_STEP_SUMMARY"):
+        logging.warning("GITHUB_STEP_SUMMARY environment variable not set")
+        return
+
     with open(os.environ["GITHUB_STEP_SUMMARY"], "a") as f:
         f.write(content)
 
@@ -982,3 +1006,30 @@ def run_logprob_check(self: unittest.TestCase, arg: Tuple):
                                 rank += 1
                             else:
                                 raise
+
+
+class CustomTestCase(unittest.TestCase):
+    def _callTestMethod(self, method):
+        _retry_execution(
+            lambda: super(CustomTestCase, self)._callTestMethod(method),
+            max_retry=_get_max_retry(),
+        )
+
+
+def _get_max_retry():
+    return int(os.environ.get("SGLANG_TEST_MAX_RETRY", "2" if is_in_ci() else "0"))
+
+
+def _retry_execution(fn, max_retry: int):
+    if max_retry == 0:
+        fn()
+        return
+
+    try:
+        fn()
+    except Exception as e:
+        print(
+            f"retry_execution failed once and will retry. This may be an error or a flaky test. Error: {e}"
+        )
+        traceback.print_exc()
+        _retry_execution(fn, max_retry=max_retry - 1)

sglang/utils.py

@@ -22,6 +22,7 @@ from typing import Any, Callable, List, Optional, Tuple, Type, Union
 import numpy as np
 import requests
 from IPython.display import HTML, display
+from pydantic import BaseModel
 from tqdm import tqdm
 
 from sglang.srt.utils import kill_process_tree
@@ -29,6 +30,36 @@ from sglang.srt.utils import kill_process_tree
 logger = logging.getLogger(__name__)
 
 
+def convert_json_schema_to_str(json_schema: Union[dict, str, Type[BaseModel]]) -> str:
+    """Convert a JSON schema to a string.
+    Parameters
+    ----------
+    json_schema
+        The JSON schema.
+    Returns
+    -------
+    str
+        The JSON schema converted to a string.
+    Raises
+    ------
+    ValueError
+        If the schema is not a dictionary, a string or a Pydantic class.
+    """
+    if isinstance(json_schema, dict):
+        schema_str = json.dumps(json_schema)
+    elif isinstance(json_schema, str):
+        schema_str = json_schema
+    elif issubclass(json_schema, BaseModel):
+        schema_str = json.dumps(json_schema.model_json_schema())
+    else:
+        raise ValueError(
+            f"Cannot parse schema {json_schema}. The schema must be either "
+            + "a Pydantic class, a dictionary or a string that contains the JSON "
+            + "schema specification"
+        )
+    return schema_str
+
+
 def get_exception_traceback():
     etype, value, tb = sys.exc_info()
     err_str = "".join(traceback.format_exception(etype, value, tb))

sglang/version.py

@@ -1 +1 @@
-__version__ = "0.4.4"
+__version__ = "0.4.4.post2"

{sglang-0.4.4.dist-info → sglang-0.4.4.post2.dist-info}/METADATA

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: sglang
-Version: 0.4.4
+Version: 0.4.4.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
@@ -235,17 +235,17 @@ 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: soundfile==0.13.1; extra == "runtime-common"
 Requires-Dist: torchao>=0.7.0; extra == "runtime-common"
-Requires-Dist: transformers==4.48.3; extra == "runtime-common"
+Requires-Dist: transformers==4.50.0; extra == "runtime-common"
 Requires-Dist: uvicorn; extra == "runtime-common"
 Requires-Dist: uvloop; extra == "runtime-common"
-Requires-Dist: xgrammar==0.1.15; extra == "runtime-common"
+Requires-Dist: xgrammar==0.1.16; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"
-Requires-Dist: sgl-kernel==0.0.5; extra == "srt"
+Requires-Dist: sgl-kernel==0.0.5.post3; extra == "srt"
 Requires-Dist: flashinfer_python==0.2.3; extra == "srt"
 Requires-Dist: torch==2.5.1; extra == "srt"
-Requires-Dist: vllm<=0.7.2,>=0.6.4.post1; extra == "srt"
 Requires-Dist: cuda-python; extra == "srt"
 Requires-Dist: outlines<=0.1.11,>=0.0.44; extra == "srt"
 Provides-Extra: srt-hip
@@ -271,7 +271,7 @@ Requires-Dist: anthropic>=0.20.0; extra == "anthropic"
 Provides-Extra: litellm
 Requires-Dist: litellm>=1.0.0; extra == "litellm"
 Provides-Extra: torch-memory-saver
-Requires-Dist: torch_memory_saver; extra == "torch-memory-saver"
+Requires-Dist: torch_memory_saver>=0.0.3; extra == "torch-memory-saver"
 Provides-Extra: test
 Requires-Dist: jsonlines; extra == "test"
 Requires-Dist: matplotlib; extra == "test"
@@ -319,6 +319,7 @@ 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"
+Dynamic: license-file
 
 <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>
@@ -342,6 +343,9 @@ Requires-Dist: sglang[test]; extra == "dev-cpu"
 | [**Slides**](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#slides) |
 
 ## News
+- [2025/03] Supercharge DeepSeek-R1 Inference on AMD Instinct MI300X ([AMD blog](https://rocm.blogs.amd.com/artificial-intelligence/DeepSeekR1-Part2/README.html))
+- [2025/03] SGLang Joins PyTorch Ecosystem: Efficient LLM Serving Engine ([PyTorch blog](https://pytorch.org/blog/sglang-joins-pytorch/))
+- [2025/02] Unlock DeepSeek-R1 Inference Performance on AMD Instinct™ MI300X GPU ([AMD blog](https://rocm.blogs.amd.com/artificial-intelligence/DeepSeekR1_Perf/README.html))
 - [2025/01] 🔥 SGLang provides day one support for DeepSeek V3/R1 models on NVIDIA and AMD GPUs with DeepSeek-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), [10+ other companies](https://x.com/lmsysorg/status/1887262321636221412))
 - [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/)).
@@ -362,7 +366,7 @@ SGLang is a fast serving framework for large language models and vision language
 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, overhead-free CPU scheduler, continuous batching, token attention (paged attention), tensor parallelism, FlashInfer kernels, chunked prefill, and quantization (FP8/INT4/AWQ/GPTQ).
+- **Fast Backend Runtime**: Provides efficient serving with RadixAttention for prefix caching, zero-overhead CPU scheduler, continuous batching, token attention (paged attention), speculative decoding, tensor parallelism, chunked prefill, structured outputs, and quantization (FP8/INT4/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.), embedding models (e5-mistral, gte, mcdse) and reward models (Skywork), with easy extensibility for integrating new models.
 - **Active Community**: SGLang is open-source and backed by an active community with industry adoption.