diff --git a/vllm/config/vllm.py b/vllm/config/vllm.py
index ba7d26c..c74e940 100644
--- a/vllm/config/vllm.py
+++ b/vllm/config/vllm.py
@@ -1690,6 +1690,24 @@ class VllmConfig:
                 max_cudagraph_capture_size = min(
                     self.scheduler_config.max_num_seqs * decode_query_len * 2, 512
                 )
+                # The bound above counts one query position per running
+                # sequence, which is the batch size of a pure-decode step.
+                # With chunked prefill a step also carries prompt tokens, so
+                # its width is set by the token budget rather than the
+                # sequence count. When the two disagree -- max_num_seqs 32
+                # against max_num_batched_tokens 8192 -- every mixed
+                # prefill/decode step is wider than the largest captured
+                # graph and takes the eager path instead, paying Python
+                # dispatch per layer on the steps that dominate TTFT.
+                # Widen the default to cover those steps, capped so capture
+                # stays bounded in time and memory. An explicit
+                # max_cudagraph_capture_size is untouched: this only fills in
+                # the default, and the scheduler still admits exactly
+                # max_num_seqs sequences.
+                max_cudagraph_capture_size = max(
+                    max_cudagraph_capture_size,
+                    min(self.scheduler_config.max_num_batched_tokens, 768),
+                )
             max_num_tokens = self.scheduler_config.max_num_batched_tokens
             max_cudagraph_capture_size = min(max_num_tokens, max_cudagraph_capture_size)