fix: insert correct number of image tokens for multi-image in Cambrians model (#1075)

When a benchmark (e.g. MindCube) provides multiple images with a plain
string prompt, the Cambrians model wrapper only inserted a single
<image> token into the prompt regardless of the actual image count.
This caused the downstream cambrian library to take the single-image
code path while the image features tensor was shaped for multiple
images, resulting in a dimension mismatch RuntimeError.

Use len(visual_sizes) to insert one <image> token per image so the
cambrian library correctly routes to its multi-image concatenation
path.

Author: oscarqjh

lmms_eval/models/simple/cambrians.py

@@ -350,11 +350,12 @@ def __getitem__(self, idx):
                                     real_qs += DEFAULT_IMAGE_TOKEN + "\n"
                         qs = real_qs
                     elif isinstance(qs, str):
+                        num_images = len(visual_sizes)
                         if self.model_config.mm_use_im_start_end:
-                            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + "\n" + qs
+                            image_tokens = (DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + "\n") * num_images
+                            qs = image_tokens + qs
                         else:
-                            assert len(visual_tensors) == 1, "This should not happen."
-                            qs = DEFAULT_IMAGE_TOKEN * len(visual_tensors) + "\n" + qs
+                            qs = DEFAULT_IMAGE_TOKEN * num_images + "\n" + qs
                     else:
                         raise NotImplementedError