sglang 0.4.3.post1__py3-none-any.whl → 0.4.3.post3__py3-none-any.whl

sglang/test/test_utils.py

@@ -8,10 +8,11 @@ import random
 import subprocess
 import threading
 import time
+import unittest
 from concurrent.futures import ThreadPoolExecutor
 from functools import partial
 from types import SimpleNamespace
-from typing import Callable, List, Optional
+from typing import Callable, List, Optional, Tuple
 
 import numpy as np
 import requests
@@ -34,6 +35,7 @@ DEFAULT_SMALL_MOE_MODEL_NAME_FOR_TEST = "Qwen/Qwen1.5-MoE-A2.7B"
 DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST = "Alibaba-NLP/gte-Qwen2-1.5B-instruct"
 DEFAULT_MLA_MODEL_NAME_FOR_TEST = "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct"
 DEFAULT_MLA_FP8_MODEL_NAME_FOR_TEST = "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8"
+DEFAULT_REASONING_MODEL_NAME_FOR_TEST = "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B"
 DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH = 1000
 DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP1 = "meta-llama/Llama-3.1-8B-Instruct,mistralai/Mistral-7B-Instruct-v0.3,deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct,google/gemma-2-27b-it"
 DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP2 = "meta-llama/Llama-3.1-70B-Instruct,mistralai/Mixtral-8x7B-Instruct-v0.1,Qwen/Qwen2-57B-A14B-Instruct"
@@ -41,9 +43,14 @@ DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP1 = "neuralmagic/Meta-Llama-3.1-8B-Ins
 DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP2 = "neuralmagic/Meta-Llama-3.1-70B-Instruct-FP8,neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8,neuralmagic/Qwen2-72B-Instruct-FP8,neuralmagic/Qwen2-57B-A14B-Instruct-FP8,neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8"
 DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_QUANT_TP1 = "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4,hugging-quants/Meta-Llama-3.1-8B-Instruct-GPTQ-INT4"
 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 = "lmzheng/sglang-EAGLE-llama2-chat-7B"
+DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST = "lmsys/sglang-EAGLE-llama2-chat-7B"
+
+DEFAULT_IMAGE_URL = "https://github.com/sgl-project/sglang/blob/main/test/lang/example_image.png?raw=true"
+DEFAULT_VIDEO_URL = "https://raw.githubusercontent.com/EvolvingLMMs-Lab/sglang/dev/onevision_local/assets/jobs.mp4"
 
 
 def is_in_ci():
@@ -158,45 +165,6 @@ def call_generate_guidance(
     return rets if n > 1 else rets[0]
 
 
-async def call_generate_lmql(
-    prompt, temperature, max_tokens, stop=None, n=1, max_len=4096, model=None, **kwargs
-):
-    assert model is not None
-    import lmql
-
-    if stop != None:
-
-        @lmql.query(model=model)
-        async def program(question, max_tokens, stop):
-            '''lmql
-            """{question}[ANSWER]""" where len(TOKENS(ANSWER)) < max_tokens and STOPS_AT(ANSWER, stop)
-            return ANSWER
-            '''
-
-    else:
-
-        @lmql.query(model=model)
-        async def program(question, max_tokens):
-            '''lmql
-            """{question}[ANSWER]""" where len(TOKENS(ANSWER)) < max_tokens
-            return ANSWER
-            '''
-
-    tasks = [
-        program(
-            question=prompt,
-            temperature=temperature,
-            max_tokens=max_tokens,
-            stop=stop,
-            max_len=max_len,
-            **kwargs,
-        )
-        for _ in range(n)
-    ]
-    rets = await asyncio.gather(*tasks)
-    return rets if n > 1 else rets[0]
-
-
 def call_select_lightllm(context, choices, url=None):
     assert url is not None
 
@@ -246,23 +214,6 @@ def call_select_guidance(context, choices, model=None):
     return choices.index(out["answer"])
 
 
-async def call_select_lmql(context, choices, temperature=0, max_len=4096, model=None):
-    assert model is not None
-    import lmql
-
-    @lmql.query(model=model)
-    async def program(ctx, choices):
-        '''lmql
-        """{ctx}[ANSWER]""" where ANSWER in set(choices)
-        return ANSWER
-        '''
-
-    answer = await program(
-        ctx=context, choices=choices, temperature=temperature, max_len=max_len
-    )
-    return choices.index(answer)
-
-
 def add_common_other_args_and_parse(parser: argparse.ArgumentParser):
     parser.add_argument("--parallel", type=int, default=64)
     parser.add_argument("--host", type=str, default="http://127.0.0.1")
@@ -277,7 +228,6 @@ def add_common_other_args_and_parse(parser: argparse.ArgumentParser):
             "lightllm",
             "gserver",
             "guidance",
-            "lmql",
             "srt-raw",
             "llama.cpp",
         ],
@@ -294,7 +244,6 @@ def add_common_other_args_and_parse(parser: argparse.ArgumentParser):
             "vllm": 21000,
             "outlines": 21000,
             "lightllm": 22000,
-            "lmql": 23000,
             "srt-raw": 30000,
             "gserver": 9988,
         }
@@ -342,11 +291,6 @@ def _get_call_generate(args: argparse.Namespace):
         call_generate = partial(call_generate_guidance, model=model)
         call_generate("Hello,", 1.0, 8, ".")
         return call_generate
-    elif args.backend == "lmql":
-        import lmql
-
-        model = lmql.model(args.model_path, endpoint=f"{args.host}:{args.port}")
-        return partial(call_generate_lmql, model=model)
     else:
         raise ValueError(f"Invalid backend: {args.backend}")
 
@@ -364,12 +308,6 @@ def _get_call_select(args: argparse.Namespace):
 
         call_select("Hello,", ["world", "earth"])
         return call_select
-
-    elif args.backend == "lmql":
-        import lmql
-
-        model = lmql.model(args.model_path, endpoint=f"{args.host}:{args.port}")
-        return partial(call_select_lmql, model=model)
     else:
         raise ValueError(f"Invalid backend: {args.backend}")
 
@@ -408,26 +346,49 @@ def popen_launch_server(
     other_args: list[str] = (),
     env: Optional[dict] = None,
     return_stdout_stderr: Optional[tuple] = None,
+    pd_seperated: bool = False,
 ):
     _, host, port = base_url.split(":")
     host = host[2:]
 
+    if pd_seperated:
+        command = "sglang.launch_pd_server"
+    else:
+        command = "sglang.launch_server"
+
     command = [
         "python3",
         "-m",
-        "sglang.launch_server",
+        command,
         "--model-path",
         model,
-        "--host",
-        host,
-        "--port",
-        port,
-        *other_args,
+        *[str(x) for x in other_args],
     ]
 
+    if pd_seperated:
+        command.extend(
+            [
+                "--lb-host",
+                host,
+                "--lb-port",
+                port,
+            ]
+        )
+    else:
+        command.extend(
+            [
+                "--host",
+                host,
+                "--port",
+                port,
+            ]
+        )
+
     if api_key:
         command += ["--api-key", api_key]
 
+    print(f"command={' '.join(command)}")
+
     if return_stdout_stderr:
         process = subprocess.Popen(
             command,
@@ -456,6 +417,8 @@ def popen_launch_server(
             except requests.RequestException:
                 pass
             time.sleep(10)
+
+    kill_process_tree(process.pid)
     raise TimeoutError("Server failed to start within the timeout period.")
 
 
@@ -488,9 +451,11 @@ def run_unittest_files(files: List[str], timeout_per_file: float):
     success = True
 
     for filename in files:
-        global process
+        process = None
 
         def run_one_file(filename):
+            nonlocal process
+
             filename = os.path.join(os.getcwd(), filename)
             print(f"\n\nRun:\npython3 {filename}\n\n", flush=True)
             process = subprocess.Popen(
@@ -503,7 +468,9 @@ def run_unittest_files(files: List[str], timeout_per_file: float):
             ret_code = run_with_timeout(
                 run_one_file, args=(filename,), timeout=timeout_per_file
             )
-            assert ret_code == 0
+            assert (
+                ret_code == 0
+            ), f"expected return code 0, but {filename} returned {ret_code}"
         except TimeoutError:
             kill_process_tree(process.pid)
             time.sleep(5)
@@ -532,11 +499,15 @@ def get_benchmark_args(
     dataset_path="",
     tokenizer="",
     num_prompts=500,
+    sharegpt_output_len=None,
     random_input_len=4096,
     random_output_len=2048,
+    sharegpt_context_len=None,
     request_rate=float("inf"),
     disable_stream=False,
     disable_ignore_eos=False,
+    seed: int = 0,
+    pd_seperated: bool = False,
 ):
     return SimpleNamespace(
         backend="sglang",
@@ -548,8 +519,8 @@ def get_benchmark_args(
         model=None,
         tokenizer=tokenizer,
         num_prompts=num_prompts,
-        sharegpt_output_len=None,
-        sharegpt_context_len=None,
+        sharegpt_output_len=sharegpt_output_len,
+        sharegpt_context_len=sharegpt_context_len,
         random_input_len=random_input_len,
         random_output_len=random_output_len,
         random_range_ratio=0.0,
@@ -559,12 +530,14 @@ def get_benchmark_args(
         disable_tqdm=False,
         disable_stream=disable_stream,
         return_logprob=False,
-        seed=0,
+        seed=seed,
         disable_ignore_eos=disable_ignore_eos,
         extra_request_body=None,
         apply_chat_template=False,
         profile=None,
         lora_name=None,
+        prompt_suffix="",
+        pd_seperated=pd_seperated,
     )
 
 
@@ -578,9 +551,11 @@ def run_bench_serving(
     tokenizer=None,
     random_input_len=4096,
     random_output_len=2048,
+    sharegpt_context_len=None,
     disable_stream=False,
     disable_ignore_eos=False,
     need_warmup=False,
+    seed: int = 0,
 ):
     # Launch the server
     base_url = DEFAULT_URL_FOR_TEST
@@ -600,9 +575,11 @@ def run_bench_serving(
         num_prompts=num_prompts,
         random_input_len=random_input_len,
         random_output_len=random_output_len,
+        sharegpt_context_len=sharegpt_context_len,
         request_rate=request_rate,
         disable_stream=disable_stream,
         disable_ignore_eos=disable_ignore_eos,
+        seed=seed,
     )
 
     try:
@@ -624,6 +601,7 @@ def run_bench_serving_multi(
     other_server_args,
     benchmark_args,
     need_warmup=False,
+    pd_seperated=False,
 ):
     # Launch the server
     process = popen_launch_server(
@@ -631,6 +609,7 @@ def run_bench_serving_multi(
         base_url,
         timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
         other_args=other_server_args,
+        pd_seperated=pd_seperated,
     )
 
     # run benchmark for all
@@ -663,7 +642,7 @@ def run_bench_one_batch(model, other_args):
         "128",
         "--output",
         "8",
-        *other_args,
+        *[str(x) for x in other_args],
     ]
     process = subprocess.Popen(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
 
@@ -814,7 +793,7 @@ def run_command_and_capture_output(command, env: Optional[dict] = None):
     stdout = open(STDOUT_FILENAME, "w")
     stderr = open(STDERR_FILENAME, "w")
     process = subprocess.Popen(
-        command, stdout=stdout, stderr=stderr, env=env, text=True
+        command, stdout=stdout, stderr=stdout, env=env, text=True
     )
 
     # Launch a thread to stream the output
@@ -912,3 +891,78 @@ def run_mulit_request_test(
 def write_github_step_summary(content):
     with open(os.environ["GITHUB_STEP_SUMMARY"], "a") as f:
         f.write(content)
+
+
+def run_logprob_check(self: unittest.TestCase, arg: Tuple):
+    (
+        input_len,
+        output_len,
+        temperature,
+        logprob_start_len,
+        return_logprob,
+        top_logprobs_num,
+    ) = arg
+    input_ids = list(range(input_len))
+
+    response = requests.post(
+        self.base_url + "/generate",
+        json={
+            "input_ids": input_ids,
+            "sampling_params": {
+                "temperature": temperature,
+                "max_new_tokens": output_len,
+                "ignore_eos": True,
+            },
+            "return_logprob": return_logprob,
+            "logprob_start_len": logprob_start_len,
+            "top_logprobs_num": top_logprobs_num,
+        },
+    )
+    response_json = response.json()
+
+    res = response_json
+    self.assertEqual(res["meta_info"]["prompt_tokens"], input_len)
+    self.assertEqual(res["meta_info"]["completion_tokens"], output_len)
+
+    # Test the number of tokens are correct
+    if return_logprob:
+        self.assertEqual(
+            len(res["meta_info"]["input_token_logprobs"]) + logprob_start_len,
+            res["meta_info"]["prompt_tokens"],
+        )
+        self.assertEqual(len(res["meta_info"]["output_token_logprobs"]), output_len)
+
+        if top_logprobs_num:
+            self.assertEqual(
+                len(res["meta_info"]["input_top_logprobs"]) + logprob_start_len,
+                res["meta_info"]["prompt_tokens"],
+            )
+            self.assertEqual(len(res["meta_info"]["output_top_logprobs"]), output_len)
+
+            for i in range(output_len):
+                self.assertEqual(
+                    len(res["meta_info"]["output_top_logprobs"][i]),
+                    top_logprobs_num,
+                )
+
+                # Test the top-1 tokens are the same as output tokens if temperature == 0
+                if temperature == 0:
+                    rank = 0
+                    while rank < len(res["meta_info"]["output_top_logprobs"][i]):
+                        try:
+                            self.assertListEqual(
+                                res["meta_info"]["output_token_logprobs"][i],
+                                res["meta_info"]["output_top_logprobs"][i][rank],
+                            )
+                            break
+                        except AssertionError:
+                            # There's a tie. Allow the second item in this case.
+                            if (
+                                res["meta_info"]["output_top_logprobs"][i][rank][0]
+                                == res["meta_info"]["output_top_logprobs"][i][rank + 1][
+                                    0
+                                ]
+                            ):
+                                rank += 1
+                            else:
+                                raise

sglang/utils.py

@@ -5,12 +5,15 @@ import importlib
 import json
 import logging
 import os
+import random
 import signal
+import socket
 import subprocess
 import sys
 import time
 import traceback
 import urllib.request
+import weakref
 from concurrent.futures import ThreadPoolExecutor
 from io import BytesIO
 from json import dumps
@@ -21,8 +24,14 @@ import requests
 from IPython.display import HTML, display
 from tqdm import tqdm
 
+from sglang.srt.openai_api.protocol import ChatCompletionMessageContentPart
+from sglang.srt.utils import kill_process_tree
+
 logger = logging.getLogger(__name__)
 
+# type of content fields, can be only prompts or with images/videos
+MsgContent = Union[str, List[ChatCompletionMessageContentPart]]
+
 
 def get_exception_traceback():
     etype, value, tb = sys.exc_info()
@@ -306,27 +315,12 @@ def download_and_cache_file(url: str, filename: Optional[str] = None):
     return filename
 
 
-import fcntl
-
-
 def is_in_ci():
     from sglang.test.test_utils import is_in_ci
 
     return is_in_ci()
 
 
-LOCKFILE = os.path.expanduser("~/.sglang_port_lock")
-PORT_REGISTRY = os.path.expanduser("~/.sglang_port_registry.json")
-
-if not os.path.exists(LOCKFILE):
-    with open(LOCKFILE, "w") as f:
-        pass
-
-if not os.path.exists(PORT_REGISTRY):
-    with open(PORT_REGISTRY, "w") as f:
-        json.dump([], f)
-
-
 def print_highlight(html_content: str):
     if is_in_ci():
         html_content = str(html_content).replace("\n", "<br>")
@@ -335,55 +329,44 @@ def print_highlight(html_content: str):
         print(html_content)
 
 
-def init_port_registry():
-    """Initialize the port registry file if it doesn't exist."""
-    if not os.path.exists(PORT_REGISTRY):
-        with open(PORT_REGISTRY, "w") as f:
-            json.dump([], f)
+process_socket_map = weakref.WeakKeyDictionary()
 
 
-def reserve_port(start=30000, end=40000):
+def reserve_port(host, start=30000, end=40000):
     """
-    Reserve an available port using a file lock and a registry.
-    Returns the allocated port.
+    Reserve an available port by trying to bind a socket.
+    Returns a tuple (port, lock_socket) where `lock_socket` is kept open to hold the lock.
     """
-    init_port_registry()
-    with open(LOCKFILE, "w") as lock:
-        fcntl.flock(lock, fcntl.LOCK_EX)
-        try:
-            with open(PORT_REGISTRY, "r") as f:
-                used = json.load(f)
-        except Exception:
-            used = []
-        for port in range(start, end):
-            if port not in used:
-                used.append(port)
-                with open(PORT_REGISTRY, "w") as f:
-                    json.dump(used, f)
-                return port
-    raise RuntimeError("No free port available")
-
-
-def release_port(port):
-    """Release the reserved port by removing it from the registry."""
-    with open(LOCKFILE, "w") as lock:
-        fcntl.flock(lock, fcntl.LOCK_EX)
+    candidates = list(range(start, end))
+    random.shuffle(candidates)
+
+    for port in candidates:
+        sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+        sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
         try:
-            with open(PORT_REGISTRY, "r") as f:
-                used = json.load(f)
-        except Exception:
-            used = []
-        if port in used:
-            used.remove(port)
-        with open(PORT_REGISTRY, "w") as f:
-            json.dump(used, f)
+            # Attempt to bind to the port on localhost
+            sock.bind((host, port))
+            return port, sock
+        except socket.error:
+            sock.close()  # Failed to bind, try next port
+            continue
+    raise RuntimeError("No free port available.")
+
+
+def release_port(lock_socket):
+    """
+    Release the reserved port by closing the lock socket.
+    """
+    try:
+        lock_socket.close()
+    except Exception as e:
+        print(f"Error closing socket: {e}")
 
 
 def execute_shell_command(command: str) -> subprocess.Popen:
     """
     Execute a shell command and return its process handle.
     """
-    # Replace newline continuations and split the command string.
     command = command.replace("\\\n", " ").replace("\\", " ")
     parts = command.split()
     return subprocess.Popen(parts, text=True, stderr=subprocess.STDOUT)
@@ -395,21 +378,28 @@ def launch_server_cmd(command: str, host: str = "0.0.0.0", port: int = None):
     If no port is specified, a free port is reserved.
     """
     if port is None:
-        port = reserve_port()
+        port, lock_socket = reserve_port(host)
+    else:
+        lock_socket = None
+
     full_command = f"{command} --port {port}"
     process = execute_shell_command(full_command)
+
+    if lock_socket is not None:
+        process_socket_map[process] = lock_socket
+
     return process, port
 
 
-def terminate_process(process, port=None):
+def terminate_process(process):
     """
-    Terminate the process and, if a port was reserved, release it.
+    Terminate the process and automatically release the reserved port.
     """
-    from sglang.srt.utils import kill_process_tree
-
     kill_process_tree(process.pid)
-    if port is not None:
-        release_port(port)
+
+    lock_socket = process_socket_map.pop(process, None)
+    if lock_socket is not None:
+        release_port(lock_socket)
 
 
 def wait_for_server(base_url: str, timeout: int = None) -> None:

sglang/version.py

@@ -1 +1 @@
-__version__ = "0.4.3.post1"
+__version__ = "0.4.3.post3"

{sglang-0.4.3.post1.dist-info → sglang-0.4.3.post3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: sglang
-Version: 0.4.3.post1
+Version: 0.4.3.post3
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -235,32 +235,34 @@ 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==0.1.10; extra == "runtime-common"
+Requires-Dist: xgrammar==0.1.14; extra == "runtime-common"
 Requires-Dist: ninja; extra == "runtime-common"
+Requires-Dist: transformers==4.48.3; extra == "runtime-common"
+Requires-Dist: llguidance>=0.6.15; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"
-Requires-Dist: cuda-python; extra == "srt"
-Requires-Dist: sgl-kernel>=0.0.3.post6; extra == "srt"
-Requires-Dist: torch; extra == "srt"
+Requires-Dist: sgl-kernel==0.0.3.post6; extra == "srt"
+Requires-Dist: flashinfer_python==0.2.2.post1; extra == "srt"
+Requires-Dist: torch==2.5.1; extra == "srt"
 Requires-Dist: vllm<=0.7.2,>=0.6.4.post1; extra == "srt"
-Requires-Dist: flashinfer_python>=0.2.1.post1; extra == "srt"
+Requires-Dist: cuda-python; extra == "srt"
 Requires-Dist: outlines<=0.1.11,>=0.0.44; extra == "srt"
 Provides-Extra: srt-hip
 Requires-Dist: sglang[runtime_common]; extra == "srt-hip"
+Requires-Dist: sgl-kernel==0.0.3.post6; extra == "srt-hip"
 Requires-Dist: torch; extra == "srt-hip"
 Requires-Dist: vllm==0.6.7.dev2; extra == "srt-hip"
 Requires-Dist: outlines==0.1.11; extra == "srt-hip"
-Requires-Dist: sgl-kernel>=0.0.3.post1; extra == "srt-hip"
 Provides-Extra: srt-xpu
 Requires-Dist: sglang[runtime_common]; extra == "srt-xpu"
-Requires-Dist: outlines<0.1.0,>=0.0.44; extra == "srt-xpu"
+Requires-Dist: outlines<=0.1.11,>=0.0.44; extra == "srt-xpu"
 Provides-Extra: srt-hpu
 Requires-Dist: sglang[runtime_common]; extra == "srt-hpu"
-Requires-Dist: outlines<0.1.0,>=0.0.44; extra == "srt-hpu"
+Requires-Dist: outlines<=0.1.11,>=0.0.44; extra == "srt-hpu"
 Provides-Extra: srt-cpu
 Requires-Dist: sglang[runtime_common]; extra == "srt-cpu"
+Requires-Dist: outlines<=0.1.11,>=0.0.44; extra == "srt-cpu"
 Requires-Dist: torch; extra == "srt-cpu"
-Requires-Dist: outlines<0.1.0,>=0.0.44; extra == "srt-cpu"
 Provides-Extra: openai
 Requires-Dist: openai>=1.0; extra == "openai"
 Requires-Dist: tiktoken; extra == "openai"
@@ -318,7 +320,7 @@ Provides-Extra: dev-cpu
 Requires-Dist: sglang[all_cpu]; extra == "dev-cpu"
 Requires-Dist: sglang[test]; extra == "dev-cpu"
 
-<div align="center"  id="sglangtop">
+<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)
@@ -336,10 +338,11 @@ Requires-Dist: sglang[test]; extra == "dev-cpu"
 | [**Documentation**](https://docs.sglang.ai/)
 | [**Join Slack**](https://slack.sglang.ai/)
 | [**Join Bi-Weekly Development Meeting**](https://meeting.sglang.ai/)
+| [**Roadmap**](https://github.com/sgl-project/sglang/issues/4042)
 | [**Slides**](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#slides) |
 
 ## News
-- [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))
+- [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/)).
 - [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/)).
@@ -366,7 +369,7 @@ The core features include:
 
 ## Getting Started
 - [Install SGLang](https://docs.sglang.ai/start/install.html)
-- [Quick Start](https://docs.sglang.ai/start/send_request.html)
+- [Quick Start](https://docs.sglang.ai/backend/send_request.html)
 - [Backend Tutorial](https://docs.sglang.ai/backend/openai_api_completions.html)
 - [Frontend Tutorial](https://docs.sglang.ai/frontend/frontend.html)
 - [Contribution Guide](https://docs.sglang.ai/references/contribution_guide.html)
@@ -375,10 +378,13 @@ The core features include:
 Learn more in the 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)
+[Development Roadmap (2025 H1)](https://github.com/sgl-project/sglang/issues/4042)
 
 ## Adoption and Sponsorship
-The project is supported by (alphabetically): AMD, Atlas Cloud, Baseten, Cursor, DataCrunch, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, LMSYS CORP, Meituan, Nebius, Novita AI, NVIDIA, RunPod, Stanford, UC Berkeley, UCLA, xAI, 01.AI.
+The project has been deployed to large-scale production, generating trillions of tokens every day.
+It is supported by the following institutions: AMD, Atlas Cloud, Baseten, Cursor, DataCrunch, Etched, Hyperbolic, Iflytek, Jam & Tea Studios, LinkedIn, LMSYS, Meituan, Nebius, Novita AI, NVIDIA, RunPod, Stanford, UC Berkeley, UCLA, xAI, and 01.AI.
+
+<img src="https://raw.githubusercontent.com/sgl-project/sgl-learning-materials/main/slides/adoption.png" alt="logo" width="800" margin="10px"></img>
 
 ## Contact Us