diff --git a/vllm/config/vllm.py b/vllm/config/vllm.py
index ba7d26c93b..9e59c60a94 100644
--- a/vllm/config/vllm.py
+++ b/vllm/config/vllm.py
@@ -77,6 +77,12 @@ DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES = frozenset(
 )
 
 
+# Upper bound for the token-keyed CUDA graph capture range chosen in
+# _set_cudagraph_sizes. Capture cost grows with the number and size of captured
+# graphs, so the token budget is only honoured up to this many tokens.
+_TOKEN_KEYED_CUDAGRAPH_CAPTURE_LIMIT = 1024
+
+
 class OptimizationLevel(IntEnum):
     """Optimization level enum."""
 
@@ -1690,6 +1696,32 @@ class VllmConfig:
                 max_cudagraph_capture_size = min(
                     self.scheduler_config.max_num_seqs * decode_query_len * 2, 512
                 )
+                # The bound above is derived from the sequence count, which is
+                # the right bound for a step that is purely decode: one query
+                # position per running sequence. Under chunked prefill a step
+                # also carries prompt tokens, so its batch size is bounded by
+                # the token budget instead, and the two can differ by an order
+                # of magnitude -- max_num_seqs 32 with max_num_batched_tokens
+                # 8192 stops capture at 64 while real mixed prefill/decode
+                # steps run into the hundreds of tokens. Every such step then
+                # exceeds the largest captured graph and falls back to the
+                # eager path (see the note on this method: "If batch size >
+                # largest cudagraph_capture_sizes, cudagraph will not be
+                # used"), paying full Python dispatch on every layer.
+                #
+                # Take the token budget into account as well, bounded so that
+                # capture stays cheap in time and memory. An explicit
+                # max_cudagraph_capture_size still wins: this only fills in the
+                # default, and max_num_seqs is not read differently anywhere
+                # else -- 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,
+                        _TOKEN_KEYED_CUDAGRAPH_CAPTURE_LIMIT,
+                    ),
+                )
             max_num_tokens = self.scheduler_config.max_num_batched_tokens
             max_cudagraph_capture_size = min(max_num_tokens, max_cudagraph_capture_size)