sglang 0.4.6.post3__py3-none-any.whl → 0.4.6.post5__py3-none-any.whl

sglang/bench_offline_throughput.py

@@ -17,11 +17,12 @@ import logging
 import os
 import random
 import time
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, List, Optional
 
 import numpy as np
 
 from sglang.bench_serving import (
+    DatasetRow,
     get_dataset,
     get_tokenizer,
     sample_random_requests,
@@ -194,7 +195,7 @@ class BenchArgs:
 def throughput_test_once(
     backend_name: str,
     backend,
-    reqs: List[Tuple[str, int, int]],
+    reqs: List[DatasetRow],
     ignore_eos: bool,
     extra_request_body: Dict,
     profile: bool,
@@ -203,7 +204,7 @@ def throughput_test_once(
         "backend": backend_name,
         "successful_requests": len(reqs),
         "total_latency": -1,
-        "total_input_tokens": sum(r[1] for r in reqs),
+        "total_input_tokens": sum(r.prompt_len for r in reqs),
         "total_output_tokens": -1,
         "request_throughput": -1,
         "input_throughput": -1,
@@ -211,11 +212,11 @@ def throughput_test_once(
         "total_throughput": -1,
     }
 
-    prompt = [r[0] for r in reqs]
+    prompt = [r.prompt for r in reqs]
     sampling_params = [
         {
             "temperature": 0,
-            "max_new_tokens": r[2],
+            "max_new_tokens": r.output_len,
             "ignore_eos": ignore_eos,
             **extra_request_body,
         }
@@ -259,13 +260,14 @@ 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
 
 
 def monitor_trace_file(directory, interval=1):
-
     print(f"Monitoring {directory} for new trace files...")
 
     known_files = set(os.listdir(directory))
@@ -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
 
@@ -269,6 +269,7 @@ def _maybe_prepare_dp_attn_batch(batch: ScheduleBatch, model_runner):
             batch,
             dp_size=model_runner.server_args.dp_size,
             attn_tp_size=1,
+            moe_dense_tp_size=model_runner.server_args.moe_dense_tp_size,
             tp_cpu_group=model_runner.tp_group.cpu_group,
             get_idle_batch=None,
             disable_cuda_graph=model_runner.server_args.disable_cuda_graph,
@@ -372,10 +373,10 @@ def latency_test_run_once(
 
     # Prefill
     synchronize(device)
-    tic = time.time()
+    tic = time.perf_counter()
     next_token_ids, _, batch = extend(reqs, model_runner)
     synchronize(device)
-    prefill_latency = time.time() - tic
+    prefill_latency = time.perf_counter() - tic
     tot_latency += prefill_latency
     throughput = input_len * batch_size / prefill_latency
     rank_print(
@@ -388,10 +389,10 @@ def latency_test_run_once(
     decode_latencies = []
     for i in range(output_len - 1):
         synchronize(device)
-        tic = time.time()
+        tic = time.perf_counter()
         next_token_ids, _ = decode(next_token_ids, batch, model_runner)
         synchronize(device)
-        latency = time.time() - tic
+        latency = time.perf_counter() - tic
         tot_latency += latency
         throughput = batch_size / latency
         decode_latencies.append(latency)

sglang/bench_one_batch_server.py

@@ -22,9 +22,11 @@ from typing import Tuple
 import numpy as np
 import requests
 
+from sglang.bench_serving import get_tokenizer, sample_random_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 +35,13 @@ 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 +55,19 @@ 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):
@@ -79,8 +93,8 @@ def launch_server_process(server_args: ServerArgs):
     base_url = f"http://{server_args.host}:{server_args.port}"
     timeout = 600
 
-    start_time = time.time()
-    while time.time() - start_time < timeout:
+    start_time = time.perf_counter()
+    while time.perf_counter() - start_time < timeout:
         try:
             headers = {
                 "Content-Type": "application/json; charset=utf-8",
@@ -99,36 +113,91 @@ 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,
+    tokenizer,
 ):
-    input_ids = [
-        [int(x) for x in np.random.randint(0, high=16384, size=(input_len,))]
-        for _ in range(batch_size)
-    ]
+    requests.post(url + "/flush_cache")
+    input_requests = sample_random_requests(
+        input_len=input_len,
+        output_len=output_len,
+        num_prompts=batch_size,
+        range_ratio=1.0,
+        tokenizer=tokenizer,
+        dataset_path="",
+        random_sample=True,
+        return_text=False,
+    )
+
+    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()
+    tic = time.perf_counter()
     response = requests.post(
         url + "/generate",
         json={
-            "input_ids": input_ids,
+            "input_ids": [req.prompt for req in input_requests],
             "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.perf_counter() - tic
+
+    latency = time.perf_counter() - 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 +209,21 @@ 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:
@@ -150,29 +231,45 @@ def run_benchmark(server_args: ServerArgs, bench_args: BenchArgs):
     else:
         proc, base_url = launch_server_process(server_args)
 
+    tokenizer_id = server_args.tokenizer_path or server_args.model_path
+    tokenizer = get_tokenizer(tokenizer_id)
+
     # 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="",
+            tokenizer=tokenizer,
         )
+        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,
+                    tokenizer=tokenizer,
+                )
             )
     finally:
         if proc:
@@ -180,6 +277,45 @@ 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()