diff --git a/vllm/v1/worker/gpu/cudagraph_utils.py b/vllm/v1/worker/gpu/cudagraph_utils.py
index 13ff66fac8d9..53580ecaa84b 100644
--- a/vllm/v1/worker/gpu/cudagraph_utils.py
+++ b/vllm/v1/worker/gpu/cudagraph_utils.py
@@ -528,12 +528,15 @@ def dispatch(
         num_active_loras: int,
         max_query_len: int | None = None,
         num_ubatches: int = 1,
+        allow_full: bool = True,
     ) -> BatchExecutionDescriptor:
         """Find matching cudagraph descriptor from priority-ordered candidates."""
         effective_loras = self._resolve_effective_loras(num_active_loras)
         key = (num_tokens, effective_loras)
         if self._graphs_captured and num_tokens > 0 and key in self._candidates:
             for desc in self._candidates[key]:
+                if not allow_full and desc.cg_mode == CUDAGraphMode.FULL:
+                    continue
                 if _is_compatible(
                     desc,
                     num_reqs,
diff --git a/vllm/v1/worker/gpu/dp_utils.py b/vllm/v1/worker/gpu/dp_utils.py
index fb9f4c02d3e4..b0d2bc81c68e 100644
--- a/vllm/v1/worker/gpu/dp_utils.py
+++ b/vllm/v1/worker/gpu/dp_utils.py
@@ -223,6 +223,7 @@ def dispatch_cg_and_sync_dp(
     uniform_decode: bool = False,
     allow_skip_drafts: bool = False,
     dp_sync_state: DPSyncState | None = None,
+    allow_full: bool = True,
 ) -> tuple[BatchExecutionDescriptor, DPSyncState | None]:
     """Pick a cudagraph descriptor for this batch, agreeing it across DP ranks.
 
@@ -287,6 +288,7 @@ def dispatch_cg_and_sync_dp(
             dp_sync_state.uniform_token_count if dp_sync_state else uniform_token_count,
             num_active_loras=num_active_loras,
             max_query_len=max_query_len,
+            allow_full=allow_full,
         )
 
     if dp_size == 1:
diff --git a/vllm/v1/worker/gpu/model_runner.py b/vllm/v1/worker/gpu/model_runner.py
index 31b2b6ad5132..487a2f70d31c 100644
--- a/vllm/v1/worker/gpu/model_runner.py
+++ b/vllm/v1/worker/gpu/model_runner.py
@@ -1774,12 +1774,14 @@ def execute_model(
         num_reqs = len(scheduler_output.num_scheduled_tokens)
         num_toks = scheduler_output.total_num_scheduled_tokens
         max_query_len: int | None = max(scheduler_output.num_scheduled_tokens.values())
+        allow_full = True
         batch_req_state, uniform_tok_count = self.gather_batch_req_state(
             scheduler_output, dummy_run
         )
         if batch_req_state is not None:
             num_toks = batch_req_state.num_tokens
             if batch_req_state.has_prefill:
+                allow_full = False
                 # Varlen decode graphs replay decodes only, and their bound
                 # alone would admit a short prefill.
                 max_query_len = None
@@ -1818,6 +1820,7 @@ def execute_model(
                 self.ubatch_runner is not None and not skip_attn_for_dummy_run
             ),
             uniform_decode=uniform_tok_count == self.decode_query_len,
+            allow_full=allow_full,
             allow_skip_drafts=scheduler_output.skip_drafts,
         )