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 16, 10:45 UTC
[Misc] Fix sampling from sonnet for long context case (#9235)
Andy Dai committed
Oct 10, 2024 at 17:33 UTC
94bf9ae4e9b8199636668ccbe4dabcdc3b9e5ae6
1 file changed
+4
-4
benchmarks/benchmark_serving.py
+4
-4
@@ -176,9 +176,9 @@ def sample_sonnet_requests(
176
# Sample the rest of lines per request.
177
sampled_requests: List[Tuple[str, int, int]] = []
178
for _ in range(num_requests):
179
- sampled_lines = "".join(
180
- prefix_lines +
181
- random.sample(poem_lines, num_input_lines - num_prefix_lines))
179
+ num_lines_needed = num_input_lines - num_prefix_lines
180
+ sampled_lines = "".join(prefix_lines +
181
+ random.choices(poem_lines, k=num_lines_needed))
182
183
prompt = f"{base_prompt}{sampled_lines}"
184
message = [
@@ -536,7 +536,7 @@ async def benchmark(
536
# E.g., "Time to First Token"
537
metric_header: str,
538
):
539
- # This function print and add statistics of the specified
539
+ # This function prints and adds statistics of the specified
540
# metric.
541
if metric_attribute_name not in selected_percentile_metrics:
542
return