A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 23:45 UTC
[bugfix] fix profile impact benchmark results (#21507)
Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>
rongfu.leng committed
Jul 28, 2025 at 13:44 UTC
18cc33dd603cce058867f5730919041fa3617eb7
3 files changed
+41
-42
benchmarks/benchmark_serving.py
+14
-14
@@ -396,20 +396,6 @@ async def benchmark(
396
tasks.append(asyncio.create_task(task))
397
outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
398
399
- if profile:
400
- print("Stopping profiler...")
401
- profile_input = RequestFuncInput(
402
- model=model_id,
403
- prompt=test_prompt,
404
- api_url=base_url + "/stop_profile",
405
- prompt_len=test_prompt_len,
406
- output_len=test_output_len,
407
- logprobs=logprobs,
408
- )
409
- profile_output = await request_func(request_func_input=profile_input)
410
- if profile_output.success:
411
- print("Profiler stopped")
412
-
399
if pbar is not None:
400
pbar.close()
401
@@ -518,6 +504,20 @@ async def benchmark(
504
505
print("=" * 50)
506
507
+ if profile:
508
+ print("Stopping profiler...")
509
+ profile_input = RequestFuncInput(
510
+ model=model_id,
511
+ prompt=test_prompt,
512
+ api_url=base_url + "/stop_profile",
513
+ prompt_len=test_prompt_len,
514
+ output_len=test_output_len,
515
+ logprobs=logprobs,
516
+ )
517
+ profile_output = await request_func(request_func_input=profile_input)
518
+ if profile_output.success:
519
+ print("Profiler stopped")
520
+
521
return result
522
523
benchmarks/benchmark_serving_structured_output.py
+14
-14
@@ -538,20 +538,6 @@ async def benchmark(
538
)
539
outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
540
541
- if profile:
542
- print("Stopping profiler...")
543
- profile_input = RequestFuncInput(
544
- model=model_id,
545
- prompt=test_request.prompt,
546
- api_url=base_url + "/stop_profile",
547
- prompt_len=test_request.prompt_len,
548
- output_len=test_request.expected_output_len,
549
- extra_body={test_request.structure_type: test_request.schema},
550
- )
551
- profile_output = await request_func(request_func_input=profile_input)
552
- if profile_output.success:
553
- print("Profiler stopped")
554
-
541
if pbar is not None:
542
pbar.close()
543
@@ -666,6 +652,20 @@ async def benchmark(
652
653
print("=" * 50)
654
655
+ if profile:
656
+ print("Stopping profiler...")
657
+ profile_input = RequestFuncInput(
658
+ model=model_id,
659
+ prompt=test_request.prompt,
660
+ api_url=base_url + "/stop_profile",
661
+ prompt_len=test_request.prompt_len,
662
+ output_len=test_request.expected_output_len,
663
+ extra_body={test_request.structure_type: test_request.schema},
664
+ )
665
+ profile_output = await request_func(request_func_input=profile_input)
666
+ if profile_output.success:
667
+ print("Profiler stopped")
668
+
669
return result, ret
670
671
vllm/benchmarks/serve.py
+13
-14
@@ -470,20 +470,6 @@ async def benchmark(
470
pbar=pbar)))
471
outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
472
473
- if profile:
474
- print("Stopping profiler...")
475
- profile_input = RequestFuncInput(
476
- model=model_id,
477
- prompt=test_prompt,
478
- api_url=base_url + "/stop_profile",
479
- prompt_len=test_prompt_len,
480
- output_len=test_output_len,
481
- logprobs=logprobs,
482
- )
483
- profile_output = await request_func(request_func_input=profile_input)
484
- if profile_output.success:
485
- print("Profiler stopped")
486
-
473
if pbar is not None:
474
pbar.close()
475
@@ -576,6 +562,19 @@ async def benchmark(
562
563
print("=" * 50)
564
565
+ if profile:
566
+ print("Stopping profiler...")
567
+ profile_input = RequestFuncInput(
568
+ model=model_id,
569
+ prompt=test_prompt,
570
+ api_url=base_url + "/stop_profile",
571
+ prompt_len=test_prompt_len,
572
+ output_len=test_output_len,
573
+ logprobs=logprobs,
574
+ )
575
+ profile_output = await request_func(request_func_input=profile_input)
576
+ if profile_output.success:
577
+ print("Profiler stopped")
578
return result
579
580