updated

Author: anxiangsir

onevision_encoder/configuration_onevision_encoder.py

@@ -77,7 +77,7 @@ def __init__(
         attention_dropout=0.0,
         initializer_range=0.02,
         rope_theta=10000.0,
-        rope_temporal_size=None,
+        rope_temporal_size=64,
         use_head=True,
         **kwargs,
     ):

onevision_encoder/modeling_onevision_encoder.py

@@ -547,10 +547,7 @@ def forward(
         # Determine video dimensions for RoPE
         # Note: pixel_values passed to embeddings can be 4D or 5D
         if pixel_values.dim() == 5:
-             # fix: use config.rope_temporal_size if set, otherwise use actual frames
-             # legacy behavior was hardcoded t_frames=64 (for padded 64-frame videos)
-             actual_frames = pixel_values.shape[2]
-             t_frames = self.config.rope_temporal_size if self.config.rope_temporal_size else actual_frames
+             t_frames = self.config.rope_temporal_size
              height = pixel_values.shape[3]
              width = pixel_values.shape[4]
         else: