diff --git a/vllm/config/vllm.py b/vllm/config/vllm.py
index ba7d26c93..5421594e3 100644
--- a/vllm/config/vllm.py
+++ b/vllm/config/vllm.py
@@ -1690,6 +1690,32 @@ class VllmConfig:
                 max_cudagraph_capture_size = min(
                     self.scheduler_config.max_num_seqs * decode_query_len * 2, 512
                 )
+                # The bound above sizes a step that is purely decode: one
+                # query position per running sequence, times the speculative
+                # width. Under chunked prefill a step also carries prompt
+                # tokens, so its size is bounded by the token budget instead,
+                # and the two differ by more than an order of magnitude --
+                # max_num_seqs 32 with max_num_batched_tokens 8192 stops
+                # capture in the low hundreds while a mixed step runs to
+                # thousands of tokens. Per this method's own contract, "if
+                # batch size > largest cudagraph_capture_sizes, cudagraph will
+                # not be used", so every mixed step then pays full Python
+                # dispatch per layer.
+                #
+                # Raise the default toward the token budget. The ceiling is
+                # what keeps capture affordable: graphs cost startup time,
+                # which is not on the measured path, but they also cost device
+                # memory, which competes with the KV cache and is the one
+                # resource that can fail a start outright. An explicit
+                # max_cudagraph_capture_size still wins outright; this only
+                # fills in the default.
+                max_cudagraph_capture_size = max(
+                    max_cudagraph_capture_size,
+                    min(
+                        self.scheduler_config.max_num_batched_tokens,
+                        _cudagraph_capture_ceiling(),
+                    ),
+                )
             max_num_tokens = self.scheduler_config.max_num_batched_tokens
             max_cudagraph_capture_size = min(max_num_tokens, max_cudagraph_capture_size)
 
@@ -2292,3 +2318,38 @@ def get_layers_from_vllm_config(
         for layer_name in layer_names
         if isinstance(layer := forward_context.get(layer_name), layer_type)
     }
+
+
+# Ceiling on the token-keyed CUDA graph capture range chosen in
+# _set_cudagraph_sizes when the config pins no explicit size. Capture cost
+# grows with the number and the size of the captured graphs, so the token
+# budget is only honoured this far. Overridable so a sweep needs one build
+# rather than one build per ceiling.
+_CUDAGRAPH_CAPTURE_CEILING_ENV = "VLLM_MIXED_STEP_CAPTURE_CEILING"
+_CUDAGRAPH_CAPTURE_CEILING_DEFAULT = 1536
+
+
+def _cudagraph_capture_ceiling() -> int:
+    """Capture ceiling, from the environment when set and usable."""
+    raw = os.environ.get(_CUDAGRAPH_CAPTURE_CEILING_ENV)
+    if raw is None:
+        return _CUDAGRAPH_CAPTURE_CEILING_DEFAULT
+    try:
+        ceiling = int(raw)
+    except ValueError:
+        logger.warning(
+            "%s=%r is not an integer; using %d",
+            _CUDAGRAPH_CAPTURE_CEILING_ENV,
+            raw,
+            _CUDAGRAPH_CAPTURE_CEILING_DEFAULT,
+        )
+        return _CUDAGRAPH_CAPTURE_CEILING_DEFAULT
+    if ceiling < 1:
+        logger.warning(
+            "%s=%d is not positive; using %d",
+            _CUDAGRAPH_CAPTURE_CEILING_ENV,
+            ceiling,
+            _CUDAGRAPH_CAPTURE_CEILING_DEFAULT,
+        )
+        return _CUDAGRAPH_CAPTURE_CEILING_DEFAULT
+    return ceiling