diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index f3279fa5fa8ce..44de1469d1b10 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -2270,8 +2270,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             # Only rank 0 should print progress bar during capture
             compilation_cases = reversed(self.cudagraph_batch_sizes)
             if is_global_first_rank():
-                compilation_cases = tqdm(list(compilation_cases),
-                                         desc="Capturing CUDA graph shapes")
+                compilation_cases = tqdm(
+                    list(compilation_cases),
+                    disable=not self.load_config.use_tqdm_on_load,
+                    desc="Capturing CUDA graph shapes")
             for num_tokens in compilation_cases:
                 # We skip EPLB here since we don't want to record dummy metrics
                 for _ in range(
diff --git a/vllm/worker/model_runner.py b/vllm/worker/model_runner.py
index 9d936f3dbf01c..4fe70a0abf808 100644
--- a/vllm/worker/model_runner.py
+++ b/vllm/worker/model_runner.py
@@ -1587,6 +1587,7 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
                 if get_tensor_model_parallel_rank() == 0:
                     compilation_cases = tqdm(
                         list(compilation_cases),
+                        disable=not self.load_config.use_tqdm_on_load,
                         desc="Capturing CUDA graph shapes")
                 for batch_size, use_inputs_embeds in compilation_cases:
                     attn_metadata = (