diff --git a/relax/engine/sft/dataset/multimodal.py b/relax/engine/sft/dataset/multimodal.py
index 7a91e28d8..46b992e48 100644
--- a/relax/engine/sft/dataset/multimodal.py
+++ b/relax/engine/sft/dataset/multimodal.py
@@ -37,7 +37,7 @@ def _fetch_media(sample: CanonicalSample, rendered_text: str) -> tuple[dict[str,
     if sample.images:
         mm_inputs["images"] = [load_image(p) for p in sample.images]
     if sample.videos:
-        mm_inputs["videos"] = [load_video(p) for p in sample.videos]
+        mm_inputs["videos"] = [load_video(p, use_audio_in_video=False)[0] for p in sample.videos]
     if sample.audios:
         mm_inputs["audios"] = [load_audio(p) for p in sample.audios]
     return mm_inputs, rendered_text
diff --git a/relax/engine/sft/predict/loop.py b/relax/engine/sft/predict/loop.py
index d481f1ced..e4f2f9746 100644
--- a/relax/engine/sft/predict/loop.py
+++ b/relax/engine/sft/predict/loop.py
@@ -47,7 +47,7 @@ def _build_multimodal_inputs(sample: CanonicalSample) -> dict | None:
 
     return {
         "images": [load_image(p) for p in (sample.images or [])],
-        "videos": [load_video(p) for p in (sample.videos or [])],
+        "videos": [load_video(p, use_audio_in_video=False)[0] for p in (sample.videos or [])],
         "audio": [load_audio(p) for p in (sample.audios or [])],
     }