sglang 0.4.1.post6__py3-none-any.whl → 0.4.1.post7__py3-none-any.whl

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/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/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
 
 
@@ -39,14 +40,15 @@ 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
     disable_ignore_eos: bool = False
     extra_request_body: Optional[str] = None
     seed: int = 1
@@ -82,6 +84,12 @@ 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,
@@ -102,35 +110,35 @@ 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(

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
 
@@ -56,12 +57,12 @@ 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
@@ -77,6 +78,8 @@ 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 +98,19 @@ 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. The endpoint must be launched with "
+            "SGLANG_TORCH_PROFILER_DIR to enable 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 +232,7 @@ 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 +303,16 @@ 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 +334,17 @@ 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 +375,13 @@ 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)
+
     # Record decode timing from 2nd output
     if output_len > 1:
         med_decode_latency = np.median(decode_latencies)
@@ -386,6 +430,8 @@ 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 +451,8 @@ def latency_test(
             il,
             ol,
             server_args.device,
+            bench_args.profile,
+            bench_args.profile_filename_prefix,
         )
         if ret is not None:
             result_list.append(ret)

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
 

sglang/bench_serving.py

@@ -452,6 +452,7 @@ def get_dataset(args, tokenizer):
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
             fixed_output_len=args.sharegpt_output_len,
+            context_len=args.sharegpt_context_len,
         )
     elif args.dataset_name == "random":
         input_requests = sample_random_requests(
@@ -464,11 +465,11 @@ def get_dataset(args, tokenizer):
         )
     elif args.dataset_name == "generated-shared-prefix":
         input_requests = sample_generated_shared_prefix_requests(
-            num_groups=args.gen_num_groups,
-            prompts_per_group=args.gen_prompts_per_group,
-            system_prompt_len=args.gen_system_prompt_len,
-            question_len=args.gen_question_len,
-            output_len=args.gen_output_len,
+            num_groups=args.gsp_num_groups,
+            prompts_per_group=args.gsp_prompts_per_group,
+            system_prompt_len=args.gsp_system_prompt_len,
+            question_len=args.gsp_question_len,
+            output_len=args.gsp_output_len,
             tokenizer=tokenizer,
         )
     else:
@@ -560,6 +561,7 @@ def sample_sharegpt_requests(
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
     fixed_output_len: Optional[int] = None,
+    context_len: Optional[int] = None,
 ) -> List[Tuple[str, int, int]]:
     if fixed_output_len is not None and fixed_output_len < 4:
         raise ValueError("output_len too small")
@@ -597,14 +599,15 @@ def sample_sharegpt_requests(
         output_len = (
             len(completion_token_ids) if fixed_output_len is None else fixed_output_len
         )
-        if prompt_len < 4 or output_len < 4:
+
+        if prompt_len < 1 or output_len < 1:
             # Prune too short sequences.
             continue
-        if prompt_len > 1024 or (
-            prompt_len + output_len > 2048 and fixed_output_len is None
-        ):
+
+        if context_len and prompt_len + output_len > context_len:
             # Prune too long sequences.
             continue
+
         filtered_dataset.append((prompt, prompt_len, output_len))
 
     print(f"#Input tokens: {np.sum([x[1] for x in filtered_dataset])}")
@@ -706,8 +709,8 @@ def get_gen_prefix_cache_path(args, tokenizer):
 
     # Create a unique cache filename based on the generation parameters
     cache_key = (
-        f"gen_prefix_{args.gen_num_groups}_{args.gen_prompts_per_group}_"
-        f"{args.gen_system_prompt_len}_{args.gen_question_len}_{args.gen_output_len}_"
+        f"gen_shared_prefix_{args.gsp_num_groups}_{args.gsp_prompts_per_group}_"
+        f"{args.gsp_system_prompt_len}_{args.gsp_question_len}_{args.gsp_output_len}_"
         f"{tokenizer.__class__.__name__}.pkl"
     )
     return cache_dir / cache_key
@@ -1374,6 +1377,12 @@ if __name__ == "__main__":
         default=None,
         help="Output length for each request. Overrides the output length from the ShareGPT dataset.",
     )
+    parser.add_argument(
+        "--sharegpt-context-len",
+        type=int,
+        default=None,
+        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,
@@ -1453,49 +1462,49 @@ if __name__ == "__main__":
         help="Append given JSON object to the request payload. You can use this to specify"
         "additional generate params like sampling params.",
     )
+    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(
+        "--lora-name",
+        type=str,
+        default=None,
+        help="The name of LoRA adapter",
+    )
 
     group = parser.add_argument_group("generated-shared-prefix dataset arguments")
     group.add_argument(
-        "--gen-num-groups",
+        "--gsp-num-groups",
         type=int,
         default=64,
         help="Number of system prompt groups for generated-shared-prefix dataset",
     )
     group.add_argument(
-        "--gen-prompts-per-group",
+        "--gsp-prompts-per-group",
         type=int,
         default=16,
         help="Number of prompts per system prompt group for generated-shared-prefix dataset",
     )
     group.add_argument(
-        "--gen-system-prompt-len",
+        "--gsp-system-prompt-len",
         type=int,
         default=2048,
         help="Target length in tokens for system prompts in generated-shared-prefix dataset",
     )
     group.add_argument(
-        "--gen-question-len",
+        "--gsp-question-len",
         type=int,
         default=128,
         help="Target length in tokens for questions in generated-shared-prefix dataset",
     )
     group.add_argument(
-        "--gen-output-len",
+        "--gsp-output-len",
         type=int,
         default=256,
         help="Target length in tokens for outputs in generated-shared-prefix dataset",
     )
-    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(
-        "--lora-name",
-        type=str,
-        default=None,
-        help="The name of LoRA adapter",
-    )
     args = parser.parse_args()
     run_benchmark(args)

sglang/lang/backend/runtime_endpoint.py

@@ -1,6 +1,11 @@
+import atexit
 import json
+import multiprocessing
 import warnings
-from typing import List, Optional
+from typing import Dict, List, Optional, Union
+
+import aiohttp
+import requests
 
 from sglang.global_config import global_config
 from sglang.lang.backend.base_backend import BaseBackend
@@ -251,11 +256,12 @@ class RuntimeEndpoint(BaseBackend):
         }
         obj = self._generate_http_request(s, data)
 
-        normalized_prompt_logprobs = [
-            r["meta_info"]["normalized_prompt_logprob"] for r in obj
-        ]
         input_token_logprobs = [r["meta_info"]["input_token_logprobs"] for r in obj]
         output_token_logprobs = [r["meta_info"]["output_token_logprobs"] for r in obj]
+        normalized_prompt_logprobs = [
+            compute_normalized_prompt_logprobs(r["meta_info"]["input_token_logprobs"])
+            for r in obj
+        ]
 
         # Remove extra token if no token healing occurred
         for i in range(len(input_token_logprobs)):
@@ -319,3 +325,176 @@ class RuntimeEndpoint(BaseBackend):
     def _assert_success(self, res):
         if res.status_code != 200:
             raise RuntimeError(res.json())
+
+
+def compute_normalized_prompt_logprobs(input_logprobs):
+    values = [x[0] for x in input_logprobs if x[0]]
+    return sum(values) / len(values)
+
+
+class Runtime:
+    """
+    A wrapper for the HTTP server.
+    This is used for launching the server in a python program without
+    using the commond line interface.
+
+    It is mainly used for the frontend language.
+    You should use the Engine class if you want to do normal offline processing without the frontend language.
+    """
+
+    def __init__(
+        self,
+        log_level: str = "error",
+        *args,
+        **kwargs,
+    ):
+        """See the arguments in server_args.py::ServerArgs"""
+        # We delay the import of any `sglang.srt` components in `sglang.lang`, so users can run
+        # client code without installing SRT server and its dependency if they want.
+        from sglang.srt.entrypoints.http_server import launch_server
+        from sglang.srt.server_args import ServerArgs
+        from sglang.srt.utils import is_port_available
+
+        self.server_args = ServerArgs(*args, log_level=log_level, **kwargs)
+
+        # Pre-allocate ports
+        for port in range(self.server_args.port, 40000):
+            if is_port_available(port):
+                break
+        self.server_args.port = port
+
+        self.url = self.server_args.url()
+        self.generate_url = self.url + "/generate"
+
+        # NOTE: We store pid instead of proc to fix some issues during __delete__
+        self.pid = None
+        pipe_reader, pipe_writer = multiprocessing.Pipe(duplex=False)
+
+        proc = multiprocessing.Process(
+            target=launch_server,
+            args=(self.server_args, pipe_writer),
+        )
+        proc.start()
+        pipe_writer.close()
+        self.pid = proc.pid
+
+        # Before python program terminates, call shutdown implicitly. Therefore, users don't have to explicitly call .shutdown()
+        atexit.register(self.shutdown)
+
+        # TODO: remove this pipe_writer mechanism and use `/health_generate` instead.
+        try:
+            init_state = pipe_reader.recv()
+        except EOFError:
+            init_state = ""
+
+        if init_state != "ready":
+            self.shutdown()
+            raise RuntimeError(
+                "Initialization failed. Please see the error messages above."
+            )
+
+        self.endpoint = RuntimeEndpoint(self.url)
+
+    def shutdown(self):
+        from sglang.srt.utils import kill_process_tree
+
+        if self.pid is not None:
+            kill_process_tree(self.pid)
+            self.pid = None
+
+    def cache_prefix(self, prefix: str):
+        self.endpoint.cache_prefix(prefix)
+
+    def get_tokenizer(self):
+        from sglang.srt.hf_transformers_utils import get_tokenizer
+
+        return get_tokenizer(
+            self.server_args.tokenizer_path,
+            tokenizer_mode=self.server_args.tokenizer_mode,
+            trust_remote_code=self.server_args.trust_remote_code,
+            revision=self.server_args.revision,
+        )
+
+    async def async_generate(
+        self,
+        prompt: str,
+        sampling_params: Optional[Dict] = None,
+    ):
+        if self.server_args.skip_tokenizer_init:
+            json_data = {
+                "input_ids": prompt,
+                "sampling_params": sampling_params,
+                "stream": True,
+            }
+        else:
+            json_data = {
+                "text": prompt,
+                "sampling_params": sampling_params,
+                "stream": True,
+            }
+        pos = 0
+
+        timeout = aiohttp.ClientTimeout(total=3 * 3600)
+        async with aiohttp.ClientSession(timeout=timeout, trust_env=True) as session:
+            async with session.post(self.generate_url, json=json_data) as response:
+                async for chunk, _ in response.content.iter_chunks():
+                    chunk = chunk.decode("utf-8")
+                    if chunk and chunk.startswith("data:"):
+                        if chunk == "data: [DONE]\n\n":
+                            break
+                        data = json.loads(chunk[5:].strip("\n"))
+                        if "text" in data:
+                            cur = data["text"][pos:]
+                            if cur:
+                                yield cur
+                            pos += len(cur)
+                        else:
+                            yield data
+
+    add_request = async_generate
+
+    def generate(
+        self,
+        prompt: Union[str, List[str]],
+        sampling_params: Optional[Dict] = None,
+        return_logprob: Optional[Union[List[bool], bool]] = False,
+        logprob_start_len: Optional[Union[List[int], int]] = None,
+        top_logprobs_num: Optional[Union[List[int], int]] = None,
+        lora_path: Optional[List[Optional[str]]] = None,
+    ):
+        json_data = {
+            "text": prompt,
+            "sampling_params": sampling_params,
+            "return_logprob": return_logprob,
+            "logprob_start_len": logprob_start_len,
+            "top_logprobs_num": top_logprobs_num,
+            "lora_path": lora_path,
+        }
+        assert not isinstance(lora_path, list) or len(lora_path) == len(prompt)
+        response = requests.post(
+            self.url + "/generate",
+            json=json_data,
+        )
+        return json.dumps(response.json())
+
+    def encode(
+        self,
+        prompt: Union[str, List[str], List[Dict], List[List[Dict]]],
+    ):
+        json_data = {"text": prompt}
+        response = requests.post(self.url + "/encode", json=json_data)
+        return json.dumps(response.json())
+
+    async def get_server_info(self):
+        async with aiohttp.ClientSession() as session:
+            async with session.get(f"{self.url}/get_server_info") as response:
+                if response.status == 200:
+                    return await response.json()
+                else:
+                    error_data = await response.json()
+                    raise RuntimeError(
+                        f"Failed to get server info. {error_data['error']['message']}"
+                    )
+
+    def __del__(self):
+        self.shutdown()