fix(mimo-vl): pass padded_context_dim to Qwen2_5_VisionPatchMerger (#29994)

2026-07-04 08:08:25 +08:00
parent 6ce02b95ad
commit b28bc1060f
1 changed files with 5 additions and 0 deletions
@@ -301,6 +301,11 @@ class MiMoVisionTransformer(nn.Module):
        self.merger = Qwen2_5_VisionPatchMerger(
            dim=vision_config.out_hidden_size,
            context_dim=hidden_size,
+            # MiMo-VL's merger MLP is square (intermediate == context_dim * merge**2),
+            # so no dim padding is needed. The Qwen2.5-VL formula num_heads * head_dim
+            # over-sizes it here because MiMo uses qk_channels (64) for head_dim rather
+            # than hidden_size // num_heads, which would mismatch the checkpoint.
+            padded_context_dim=hidden_size,
            spatial_merge_size=spatial_merge_size,
            quant_config=quant_config,
            prefix=add_prefix("merger", prefix),