[benchmark] update tools (#7211)

benchmarks/backend_request_func.py

@@ -269,6 +269,15 @@ async def async_request_eb_openai_chat_completions(
     if request_func_input.response_format:
         payload["response_format"] = request_func_input.response_format
 
+    # 随机输入开关
+    if request_func_input.random_flag:
+        payload["max_tokens"] = request_func_input.output_len
+        payload["min_tokens"] = request_func_input.output_len
+        # 随机token_ids场景
+        if isinstance(request_func_input.prompt, list):
+            request_func_input.prompt_token_ids = request_func_input.prompt
+            request_func_input.prompt = ""
+
     # 支持传入prompt_token_ids
     if request_func_input.prompt_token_ids:
         # 不走messages

benchmarks/benchmark_dataset.py

@@ -818,3 +818,61 @@ class RandomTextDataset(BenchmarkDataset):
                 )
             )
         return samples
+
+
+class RandomTokenDataset(BenchmarkDataset):
+    """
+    Generates random English words for pure text benchmarking.
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def sample(
+        self,
+        num_requests: int,
+        lora_path: Optional[str] = None,
+        max_loras: Optional[int] = None,
+        random_input_len: Optional[int] = None,
+        random_output_len: Optional[int] = None,
+        random_range_ratio: Optional[float] = None,
+        enable_multimodal_chat: bool = False,
+        **kwargs,
+    ) -> list:
+        samples = []
+
+        def sample_len(base_len: int, ratio: float) -> int:
+            if base_len is None:
+                return None
+            if ratio is None or ratio <= 0:
+                return base_len
+            lo = max(1, int(base_len * (1 - ratio)))
+            hi = int(base_len * (1 + ratio))
+            return random.randint(lo, hi)
+
+        for i in range(1, num_requests + 1):
+            # [length * (1 - range_ratio), length * (1 + range_ratio)]
+            sampled_input_len = sample_len(random_input_len, random_range_ratio)
+            sampled_output_len = sample_len(random_output_len, random_range_ratio)
+
+            random.seed(21)
+            token_ids = [random.randint(2000, 10000) for _ in range(sampled_input_len)]
+            # prompt_text = " ".join(words)
+
+            data = {
+                "messages": [{"role": "user", "content": [{"type": "text", "text": ""}]}],
+                "prompt_token_ids": token_ids,
+            }
+
+            samples.append(
+                SampleRequest(
+                    no=i,
+                    json_data=data,
+                    prompt=token_ids,
+                    prompt_len=sampled_input_len,
+                    history_QA=data["messages"],
+                    expected_output_len=sampled_output_len,
+                    random_flag=True,
+                )
+            )
+        return samples

benchmarks/benchmark_serving.py

@@ -39,7 +39,13 @@ from backend_request_func import (
     RequestFuncInput,
     RequestFuncOutput,
 )
-from benchmark_dataset import EBChatDataset, EBDataset, RandomTextDataset, SampleRequest
+from benchmark_dataset import (
+    EBChatDataset,
+    EBDataset,
+    RandomTextDataset,
+    RandomTokenDataset,
+    SampleRequest,
+)
 from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from tqdm.asyncio import tqdm
 
@@ -1156,6 +1162,12 @@ def main(args: argparse.Namespace):
             random_output_len=args.random_output_len,
             random_range_ratio=args.random_range_ratio,
         ),
+        "random_token_ids": lambda: RandomTokenDataset().sample(
+            num_requests=args.num_prompts,
+            random_input_len=args.random_input_len,
+            random_output_len=args.random_output_len,
+            random_range_ratio=args.random_range_ratio,
+        ),
     }
 
     try:
@@ -1338,6 +1350,7 @@ if __name__ == "__main__":
             "EB",
             "EBChat",
             "random",
+            "random_token_ids",
         ],
         help="Name of the dataset to benchmark on.",
     )