Skip to content

add truncated output option [test] #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
kimbochen merged 1 commit into from
Jan 25, 2026
Merged

add truncated output option [test] #4

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
148 changes: 117 additions & 31 deletions benchmark_serving.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -371,10 +371,9 @@ def sample_random_requests(
range_ratio: float,
tokenizer: PreTrainedTokenizerBase,
use_chat_template: bool = False,
) -> List[Tuple[str, int, int]]:
prefix_token_ids = np.random.randint(0,
tokenizer.vocab_size,
size=prefix_len).tolist()
) -> List[Tuple[str, int, int]]:
prefix_token_ids = np.random.randint(0, tokenizer.vocab_size, size=prefix_len).tolist()

if use_chat_template:
chat_template_dummy = tokenizer.apply_chat_template(
[{"role": "user", "content": "a"}],
Expand All @@ -385,36 +384,69 @@ def sample_random_requests(
chat_template_len = len(tokenized_chat_template_dummy) - 1
input_len = input_len - chat_template_len

input_lens = np.random.randint(
int(input_len * range_ratio),
input_len + 1,
size=num_prompts,
)
output_lens = np.random.randint(
int(output_len * range_ratio),
output_len + 1,
size=num_prompts,
)
def sample_uniform(seq_len):
lower = int(seq_len * (1.0 - range_ratio))
upper = int(seq_len * (1.0 + range_ratio))
seq_lens = np.random.randint(lower, upper+1, size=num_prompts).tolist()
return seq_lens

input_lens = sample_uniform(input_len)
output_lens = sample_uniform(output_len)

# Make the first request have maximum possible length to ensure the benchmark
# fails early if max_model_len is set too low
max_input_len = int(input_len * (1.0 + range_ratio + 0.02))
max_output_len = int(output_len * (1.0 + range_ratio + 0.02))
input_lens[0] = max_input_len
output_lens[0] = max_output_len
offsets = np.random.randint(0, tokenizer.vocab_size, size=num_prompts)

input_requests = []
mismatches = []
for i in range(num_prompts):
prompt = tokenizer.decode(prefix_token_ids +
[(offsets[i] + i + j) % tokenizer.vocab_size
for j in range(input_lens[i])])
re_encoded_sequence = tokenizer.encode(prompt, add_special_tokens=False)[
:(prefix_len + input_lens[i])
]
prompt = tokenizer.decode(re_encoded_sequence)
tgt_prompt_len = prefix_len + input_lens[i]
prompt_token_ids = prefix_token_ids + [(offsets[i] + i + j) % tokenizer.vocab_size for j in range(input_lens[i])]
prompt = tokenizer.decode(prompt_token_ids)

max_retries = 10
for _ in range(max_retries):
prompt_token_ids = tokenizer.encode(prompt, add_special_tokens=False)
if len(prompt_token_ids) < tgt_prompt_len:
num_extras = tgt_prompt_len - len(prompt_token_ids)
prompt_token_ids.extend(np.random.randint(0, tokenizer.vocab_size, size=num_extras).tolist())
elif len(prompt_token_ids) > tgt_prompt_len:
prompt_token_ids = prompt_token_ids[:tgt_prompt_len]
else:
break
prompt = tokenizer.decode(prompt_token_ids)

if use_chat_template:
prompt = tokenizer.apply_chat_template(
[{"role": "user", "content": prompt}],
add_generation_prompt=True,
tokenize=False,
)
input_lens[i] += chat_template_len

input_requests.append((prompt, int(prefix_len + input_lens[i]),
int(output_lens[i]), None))
prompt_len = len(tokenizer.encode(prompt, add_special_tokens=False))
mismatches.append(prompt_len - tgt_prompt_len)
input_requests.append((prompt, prompt_len, output_lens[i], None))

header_str = f'{"-"*19} Input/Output Length Statistics {"-"*19}'
print(header_str)
print(
f' input_lens : '
f'min={min(r[1] for r in input_requests):<4d} '
f'max={max(r[1] for r in input_requests):<4d} '
f'mean={np.mean([r[1] for r in input_requests]):<7.2f} '
f'avg_token_mismatch={np.mean(mismatches):<5.2f} '
)
print(
f' output_lens: '
f'min={min(r[2] for r in input_requests):<4d} '
f'max={max(r[2] for r in input_requests):<4d} '
f'mean={np.mean([r[2] for r in input_requests]):<7.2f} '
)
print('-' * len(header_str), '\n')

return input_requests

Expand Down Expand Up @@ -584,6 +616,7 @@ async def benchmark(
request_rate: float,
burstiness: float,
disable_tqdm: bool,
num_warmups: int,
profile: bool,
selected_percentile_metrics: List[str],
selected_percentiles: List[str],
Expand All @@ -597,9 +630,9 @@ async def benchmark(
else:
raise ValueError(f"Unknown backend: {backend}")

print("Starting initial single prompt test run...")
test_prompt, test_prompt_len, test_output_len, test_mm_content = (
input_requests[0])
print(f"Starting initial single prompt test run (input length {test_prompt_len}, output length {test_output_len})...")
if backend != "openai-chat" and test_mm_content is not None:
# multi-modal benchmark is only available on OpenAI Chat backend.
raise ValueError(
Expand All @@ -616,14 +649,49 @@ async def benchmark(
multi_modal_content=test_mm_content,
ignore_eos=ignore_eos,
)

test_output = await request_func(request_func_input=test_input)
if not test_output.success:


# Take from vLLM ready_checker.py
# https://github.com/vllm-project/vllm/blob/e6ba2000aef3e61ca84bb114472badecbd533ee9/vllm/benchmarks/lib/ready_checker.py#L14
import aiohttp
timeout = 600
delay = 5
print(f"Waiting for endpoint to startup in {timeout} seconds")
for t in tqdm(range(timeout)):
if t % delay == 0:
try:
test_output = await request_func(request_func_input=test_input)
if test_output.success:
break
except aiohttp.ClientConnectorError as e:
pass
await asyncio.sleep(delay)
else:
test_output = await request_func(request_func_input=test_input)
raise ValueError(
"Initial test run failed - Please make sure benchmark arguments "
f"are correctly specified. Error: {test_output.error}")
else:
print("Initial test run completed. Starting main benchmark run...")
f"are correctly specified. Error: {test_output.error}"
)
print("Initial test run completed. Starting main benchmark run...")

if num_warmups > 0:
print(f"Warming up with {num_warmups} requests...")
warmup_pbar = None if disable_tqdm else tqdm(total=num_warmups)
warmup_semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else contextlib.nullcontext()

async def warmup_limited_req_fn():
async with warmup_semaphore:
return await request_func(request_func_input=test_input, pbar=warmup_pbar)

warmup_tasks = []
for _ in range(num_warmups):
task = asyncio.create_task(warmup_limited_req_fn())
warmup_tasks.append(task)
_ = await asyncio.gather(*warmup_tasks)

if warmup_pbar is not None:
warmup_pbar.close()
print("Warmup completed.")

if lora_modules:
# For each input request, choose a LoRA module at random.
Expand Down Expand Up @@ -931,6 +999,7 @@ def main(args: argparse.Namespace):
input_requests = [(prompt, prompt_len, output_len, None)
for prompt, prompt_formatted, prompt_len,
output_len, _ in input_requests]

else:
assert (
tokenizer.chat_template or tokenizer.default_chat_template
Expand Down Expand Up @@ -992,6 +1061,7 @@ def main(args: argparse.Namespace):
request_rate=args.request_rate,
burstiness=args.burstiness,
disable_tqdm=args.disable_tqdm,
num_warmups=args.num_warmups,
profile=args.profile,
selected_percentile_metrics=args.percentile_metrics.split(","),
selected_percentiles=[
Expand Down Expand Up @@ -1036,6 +1106,13 @@ def main(args: argparse.Namespace):
# Merge with benchmark result
result_json = {**result_json, **benchmark_result}

# Optionally exclude per-request data to reduce file size
if args.save_aggregated_only:
per_request_fields = ["input_lens", "output_lens", "ttfts", "itls",
"generated_texts", "errors"]
for field in per_request_fields:
result_json.pop(field, None)

# Save to file
base_model_id = model_id.split("/")[-1]
max_concurrency_str = (f"-concurrency{args.max_concurrency}"
Expand Down Expand Up @@ -1185,6 +1262,13 @@ def main(args: argparse.Namespace):
action="store_true",
help="Specify to save benchmark results to a json file",
)
parser.add_argument(
"--save-aggregated-only",
action="store_true",
help="When saving results, only include aggregated metrics and exclude "
"per-request data (input_lens, output_lens, ttfts, itls, generated_texts, "
"errors). This significantly reduces output file size for large benchmarks.",
)
parser.add_argument(
"--metadata",
metavar="KEY=VALUE",
Expand Down Expand Up @@ -1353,5 +1437,7 @@ def main(args: argparse.Namespace):
"launching the server. For each request, the "
"script chooses a LoRA module at random.")

parser.add_argument('--num-warmups', type=int, default=0)

args = parser.parse_args()
main(args)