diff --git a/vllm_omni/worker/cpu_diffusion_model_runner.py b/vllm_omni/worker/cpu_diffusion_model_runner.py
index 4d7f42fa1..43b684e31 100644
--- a/vllm_omni/worker/cpu_diffusion_model_runner.py
+++ b/vllm_omni/worker/cpu_diffusion_model_runner.py
@@ -37,7 +37,7 @@ logger = logging.getLogger(__name__)
 class CPUDiffusionModelRunner(OmniCPUModelRunner):
     """Diffusion model runner for vLLM-omni (non-autoregressive).
 
-    - Reuses GPUModelRunner preparation, multimodal handling, and TP/PP/DP glue.
+    - Reuses CPUModelRunner preparation, multimodal handling, and TP/PP/DP glue.
     - Does not compute logits or perform token sampling.
     - Executes diffusion process and returns tensors via `pooler_output`.
     """
diff --git a/vllm_omni/worker/cpu_model_runner.py b/vllm_omni/worker/cpu_model_runner.py
index 076a15f79..7d532cc36 100644
--- a/vllm_omni/worker/cpu_model_runner.py
+++ b/vllm_omni/worker/cpu_model_runner.py
@@ -40,9 +40,9 @@ class OmniCPUModelRunner(CPUModelRunner):
         output.
 
         The updated states are used by the `_prepare_inputs` function to create
-        the input GPU tensors for the model.
+        the input CPU tensors for the model.
 
-        The SamplingMetadata is updated and copied to the GPU if there is a
+        The SamplingMetadata is updated and copied to the CPU if there is a
         new/resumed/paused/finished request in the batch.
         """
         # Remove finished requests from the cached states.