diff --git a/vllm/patches/deepseek_v4.py b/vllm/patches/deepseek_v4.py
index 86de2081..2c1b5afc 100644
--- a/vllm/patches/deepseek_v4.py
+++ b/vllm/patches/deepseek_v4.py
@@ -508,13 +508,13 @@ class DeepseekV4MegaMoEExperts(nn.Module):
         # Warm up the CuTeDSL kernel (JIT compiles MLIR→PTX on first call).
         # This takes ~1-2 min but prevents the vLLM RPC timeout (5 min) from
         # killing the engine when the first inference request triggers compilation.
-        print("  Warming up CuTeDSL NVFP4 kernel (first call compiles)...", flush=True)
+        print("  Compiling CuTeDSL NVFP4 MegaMoE kernel (one-time JIT, ~1-2 min)...", flush=True)
         dummy_hidden = torch.randn(1, self.hidden_size, dtype=torch.bfloat16, device=self._cutedsl_runner.l1_fp4[0].device)
         dummy_ids = torch.zeros(1, 1, dtype=torch.int32, device=dummy_hidden.device)
         dummy_weights = torch.ones(1, 1, dtype=torch.float32, device=dummy_hidden.device)
         try:
             self._cutedsl_runner.run(dummy_hidden, dummy_weights, dummy_ids, expert_indices=[0])
-            print("  CuTeDSL warmup complete ✓", flush=True)
+            print("  CuTeDSL NVFP4 MegaMoE kernel compiled ✓", flush=True)
         except Exception:
             print("  CuTeDSL warmup failed (will compile on first inference)", flush=True)