sglang 0.2.9.post1__py3-none-any.whl → 0.2.11__py3-none-any.whl

sglang/srt/server.py

@@ -28,7 +28,7 @@ import sys
 import threading
 import time
 from http import HTTPStatus
-from typing import Dict, Optional
+from typing import Dict, List, Optional, Union
 
 # Fix a bug of Python threading
 setattr(threading, "_register_atexit", lambda *args, **kwargs: None)
@@ -59,6 +59,7 @@ from sglang.srt.openai_api.adapter import (
     v1_batches,
     v1_chat_completions,
     v1_completions,
+    v1_delete_file,
     v1_files_create,
     v1_retrieve_batch,
     v1_retrieve_file,
@@ -67,13 +68,13 @@ from sglang.srt.openai_api.adapter import (
 from sglang.srt.openai_api.protocol import ModelCard, ModelList
 from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.utils import (
-    API_KEY_HEADER_NAME,
-    APIKeyValidatorMiddleware,
+    add_api_key_middleware,
     allocate_init_ports,
     assert_pkg_version,
     enable_show_time_cost,
     kill_child_process,
     maybe_set_triton_cache_manager,
+    set_torch_compile_config,
     set_ulimit,
 )
 from sglang.utils import get_exception_traceback
@@ -158,6 +159,16 @@ async def openai_v1_chat_completions(raw_request: Request):
     return await v1_chat_completions(tokenizer_manager, raw_request)
 
 
+@app.get("/v1/models")
+def available_models():
+    """Show available models."""
+    served_model_names = [tokenizer_manager.served_model_name]
+    model_cards = []
+    for served_model_name in served_model_names:
+        model_cards.append(ModelCard(id=served_model_name, root=served_model_name))
+    return ModelList(data=model_cards)
+
+
 @app.post("/v1/files")
 async def openai_v1_files(file: UploadFile = File(...), purpose: str = Form("batch")):
     return await v1_files_create(
@@ -165,6 +176,12 @@ async def openai_v1_files(file: UploadFile = File(...), purpose: str = Form("bat
     )
 
 
+@app.delete("/v1/files/{file_id}")
+async def delete_file(file_id: str):
+    # https://platform.openai.com/docs/api-reference/files/delete
+    return await v1_delete_file(file_id)
+
+
 @app.post("/v1/batches")
 async def openai_v1_batches(raw_request: Request):
     return await v1_batches(tokenizer_manager, raw_request)
@@ -187,69 +204,11 @@ async def retrieve_file_content(file_id: str):
     return await v1_retrieve_file_content(file_id)
 
 
-@app.get("/v1/models")
-def available_models():
-    """Show available models."""
-    served_model_names = [tokenizer_manager.served_model_name]
-    model_cards = []
-    for served_model_name in served_model_names:
-        model_cards.append(ModelCard(id=served_model_name, root=served_model_name))
-    return ModelList(data=model_cards)
-
-
-def _set_torch_compile_config():
-    # The following configurations are for torch compile optimizations
-    import torch._dynamo.config
-    import torch._inductor.config
-
-    torch._inductor.config.coordinate_descent_tuning = True
-    torch._inductor.config.triton.unique_kernel_names = True
-    torch._inductor.config.fx_graph_cache = True  # Experimental feature to reduce compilation times, will be on by default in future
-
-    # FIXME: tmp workaround
-    torch._dynamo.config.accumulated_cache_size_limit = 256
-
-
-def set_envs_and_config(server_args: ServerArgs):
-    # Set global environments
-    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
-    os.environ["NCCL_CUMEM_ENABLE"] = "0"
-    os.environ["NCCL_NVLS_ENABLE"] = "0"
-    os.environ["TORCH_NCCL_AVOID_RECORD_STREAMS"] = "1"
-
-    # Set ulimit
-    set_ulimit()
-
-    # Enable show time cost for debugging
-    if server_args.show_time_cost:
-        enable_show_time_cost()
-
-    # Disable disk cache
-    if server_args.disable_disk_cache:
-        disable_cache()
-
-    # Fix triton bugs
-    if server_args.tp_size * server_args.dp_size > 1:
-        # FIXME: remove this after https://github.com/triton-lang/triton/pull/4295 is used as a dependency.
-        maybe_set_triton_cache_manager()
-
-    # Set torch compile config
-    if server_args.enable_torch_compile:
-        _set_torch_compile_config()
-
-    # Set global chat template
-    if server_args.chat_template:
-        # TODO: replace this with huggingface transformers template
-        load_chat_template_for_openai_api(server_args.chat_template)
-
-
 def launch_server(
     server_args: ServerArgs,
     model_overide_args: Optional[dict] = None,
     pipe_finish_writer: Optional[mp.connection.Connection] = None,
 ):
-    server_args.check_server_args()
-
     """Launch an HTTP server."""
     global tokenizer_manager
 
@@ -258,16 +217,8 @@ def launch_server(
         format="%(message)s",
     )
 
-    if not server_args.disable_flashinfer:
-        assert_pkg_version(
-            "flashinfer",
-            "0.1.3",
-            "Please uninstall the old version and "
-            "reinstall the latest version by following the instructions "
-            "at https://docs.flashinfer.ai/installation.html.",
-        )
-
-    set_envs_and_config(server_args)
+    server_args.check_server_args()
+    _set_envs_and_config(server_args)
 
     # Allocate ports
     server_args.port, server_args.additional_ports = allocate_init_ports(
@@ -284,7 +235,7 @@ def launch_server(
     )
     logger.info(f"{server_args=}")
 
-    # Handle multi-node tensor parallelism
+    # Launch processes for multi-node tensor parallelism
     if server_args.nnodes > 1:
         if server_args.node_rank != 0:
             tp_size_local = server_args.tp_size // server_args.nnodes
@@ -349,8 +300,9 @@ def launch_server(
         sys.exit(1)
     assert proc_controller.is_alive() and proc_detoken.is_alive()
 
-    if server_args.api_key and server_args.api_key != "":
-        app.add_middleware(APIKeyValidatorMiddleware, api_key=server_args.api_key)
+    # Add api key authorization
+    if server_args.api_key:
+        add_api_key_middleware(app, server_args.api_key)
 
     # Send a warmup request
     t = threading.Thread(
@@ -372,21 +324,74 @@ def launch_server(
         t.join()
 
 
+def _set_envs_and_config(server_args: ServerArgs):
+    # Set global environments
+    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+    os.environ["NCCL_CUMEM_ENABLE"] = "0"
+    os.environ["NCCL_NVLS_ENABLE"] = "0"
+    os.environ["TORCH_NCCL_AVOID_RECORD_STREAMS"] = "1"
+
+    # Set ulimit
+    set_ulimit()
+
+    # Enable show time cost for debugging
+    if server_args.show_time_cost:
+        enable_show_time_cost()
+
+    # Disable disk cache
+    if server_args.disable_disk_cache:
+        disable_cache()
+
+    # Fix triton bugs
+    if server_args.tp_size * server_args.dp_size > 1:
+        # FIXME: remove this after https://github.com/triton-lang/triton/pull/4295 is used as a dependency.
+        maybe_set_triton_cache_manager()
+
+    # Set torch compile config
+    if server_args.enable_torch_compile:
+        set_torch_compile_config()
+
+    # Set global chat template
+    if server_args.chat_template:
+        # TODO: replace this with huggingface transformers template
+        load_chat_template_for_openai_api(server_args.chat_template)
+
+    # Check flashinfer version
+    if not server_args.disable_flashinfer:
+        assert_pkg_version(
+            "flashinfer",
+            "0.1.3",
+            "Please uninstall the old version and "
+            "reinstall the latest version by following the instructions "
+            "at https://docs.flashinfer.ai/installation.html.",
+        )
+
+
 def _wait_and_warmup(server_args, pipe_finish_writer):
     headers = {}
     url = server_args.url()
     if server_args.api_key:
-        headers[API_KEY_HEADER_NAME] = server_args.api_key
+        headers["Authorization"] = f"Bearer {server_args.api_key}"
 
     # Wait until the server is launched
+    success = False
     for _ in range(120):
-        time.sleep(0.5)
+        time.sleep(1)
         try:
-            requests.get(url + "/get_model_info", timeout=5, headers=headers)
+            res = requests.get(url + "/get_model_info", timeout=5, headers=headers)
+            assert res.status_code == 200, f"{res}"
+            success = True
             break
-        except requests.exceptions.RequestException:
+        except (AssertionError, requests.exceptions.RequestException) as e:
+            last_traceback = get_exception_traceback()
             pass
 
+    if not success:
+        if pipe_finish_writer is not None:
+            pipe_finish_writer.send(last_traceback)
+        print(f"Initialization failed. warmup error: {last_traceback}", flush=True)
+        sys.exit(1)
+
     # Send a warmup request
     try:
         for _ in range(server_args.dp_size):
@@ -402,12 +407,13 @@ def _wait_and_warmup(server_args, pipe_finish_writer):
                 headers=headers,
                 timeout=600,
             )
-            assert res.status_code == 200
+            assert res.status_code == 200, f"{res}"
     except Exception as e:
+        last_traceback = get_exception_traceback()
         if pipe_finish_writer is not None:
-            pipe_finish_writer.send(get_exception_traceback())
-        print(f"Initialization failed. warmup error: {e}", flush=True)
-        raise e
+            pipe_finish_writer.send(last_traceback)
+        print(f"Initialization failed. warmup error: {last_traceback}", flush=True)
+        sys.exit(1)
 
     logger.info("The server is fired up and ready to roll!")
     if pipe_finish_writer is not None:
@@ -481,10 +487,10 @@ class Runtime:
             trust_remote_code=self.server_args.trust_remote_code,
         )
 
-    async def add_request(
+    async def async_generate(
         self,
         prompt: str,
-        sampling_params: Dict,
+        sampling_params: Optional[Dict] = None,
     ):
         json_data = {
             "text": prompt,
@@ -507,5 +513,26 @@ class Runtime:
                             yield cur
                         pos += len(cur)
 
+    add_request = async_generate
+
+    def generate(
+        self,
+        prompt: str,
+        sampling_params: Optional[Dict] = None,
+        return_logprob: Optional[Union[List[bool], bool]] = False,
+        top_logprobs_num: Optional[Union[List[int], int]] = None,
+    ):
+        json_data = {
+            "text": prompt,
+            "sampling_params": sampling_params,
+            "return_logprob": return_logprob,
+            "top_logprobs_num": top_logprobs_num,
+        }
+        response = requests.post(
+            self.url + "/generate",
+            json=json_data,
+        )
+        return json.dumps(response.json())
+
     def __del__(self):
         self.shutdown()

sglang/srt/server_args.py

@@ -61,7 +61,7 @@ class ServerArgs:
     show_time_cost: bool = False
 
     # Other
-    api_key: str = ""
+    api_key: Optional[str] = None
     file_storage_pth: str = "SGlang_storage"
 
     # Data parallelism
@@ -80,6 +80,7 @@ class ServerArgs:
     disable_disk_cache: bool = False
     enable_torch_compile: bool = False
     enable_p2p_check: bool = False
+    enable_mla: bool = False
     attention_reduce_in_fp32: bool = False
     efficient_weight_load: bool = False
 
@@ -263,6 +264,7 @@ class ServerArgs:
             help="How conservative the schedule policy is. A larger value means more conservative scheduling. Use a larger value if you see requests being retracted frequently.",
         )
         parser.add_argument(
+            "--tensor-parallel-size",
             "--tp-size",
             type=int,
             default=ServerArgs.tp_size,
@@ -306,7 +308,7 @@ class ServerArgs:
             "--api-key",
             type=str,
             default=ServerArgs.api_key,
-            help="Set API key of the server.",
+            help="Set API key of the server. It is also used in the OpenAI API compatible server.",
         )
         parser.add_argument(
             "--file-storage-pth",
@@ -317,6 +319,7 @@ class ServerArgs:
 
         # Data parallelism
         parser.add_argument(
+            "--data-parallel-size",
             "--dp-size",
             type=int,
             default=ServerArgs.dp_size,
@@ -393,6 +396,11 @@ class ServerArgs:
             action="store_true",
             help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
         )
+        parser.add_argument(
+            "--enable-mla",
+            action="store_true",
+            help="Enable Multi-head Latent Attention (MLA) for DeepSeek-V2",
+        )
         parser.add_argument(
             "--attention-reduce-in-fp32",
             action="store_true",
@@ -407,6 +415,8 @@ class ServerArgs:
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):
+        args.tp_size = args.tensor_parallel_size
+        args.dp_size = args.data_parallel_size
         attrs = [attr.name for attr in dataclasses.fields(cls)]
         return cls(**{attr: getattr(args, attr) for attr in attrs})
 

sglang/srt/utils.py

@@ -539,26 +539,6 @@ class CustomCacheManager(FileCacheManager):
                 raise RuntimeError("Could not create or locate cache dir")
 
 
-API_KEY_HEADER_NAME = "X-API-Key"
-
-
-class APIKeyValidatorMiddleware(BaseHTTPMiddleware):
-    def __init__(self, app, api_key: str):
-        super().__init__(app)
-        self.api_key = api_key
-
-    async def dispatch(self, request, call_next):
-        # extract API key from the request headers
-        api_key_header = request.headers.get(API_KEY_HEADER_NAME)
-        if not api_key_header or api_key_header != self.api_key:
-            return JSONResponse(
-                status_code=403,
-                content={"detail": "Invalid API Key"},
-            )
-        response = await call_next(request)
-        return response
-
-
 def get_ip_address(ifname):
     """
     Get the IP address of a network interface.
@@ -642,6 +622,19 @@ def receive_addrs(model_port_args, server_args):
     dist.destroy_process_group()
 
 
+def set_torch_compile_config():
+    # The following configurations are for torch compile optimizations
+    import torch._dynamo.config
+    import torch._inductor.config
+
+    torch._inductor.config.coordinate_descent_tuning = True
+    torch._inductor.config.triton.unique_kernel_names = True
+    torch._inductor.config.fx_graph_cache = True  # Experimental feature to reduce compilation times, will be on by default in future
+
+    # FIXME: tmp workaround
+    torch._dynamo.config.accumulated_cache_size_limit = 256
+
+
 def set_ulimit(target_soft_limit=65535):
     resource_type = resource.RLIMIT_NOFILE
     current_soft, current_hard = resource.getrlimit(resource_type)
@@ -700,3 +693,15 @@ def monkey_patch_vllm_qvk_linear_loader():
         origin_weight_loader(self, param, loaded_weight, loaded_shard_id)
 
     setattr(QKVParallelLinear, "weight_loader", weight_loader_srt)
+
+
+def add_api_key_middleware(app, api_key):
+    @app.middleware("http")
+    async def authentication(request, call_next):
+        if request.method == "OPTIONS":
+            return await call_next(request)
+        if request.url.path.startswith("/health"):
+            return await call_next(request)
+        if request.headers.get("Authorization") != "Bearer " + api_key:
+            return JSONResponse(content={"error": "Unauthorized"}, status_code=401)
+        return await call_next(request)

sglang/test/run_eval.py

@@ -10,7 +10,6 @@ import time
 
 from sglang.test.simple_eval_common import (
     ChatCompletionSampler,
-    download_dataset,
     make_report,
     set_ulimit,
 )
@@ -27,14 +26,26 @@ def run_eval(args):
     if args.eval_name == "mmlu":
         from sglang.test.simple_eval_mmlu import MMLUEval
 
-        dataset_path = "mmlu.csv"
-
-        if not os.path.exists(dataset_path):
-            download_dataset(
-                dataset_path,
-                "https://openaipublic.blob.core.windows.net/simple-evals/mmlu.csv",
-            )
-        eval_obj = MMLUEval(dataset_path, args.num_examples, args.num_threads)
+        filename = "https://openaipublic.blob.core.windows.net/simple-evals/mmlu.csv"
+        eval_obj = MMLUEval(filename, args.num_examples, args.num_threads)
+    elif args.eval_name == "math":
+        from sglang.test.simple_eval_math import MathEval
+
+        equality_checker = ChatCompletionSampler(model="gpt-4-turbo")
+
+        filename = (
+            "https://openaipublic.blob.core.windows.net/simple-evals/math_test.csv"
+        )
+        eval_obj = MathEval(
+            filename, equality_checker, args.num_examples, args.num_threads
+        )
+    elif args.eval_name == "gpqa":
+        from sglang.test.simple_eval_gpqa import GPQAEval
+
+        filename = (
+            "https://openaipublic.blob.core.windows.net/simple-evals/gpqa_diamond.csv"
+        )
+        eval_obj = GPQAEval(filename, args.num_examples, args.num_threads)
     elif args.eval_name == "humaneval":
         from sglang.test.simple_eval_humaneval import HumanEval
 
@@ -97,7 +108,7 @@ if __name__ == "__main__":
     )
     parser.add_argument("--eval-name", type=str, default="mmlu")
     parser.add_argument("--num-examples", type=int)
-    parser.add_argument("--num-threads", type=int, default=64)
+    parser.add_argument("--num-threads", type=int, default=512)
     set_ulimit()
     args = parser.parse_args()