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, 07:00 UTC
Update default max_num_batch_tokens for chunked prefill to 2048 (#10544)
Michael Goin committed
Nov 23, 2024 at 00:14 UTC
02a43f82a97e37581b48f1c177d3393aca4fe3f2
1 file changed
+3
-3
vllm/config.py
+3
-3
@@ -1133,9 +1133,9 @@ class SchedulerConfig:
1133
# max_num_batched_tokens.
1134
self.max_num_batched_tokens = max(self.max_model_len, 2048)
1135
else:
1136
- # It is the values that have the best balance between ITL
1137
- # and TTFT on A100. Note it is not optimized for throughput.
1138
- self.max_num_batched_tokens = 512
1136
+ # This value is chosen to have a balance between ITL
1137
+ # and TTFT. Note it is not optimized for throughput.
1138
+ self.max_num_batched_tokens = 2048
1139
else:
1140
# If max_model_len is too short, use 2048 as the default value
1141
# for higher throughput.