Replace time.time() to time.perf_counter() for benchmarking. (#6178)

Signed-off-by: Lifu Huang <lifu.hlf@gmail.com>

benchmark/llm_judge/bench_sglang.py

@@ -59,7 +59,7 @@ def main(args):
     backend = select_sglang_backend(args)
 
     # Run requests
-    tic = time.time()
+    tic = time.perf_counter()
     states = multi_dimension_judge.run_batch(
         arguments,
         temperature=0,
@@ -67,7 +67,7 @@ def main(args):
         num_threads=args.parallel,
         progress_bar=True,
     )
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     print(f"Latency: {latency:.3f}")
 

benchmark/long_json_decode/bench_other.py

@@ -45,7 +45,7 @@ def main(args):
     def get_one_answer(i):
         states[i] = json_decode(generate=call_generate, **arguments[i])
 
-    tic = time.time()
+    tic = time.perf_counter()
     if args.parallel == 1:
         for i in tqdm(range(len(arguments))):
             get_one_answer(i)
@@ -58,7 +58,7 @@ def main(args):
                 )
             )
 
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     # Compute accuracy
     print(f"Latency: {latency:.3f}")

benchmark/long_json_decode/bench_sglang.py

@@ -46,11 +46,11 @@ def main(args):
     sgl.set_default_backend(backend)
 
     # Run requests
-    tic = time.time()
+    tic = time.perf_counter()
     states = json_decode.run_batch(
         arguments, temperature=0, num_threads=args.parallel, progress_bar=True
     )
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     # Compute accuracy
     print(f"Latency: {latency:.3f}")

benchmark/mmlu/bench_other.py

@@ -76,7 +76,7 @@ def evaluate(args, subject, dev_df, test_df, call_generate):
             pred = call_generate(prompts[i], temperature=0, max_tokens=max_tokens)
             preds[i] = pred.strip()[0]
 
-        tic = time.time()
+        tic = time.perf_counter()
         if args.parallel == 1:
             for i in range(len(prompts)):
                 get_one_answer(i)
@@ -94,9 +94,9 @@ def evaluate(args, subject, dev_df, test_df, call_generate):
                 for j in range(len(rets)):
                     preds[i + j] = rets[j].strip()[0]
 
-        tic = time.time()
+        tic = time.perf_counter()
         asyncio.run(batched_call(batch_size=args.parallel))
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     # Compute accuracy
     cors = [pred == label for pred, label in zip(preds, labels)]

benchmark/mmlu/bench_sglang.py

@@ -116,7 +116,7 @@ def main(args):
     backend = select_sglang_backend(args)
 
     # Run
-    tic = time.time()
+    tic = time.perf_counter()
     states = few_shot_mmlu.run_batch(
         arguments,
         temperature=0,
@@ -128,7 +128,7 @@ def main(args):
     preds = [
         s["answer"].strip()[0] if len(s["answer"].strip()) > 0 else "" for s in states
     ]
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     # Compute accuracy
     cors = [pred == label for pred, label in zip(preds, labels)]

benchmark/mmmu/bench_sglang.py

@@ -119,7 +119,7 @@ async def eval_mmmu(args) -> None:
         api_key="sk", base_url=f"http://127.0.0.1:{args.port}/v1"
     )
     semaphore = asyncio.Semaphore(args.concurrency)
-    start = time.time()
+    start = time.perf_counter()
     base_url = f"http://127.0.0.1:{args.port}"
 
     if args.profile:
@@ -147,7 +147,7 @@ async def eval_mmmu(args) -> None:
         if profile_output.success:
             print("Profiler stopped")
 
-    print(f"Benchmark time: {time.time() - start}")
+    print(f"Benchmark time: {time.perf_counter() - start}")
     args.output_path = f"./val_sglang.json"
     save_json(args.output_path, out_samples)
     eval_result(model_answer_path=args.output_path, answer_dict=answer_dict)

benchmark/mtbench/bench_other.py

@@ -66,7 +66,7 @@ def main(args):
         answers[i] = cur_answers
 
     # Run requests
-    tic = time.time()
+    tic = time.perf_counter()
     if args.parallel == 1:
         for i in tqdm(range(len(questions))):
             get_answer(i)
@@ -79,7 +79,7 @@ def main(args):
                 )
             )
 
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     print(f"#questions: {len(questions)}, Latency: {latency:.2f}")
 

benchmark/mtbench/bench_sglang.py

@@ -57,7 +57,7 @@ def main(args):
     sgl.set_default_backend(backend)
 
     # Run requests
-    tic = time.time()
+    tic = time.perf_counter()
     rets = answer_mt_bench.run_batch(
         arguments,
         temperature=0,
@@ -66,7 +66,7 @@ def main(args):
         progress_bar=True,
     )
     answers = [[s["answer_1"], s["answer_2"]] for s in rets]
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     print(f"#questions: {len(questions)}, Latency: {latency:.2f}")
 

benchmark/mtbench/bench_sglang_eagle.py

@@ -68,7 +68,7 @@ def main(args):
     sgl.set_default_backend(backend)
 
     # Run requests
-    tic = time.time()
+    tic = time.perf_counter()
     rets = answer_mt_bench.run_batch(
         arguments,
         temperature=0,
@@ -78,7 +78,7 @@ def main(args):
     )
     answers = [[s["answer_1"], s["answer_2"]] for s in rets]
 
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
     num_output_tokens = sum(
         s.get_meta_info("answer_1")["completion_tokens"]
         + s.get_meta_info("answer_2")["completion_tokens"]

benchmark/multi_chain_reasoning/bench_other.py

@@ -113,7 +113,7 @@ def main(args):
             answer = multi_chain_gsm8k(questions[i], args.num_chains, call_generate)
             states[i] = answer
 
-        tic = time.time()
+        tic = time.perf_counter()
         if args.parallel == 1:
             for i in tqdm(range(len(questions))):
                 get_one_answer(i)
@@ -134,7 +134,7 @@ def main(args):
             )
             states[i] = answer
 
-        tic = time.time()
+        tic = time.perf_counter()
         loop = asyncio.get_event_loop()
         batches = [
             list(range(i, min(i + args.parallel, len(questions))))
@@ -144,7 +144,7 @@ def main(args):
             tasks = [get_one_answer_asyncio(k) for k in bt]
             loop.run_until_complete(asyncio.gather(*tasks))
 
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     preds = []
     for i in range(len(states)):

benchmark/multi_chain_reasoning/bench_sglang.py

@@ -90,7 +90,7 @@ def main(args):
     backend = select_sglang_backend(args)
 
     # Run requests
-    tic = time.time()
+    tic = time.perf_counter()
     states = multi_chain_gsm8k.run_batch(
         arguments,
         temperature=0,
@@ -98,7 +98,7 @@ def main(args):
         num_threads=args.parallel,
         progress_bar=True,
     )
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     preds = []
     for i in range(len(states)):

benchmark/multi_document_qa/bench_other.py

@@ -61,7 +61,7 @@ def main(args):
     def get_one_answer(i):
         states[i] = multi_document_qa(generate=call_generate, **arguments[i])
 
-    tic = time.time()
+    tic = time.perf_counter()
     if args.parallel == 1:
         for i in tqdm(range(len(labels))):
             get_one_answer(i)
@@ -74,7 +74,7 @@ def main(args):
                 )
             )
 
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     # Compute accuracy
     print(states)

benchmark/multi_document_qa/bench_sglang.py

@@ -49,11 +49,11 @@ def main(args):
     sgl.set_default_backend(backend)
 
     # Run requests
-    tic = time.time()
+    tic = time.perf_counter()
     states = multi_document_qa.run_batch(
         arguments, temperature=0, num_threads=args.parallel, progress_bar=True
     )
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     # Compute accuracy
     print([s["answer"] for s in states])

benchmark/multi_turn_chat/bench_other.py

@@ -35,7 +35,7 @@ def main(args):
     def get_one_answer(i):
         states[i] = multi_turns(generate=call_generate, **multi_qas[i])
 
-    tic = time.time()
+    tic = time.perf_counter()
     if args.parallel == 1:
         for i in tqdm(range(len(multi_qas))):
             get_one_answer(i)
@@ -50,7 +50,7 @@ def main(args):
             for _ in rets:
                 pass
 
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     # Compute accuracy
     print(f"Latency: {latency:.3f}")

benchmark/multi_turn_chat/bench_sglang.py

@@ -27,7 +27,7 @@ def main(args):
 
     backend = select_sglang_backend(args)
 
-    tic = time.time()
+    tic = time.perf_counter()
     states = multi_turns.run_batch(
         multi_qas,
         temperature=0,
@@ -35,7 +35,7 @@ def main(args):
         num_threads=args.parallel,
         progress_bar=True,
     )
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     print(f"Latency: {latency:.3f}")
 

benchmark/multi_turn_chat/long_prompt_multi_turn.py

@@ -84,7 +84,7 @@ def main(args):
 
     backend = select_sglang_backend(args)
 
-    tic = time.time()
+    tic = time.perf_counter()
     states = multi_turns.run_batch(
         multi_qas,
         temperature=0,
@@ -92,7 +92,7 @@ def main(args):
         num_threads="auto",
         progress_bar=True,
     )
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     print(f"Latency: {latency:.3f}")
 

benchmark/react/bench_other.py

@@ -146,7 +146,7 @@ def main(args):
 
             states.append(answer)
 
-    tic = time.time()
+    tic = time.perf_counter()
 
     if args.backend != "lmql":
         if args.parallel == 1:
@@ -173,7 +173,7 @@ def main(args):
             tasks = [run_single_agent_async(arg) for arg in bt]
             loop.run_until_complete(asyncio.gather(*tasks))
 
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     print(f"Latency: {latency:.3f}")
 

benchmark/react/bench_sglang.py

@@ -115,14 +115,14 @@ def main(args):
     sgl.set_default_backend(backend)
 
     states = []
-    tic = time.time()
+    tic = time.perf_counter()
     states = webthink.run_batch(
         arguments,
         temperature=0,
         num_threads=args.parallel,
         progress_bar=True,
     )
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     # Compute accuracy
     print(f"Latency: {latency:.3f}")

benchmark/reasoning_benchmark/bench_sglang.py

@@ -51,7 +51,7 @@ def main(args):
     )
 
     # Run requests
-    tic = time.time()
+    tic = time.perf_counter()
     states = reasoning_gen.run_batch(
         questions,
         num_threads=args.parallel,
@@ -60,7 +60,7 @@ def main(args):
         max_new_tokens=32768,
         top_p=0.95,
     )
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     # Extract results and record outcomes in a list.
     outcomes = []

benchmark/tip_suggestion/bench_other.py

@@ -68,7 +68,7 @@ def main(args):
     call_generate = partial(get_call_generate(args), temperature=0)
 
     # Run requests
-    tic = time.time()
+    tic = time.perf_counter()
     if args.backend != "lmql":
 
         def get_one_answer(i):
@@ -102,7 +102,7 @@ def main(args):
             loop.run_until_complete(
                 asyncio.gather(*[get_one_answer_async(i) for i in batch])
             )
-    latency = time.time() - tic
+    latency = time.perf_counter() - tic
 
     # Compute accuracy
     print(f"Latency: {latency:.3f}")