Add validation for minimum embed_dim in init_resampler

Resampler is initialized with the hard-coded line `num_heads = embed_dim // 128`. When running with `config.hidden_size < 128`, it cryptically fails later on while giving an error message stating that `num_heads=0 while embed_dim={embed_dimension value}`

This change adds explicit validation with a helpful error message about the minimum embed_dim size so no more cryptic errors pop up and it is easier to debug.

Affects use cases:
- Model compression/distillation research
- Creating tiny test models for CI/CD

modeling_minicpmo.py

@@ -203,10 +203,18 @@ class MiniCPMO(MiniCPMOPreTrainedModel):
         return model
 
     def init_resampler(self, embed_dim, vision_dim):
+        MIN_EMBED_DIM = 128
+        if embed_dim < MIN_EMBED_DIM:
+            raise ValueError(
+                f"Resampler requires embed_dim >= {MIN_EMBED_DIM} "
+                f"(needed for num_heads = embed_dim // 128 >= 1). "
+                f"Got embed_dim={embed_dim}."
+            )
+    
         return Resampler(
             num_queries=self.config.query_num,
             embed_dim=embed_dim,
-            num_heads=embed_dim // 128,
+            num_heads=embed_dim // MIN_EMBED_DIM,
             kv_dim=vision_dim,
             adaptive=True,
         )