Update default max_num_batch_tokens for chunked prefill to 2048 (#10544)

2024-11-23 00:14:19 -05:00
parent cfea9c04ef
commit 02a43f82a9
1 changed files with 3 additions and 3 deletions
--- a/vllm/config.py
+++ b/vllm/config.py
@@ -1133,9 +1133,9 @@ class SchedulerConfig:
                    # max_num_batched_tokens.
                    self.max_num_batched_tokens = max(self.max_model_len, 2048)
                else:
-                    # It is the values that have the best balance between ITL
-                    # and TTFT on A100. Note it is not optimized for throughput.
-                    self.max_num_batched_tokens = 512
+                    # This value is chosen to have a balance between ITL
+                    # and TTFT. Note it is not optimized for throughput.
+                    self.max_num_batched_tokens = 2048
            else:
                # If max_model_len is too short, use 2048 as the default value
                # for higher throughput.