sglang 0.4.6.post3__py3-none-any.whl → 0.4.6.post4__py3-none-any.whl

sglang/bench_offline_throughput.py

@@ -259,7 +259,9 @@ def throughput_test_once(
         measurement_results["total_input_tokens"]
         + measurement_results["total_output_tokens"]
     ) / latency
-    measurement_results["last_gen_throughput"] = server_info["last_gen_throughput"]
+    measurement_results["last_gen_throughput"] = server_info["internal_states"][0][
+        "last_gen_throughput"
+    ]
 
     return measurement_results
 
@@ -315,7 +317,7 @@ def throughput_test(
     tokenizer_id = server_args.tokenizer_path or server_args.model_path
     tokenizer = get_tokenizer(tokenizer_id)
 
-    # Set global environmnets
+    # Set global environments
     set_ulimit()
     random.seed(bench_args.seed)
     np.random.seed(bench_args.seed)

sglang/bench_one_batch.py

@@ -246,7 +246,7 @@ def extend(reqs, model_runner):
     _maybe_prepare_dp_attn_batch(batch, model_runner)
     model_worker_batch = batch.get_model_worker_batch()
     forward_batch = ForwardBatch.init_new(model_worker_batch, model_runner)
-    logits_output = model_runner.forward(forward_batch)
+    logits_output, _ = model_runner.forward(forward_batch)
     next_token_ids = model_runner.sample(logits_output, forward_batch)
     return next_token_ids, logits_output.next_token_logits, batch
 
@@ -258,7 +258,7 @@ def decode(input_token_ids, batch, model_runner):
     _maybe_prepare_dp_attn_batch(batch, model_runner)
     model_worker_batch = batch.get_model_worker_batch()
     forward_batch = ForwardBatch.init_new(model_worker_batch, model_runner)
-    logits_output = model_runner.forward(forward_batch)
+    logits_output, _ = model_runner.forward(forward_batch)
     next_token_ids = model_runner.sample(logits_output, forward_batch)
     return next_token_ids, logits_output.next_token_logits
 

sglang/bench_one_batch_server.py

@@ -25,6 +25,7 @@ import requests
 from sglang.srt.entrypoints.http_server import launch_server
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import kill_process_tree
+from sglang.test.test_utils import is_in_ci, write_github_step_summary
 
 
 @dataclasses.dataclass
@@ -33,9 +34,13 @@ class BenchArgs:
     batch_size: Tuple[int] = (1,)
     input_len: Tuple[int] = (1024,)
     output_len: Tuple[int] = (16,)
+    temperature: float = 0.0
+    return_logprob: bool = False
+    input_len_step_percentage: float = 0.0
     result_filename: str = "result.jsonl"
     base_url: str = ""
     skip_warmup: bool = False
+    show_report: bool = False
 
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser):
@@ -49,11 +54,19 @@ class BenchArgs:
         parser.add_argument(
             "--output-len", type=int, nargs="+", default=BenchArgs.output_len
         )
+        parser.add_argument("--temperature", type=float, default=BenchArgs.temperature)
+        parser.add_argument("--return-logprob", action="store_true")
+        parser.add_argument(
+            "--input-len-step-percentage",
+            type=float,
+            default=BenchArgs.input_len_step_percentage,
+        )
         parser.add_argument(
             "--result-filename", type=str, default=BenchArgs.result_filename
         )
         parser.add_argument("--base-url", type=str, default=BenchArgs.base_url)
         parser.add_argument("--skip-warmup", action="store_true")
+        parser.add_argument("--show-report", action="store_true")
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):
@@ -99,36 +112,89 @@ def run_one_case(
     batch_size: int,
     input_len: int,
     output_len: int,
+    temperature: float,
+    return_logprob: bool,
+    input_len_step_percentage: float,
     run_name: str,
     result_filename: str,
 ):
+    requests.post(url + "/flush_cache")
+    input_lens = [
+        int(input_len * (1 + (i - (batch_size - 1) / 2) * input_len_step_percentage))
+        for i in range(batch_size)
+    ]
     input_ids = [
-        [int(x) for x in np.random.randint(0, high=16384, size=(input_len,))]
-        for _ in range(batch_size)
+        [int(x) for x in np.random.randint(0, high=16384, size=(input_lens[i],))]
+        for i in range(batch_size)
     ]
 
+    use_structured_outputs = False
+    if use_structured_outputs:
+        texts = []
+        for _ in range(batch_size):
+            texts.append(
+                "Human: What is the capital city of france? can you give as many trivial information as possible about that city? answer in json.\n"
+                * 50
+                + "Assistant:"
+            )
+        json_schema = "$$ANY$$"
+    else:
+        json_schema = None
+
     tic = time.time()
     response = requests.post(
         url + "/generate",
         json={
+            # "text": texts,
             "input_ids": input_ids,
             "sampling_params": {
-                "temperature": 0,
+                "temperature": temperature,
                 "max_new_tokens": output_len,
                 "ignore_eos": True,
+                "json_schema": json_schema,
             },
+            "return_logprob": return_logprob,
+            "stream": True,
         },
+        stream=True,
     )
-    latency = time.time() - tic
 
-    _ = response.json()
-    output_throughput = batch_size * output_len / latency
+    # The TTFT of the last request in the batch
+    ttft = 0.0
+    for chunk in response.iter_lines(decode_unicode=False):
+        chunk = chunk.decode("utf-8")
+        if chunk and chunk.startswith("data:"):
+            if chunk == "data: [DONE]":
+                break
+            data = json.loads(chunk[5:].strip("\n"))
+            if "error" in data:
+                raise RuntimeError(f"Request has failed. {data}.")
+
+            assert (
+                data["meta_info"]["finish_reason"] is None
+                or data["meta_info"]["finish_reason"]["type"] == "length"
+            )
+            if data["meta_info"]["completion_tokens"] == 1:
+                ttft = time.time() - tic
+
+    latency = time.time() - tic
+    input_throughput = batch_size * input_len / ttft
+    output_throughput = batch_size * output_len / (latency - ttft)
     overall_throughput = batch_size * (input_len + output_len) / latency
 
+    server_info = requests.get(url + "/get_server_info").json()
+    acc_length = server_info["internal_states"][0].get("avg_spec_accept_length", None)
+    last_gen_throughput = server_info["internal_states"][0]["last_gen_throughput"]
+
     print(f"batch size: {batch_size}")
+    print(f"input_len: {input_len}")
+    print(f"output_len: {output_len}")
     print(f"latency: {latency:.2f} s")
-    print(f"output throughput: {output_throughput:.2f} token/s")
-    print(f"(input + output) throughput: {overall_throughput:.2f} token/s")
+    print(f"ttft: {ttft:.2f} s")
+    print(f"Last generation throughput: {last_gen_throughput:.2f} tok/s")
+    print(f"Input throughput: {input_throughput:.2f} tok/s")
+    if output_len != 1:
+        print(f"output throughput: {output_throughput:.2f} tok/s")
 
     if result_filename:
         with open(result_filename, "a") as fout:
@@ -140,9 +206,21 @@ def run_one_case(
                 "latency": round(latency, 4),
                 "output_throughput": round(output_throughput, 2),
                 "overall_throughput": round(overall_throughput, 2),
+                "last_gen_throughput": round(last_gen_throughput, 2),
             }
             fout.write(json.dumps(res) + "\n")
 
+    return (
+        batch_size,
+        latency,
+        ttft,
+        input_throughput,
+        output_throughput,
+        overall_throughput,
+        last_gen_throughput,
+        acc_length,
+    )
+
 
 def run_benchmark(server_args: ServerArgs, bench_args: BenchArgs):
     if bench_args.base_url:
@@ -152,27 +230,38 @@ def run_benchmark(server_args: ServerArgs, bench_args: BenchArgs):
 
     # warmup
     if not bench_args.skip_warmup:
+        print("=" * 8 + " Warmup Begin " + "=" * 8)
         run_one_case(
             base_url,
             batch_size=16,
             input_len=1024,
             output_len=16,
+            temperature=bench_args.temperature,
+            return_logprob=bench_args.return_logprob,
+            input_len_step_percentage=bench_args.input_len_step_percentage,
             run_name="",
             result_filename="",
         )
+        print("=" * 8 + " Warmup End   " + "=" * 8 + "\n")
 
     # benchmark
+    result = []
     try:
         for bs, il, ol in itertools.product(
             bench_args.batch_size, bench_args.input_len, bench_args.output_len
         ):
-            run_one_case(
-                base_url,
-                bs,
-                il,
-                ol,
-                bench_args.run_name,
-                bench_args.result_filename,
+            result.append(
+                run_one_case(
+                    base_url,
+                    bs,
+                    il,
+                    ol,
+                    temperature=bench_args.temperature,
+                    return_logprob=bench_args.return_logprob,
+                    input_len_step_percentage=bench_args.input_len_step_percentage,
+                    run_name=bench_args.run_name,
+                    result_filename=bench_args.result_filename,
+                )
             )
     finally:
         if proc:
@@ -180,6 +269,45 @@ def run_benchmark(server_args: ServerArgs, bench_args: BenchArgs):
 
     print(f"\nResults are saved to {bench_args.result_filename}")
 
+    if not bench_args.show_report:
+        return
+
+    summary = " | batch size | latency (s) | input throughput (tok/s)  | output throughput (tok/s) | acc length | ITL (ms) | input price ($/1M) | output price ($/1M) |\n"
+    summary += "| ---------- | ----------- | ------------------------- | ------------------------- | ---------- | -------- | ------------------ | ------------------- |\n"
+
+    for (
+        batch_size,
+        latency,
+        ttft,
+        input_throughput,
+        output_throughput,
+        overall_throughput,
+        last_gen_throughput,
+        acc_length,
+    ) in result:
+        hourly_cost = 2 * server_args.tp_size  # $2/hour for one H100
+        input_util = 0.7
+        accept_length = round(acc_length, 2) if acc_length is not None else "n/a"
+        line = (
+            f"| {batch_size} | "
+            f"{latency:.2f} | "
+            f"{input_throughput:.2f} | "
+            f"{output_throughput:.2f} | "
+            f"{accept_length} | "
+            f"{1 / (output_throughput/batch_size) * 1000:.2f} | "
+            f"{1e6 / (input_throughput * input_util) / 3600 * hourly_cost:.2f} | "
+            f"{1e6 / output_throughput / 3600 * hourly_cost:.2f} |\n"
+        )
+        summary += line
+
+    # print metrics table
+    print(summary)
+
+    if is_in_ci():
+        write_github_step_summary(
+            f"### Test Nightly Benchmark (bench_one_batch) \n{summary}"
+        )
+
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()

sglang/bench_serving.py

@@ -1103,7 +1103,7 @@ async def benchmark(
     lora_names: List[str],
     extra_request_body: Dict[str, Any],
     profile: bool,
-    pd_seperated: bool = False,
+    pd_separated: bool = False,
     flush_cache: bool = False,
     warmup_requests: int = 1,
 ):
@@ -1239,12 +1239,14 @@ async def benchmark(
 
     if "sglang" in backend:
         server_info = requests.get(base_url + "/get_server_info")
-        if pd_seperated:
-            accept_length = server_info.json()["decode"][0].get(
+        if pd_separated:
+            accept_length = server_info.json()["decode"][0]["internal_states"][0].get(
                 "avg_spec_accept_length", None
             )
         else:
-            accept_length = server_info.json().get("avg_spec_accept_length", None)
+            accept_length = server_info.json()["internal_states"][0].get(
+                "avg_spec_accept_length", None
+            )
     else:
         accept_length = None
 
@@ -1263,7 +1265,7 @@ async def benchmark(
     print("{:<40} {:<10}".format("Traffic request rate:", request_rate))
     print(
         "{:<40} {:<10}".format(
-            "Max reqeuest concurrency:",
+            "Max request concurrency:",
             max_concurrency if max_concurrency else "not set",
         )
     )
@@ -1541,7 +1543,7 @@ def run_benchmark(args_: argparse.Namespace):
             lora_names=args.lora_name,
             extra_request_body=extra_request_body,
             profile=args.profile,
-            pd_seperated=args.pd_seperated,
+            pd_separated=args.pd_separated,
             flush_cache=args.flush_cache,
         )
     )
@@ -1720,7 +1722,7 @@ if __name__ == "__main__":
         help="Suffix applied to the end of all user prompts, followed by assistant prompt suffix.",
     )
     parser.add_argument(
-        "--pd-seperated",
+        "--pd-separated",
         action="store_true",
         help="Benchmark PD disaggregation server",
     )

sglang/compile_deep_gemm.py

@@ -129,7 +129,7 @@ def launch_server_process_and_send_one_request(
 
 
 def refine_server_args(server_args: ServerArgs, compile_args: CompileArgs):
-    # Disbale cuda graph and torch compile to save time
+    # Disable cuda graph and torch compile to save time
     server_args.disable_cuda_graph = True
     server_args.enable_torch_compile = False
     print(f"Disable CUDA Graph and Torch Compile to save time...")

sglang/eval/loogle_eval.py

@@ -0,0 +1,157 @@
+import argparse
+import asyncio
+import os
+import pickle
+from pathlib import Path
+from typing import List
+
+import openai
+import torch
+from bert_score import BERTScorer
+from datasets import load_dataset
+from tqdm import tqdm
+
+
+def get_client(api_url: str) -> openai.AsyncOpenAI:
+    if os.getenv("OPENAI_API_KEY") is None:
+        os.environ["OPENAI_API_KEY"] = "EMPTY"
+    return openai.AsyncOpenAI(base_url=api_url)
+
+
+def get_dataset():
+    return load_dataset("bigai-nlco/LooGLE", "longdep_qa", split="test")
+
+
+async def fetch_response(
+    client: openai.AsyncOpenAI,
+    context: str,
+    question: str,
+    semaphore: asyncio.Semaphore,
+    index: int,
+    model: str,
+    output_dir: Path,
+):
+    output_file = output_dir / f"response_{index}.pkl"
+    if output_file.exists():
+        return
+
+    prompt = (
+        "Please answer the question based on the long texts below.\n"
+        f"{context}\n"
+        f"Question: {question}\n"
+        "Answer:"
+    )
+    messages = [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": prompt},
+    ]
+
+    async with semaphore:
+        try:
+            response = await client.chat.completions.create(
+                model=model,
+                messages=messages,
+                temperature=0.0,
+                max_tokens=512,
+            )
+        except openai.BadRequestError as e:
+            with open(output_file, "wb") as f:
+                pickle.dump({"error": str(e)}, f)
+            return
+
+    with open(output_file, "wb") as f:
+        pickle.dump(response, f)
+
+
+async def benchmark(args):
+    dataset = get_dataset()
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    client = get_client(args.api_url)
+    semaphore = asyncio.Semaphore(args.max_concurrency)
+
+    tasks: List[asyncio.Task] = []
+    for idx, ex in enumerate(dataset):
+        tasks.append(
+            asyncio.create_task(
+                fetch_response(
+                    client,
+                    ex["context"],
+                    ex["question"],
+                    semaphore,
+                    idx,
+                    args.model,
+                    output_dir,
+                )
+            )
+        )
+
+    for _ in tqdm(
+        asyncio.as_completed(tasks), total=len(tasks), desc="Running benchmark"
+    ):
+        await _
+
+
+def analyse(args):
+    dataset = get_dataset()
+    output_dir = Path(args.output_dir)
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    scorer = BERTScorer(lang="en", device=device)
+
+    hyps: List[str] = []
+    refs: List[str] = []
+    for idx, ex in enumerate(tqdm(dataset, desc="Loading responses")):
+        pkl_file = output_dir / f"response_{idx}.pkl"
+        if not pkl_file.exists():
+            raise FileNotFoundError(pkl_file)
+
+        response = pickle.load(open(pkl_file, "rb"))
+        if isinstance(response, dict) and "error" in response:
+            continue
+
+        hyps.append(response.choices[0].message.content.strip())
+        refs.append(ex["answer"])
+
+    if not hyps:
+        print("No valid responses to score!")
+        return
+
+    batch_size = 64
+    all_f1: List[float] = []
+    for i in tqdm(range(0, len(hyps), batch_size), desc="Scoring batches"):
+        h_batch = hyps[i : i + batch_size]
+        r_batch = refs[i : i + batch_size]
+        _, _, f1_scores = scorer.score(h_batch, r_batch, verbose=False)
+        all_f1.extend([float(x) for x in f1_scores])
+
+    avg = sum(all_f1) / len(all_f1)
+    print(f"Average BERTScore (F1): {avg:.2%}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Run benchmark and evaluation in one go."
+    )
+    parser.add_argument(
+        "--api-url",
+        default="http://127.0.0.1:30000/v1",
+        help="OpenAI‑compatible API base URL",
+    )
+    parser.add_argument(
+        "--model",
+        default="meta-llama/Llama-4-Maverick-17B-128E-Instruct",
+        help="Model name or ID, only used for model name",
+    )
+    parser.add_argument(
+        "--max-concurrency", type=int, default=144, help="Maximum concurrent requests"
+    )
+    parser.add_argument(
+        "--output-dir", default="tmp-output-dir", help="Directory for cached responses"
+    )
+    args = parser.parse_args()
+
+    asyncio.run(benchmark(args))
+
+    analyse(args)