Fix CUDA index out of bounds for q_idx in VLM token type masking for Gemma3, PaliGemma, and example modular (huggingface#41757)

* Fix CUDA index out of bounds for q_idx in Gemma3 token type masking

* Fix CUDA index out of bounds for q_idx in modular modeling_new_task_model

* Revert "Fix CUDA index out of bounds for q_idx in Gemma3 token type masking"

This reverts commit f8e5c2a.

* Fix CUDA index out of bounds for q_idx in PaliGemma token type masking

* Fix CUDA index out of bounds for q_idx in Gemma3 token type masking

Author: albertvillanova

examples/modular-transformers/modeling_new_task_model.py

@@ -125,15 +125,23 @@ def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
         # If it's 1 for both query and key/value, we are in an image block
         # NOTE: static cache shape goes beyond input seq length, while token_type_ids.shape[1] == input seq length
         # Since vmap doesn't support `if statement` we workaround it with `torch.where`
-        safe_idx = torch.where(kv_idx < token_type_ids.shape[1], kv_idx, 0)
-        token_type_ids_at_kv_idx = token_type_ids[batch_idx, safe_idx]
+        safe_q_idx = torch.where(q_idx < token_type_ids.shape[1], q_idx, 0)
+        safe_kv_idx = torch.where(kv_idx < token_type_ids.shape[1], kv_idx, 0)
+
+        token_type_ids_at_q_idx = token_type_ids[batch_idx, safe_q_idx]
+        token_type_ids_at_q_idx = torch.where(q_idx < token_type_ids.shape[1], token_type_ids_at_q_idx, 0)
+
+        token_type_ids_at_kv_idx = token_type_ids[batch_idx, safe_kv_idx]
         token_type_ids_at_kv_idx = torch.where(kv_idx < token_type_ids.shape[1], token_type_ids_at_kv_idx, 0)
 
-        image_group_ids_at_kv_idx = image_group_ids[batch_idx, safe_idx]
+        image_group_ids_at_q_idx = image_group_ids[batch_idx, safe_q_idx]
+        image_group_ids_at_q_idx = torch.where(q_idx < image_group_ids.shape[1], image_group_ids_at_q_idx, -1)
+
+        image_group_ids_at_kv_idx = image_group_ids[batch_idx, safe_kv_idx]
         image_group_ids_at_kv_idx = torch.where(kv_idx < image_group_ids.shape[1], image_group_ids_at_kv_idx, -1)
 
-        is_image_block = (token_type_ids[batch_idx, q_idx] == 1) & (token_type_ids_at_kv_idx == 1)
-        same_image_block = image_group_ids[batch_idx, q_idx] == image_group_ids_at_kv_idx
+        is_image_block = (token_type_ids_at_q_idx == 1) & (token_type_ids_at_kv_idx == 1)
+        same_image_block = image_group_ids_at_q_idx == image_group_ids_at_kv_idx
 
         # This is bidirectional attention whenever we are dealing with image tokens
         return is_image_block & same_image_block

src/transformers/models/gemma3/modeling_gemma3.py

@@ -768,15 +768,23 @@ def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
         # If it's 1 for both query and key/value, we are in an image block
         # NOTE: static cache shape goes beyond input seq length, while token_type_ids.shape[1] == input seq length
         # Since vmap doesn't support `if statement` we workaround it with `torch.where`
-        safe_idx = torch.where(kv_idx < token_type_ids.shape[1], kv_idx, 0)
-        token_type_ids_at_kv_idx = token_type_ids[batch_idx, safe_idx]
+        safe_q_idx = torch.where(q_idx < token_type_ids.shape[1], q_idx, 0)
+        safe_kv_idx = torch.where(kv_idx < token_type_ids.shape[1], kv_idx, 0)
+
+        token_type_ids_at_q_idx = token_type_ids[batch_idx, safe_q_idx]
+        token_type_ids_at_q_idx = torch.where(q_idx < token_type_ids.shape[1], token_type_ids_at_q_idx, 0)
+
+        token_type_ids_at_kv_idx = token_type_ids[batch_idx, safe_kv_idx]
         token_type_ids_at_kv_idx = torch.where(kv_idx < token_type_ids.shape[1], token_type_ids_at_kv_idx, 0)
 
-        image_group_ids_at_kv_idx = image_group_ids[batch_idx, safe_idx]
+        image_group_ids_at_q_idx = image_group_ids[batch_idx, safe_q_idx]
+        image_group_ids_at_q_idx = torch.where(q_idx < image_group_ids.shape[1], image_group_ids_at_q_idx, -1)
+
+        image_group_ids_at_kv_idx = image_group_ids[batch_idx, safe_kv_idx]
         image_group_ids_at_kv_idx = torch.where(kv_idx < image_group_ids.shape[1], image_group_ids_at_kv_idx, -1)
 
-        is_image_block = (token_type_ids[batch_idx, q_idx] == 1) & (token_type_ids_at_kv_idx == 1)
-        same_image_block = image_group_ids[batch_idx, q_idx] == image_group_ids_at_kv_idx
+        is_image_block = (token_type_ids_at_q_idx == 1) & (token_type_ids_at_kv_idx == 1)
+        same_image_block = image_group_ids_at_q_idx == image_group_ids_at_kv_idx
 
         # This is bidirectional attention whenever we are dealing with image tokens
         return is_image_block & same_image_block

src/transformers/models/paligemma/modeling_paligemma.py

@@ -116,15 +116,23 @@ def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
         # If it's 1 for both query and key/value, we are in an image block
         # NOTE: static cache shape goes beyond input seq length, while token_type_ids.shape[1] == input seq length
         # Since vmap doesn't support `if statement` we workaround it with `torch.where`
-        safe_idx = torch.where(kv_idx < token_type_ids.shape[1], kv_idx, 0)
-        token_type_ids_at_kv_idx = token_type_ids[batch_idx, safe_idx]
+        safe_q_idx = torch.where(q_idx < token_type_ids.shape[1], q_idx, 0)
+        safe_kv_idx = torch.where(kv_idx < token_type_ids.shape[1], kv_idx, 0)
+
+        token_type_ids_at_q_idx = token_type_ids[batch_idx, safe_q_idx]
+        token_type_ids_at_q_idx = torch.where(q_idx < token_type_ids.shape[1], token_type_ids_at_q_idx, 0)
+
+        token_type_ids_at_kv_idx = token_type_ids[batch_idx, safe_kv_idx]
         token_type_ids_at_kv_idx = torch.where(kv_idx < token_type_ids.shape[1], token_type_ids_at_kv_idx, 0)
 
-        image_group_ids_at_kv_idx = image_group_ids[batch_idx, safe_idx]
+        image_group_ids_at_q_idx = image_group_ids[batch_idx, safe_q_idx]
+        image_group_ids_at_q_idx = torch.where(q_idx < image_group_ids.shape[1], image_group_ids_at_q_idx, -1)
+
+        image_group_ids_at_kv_idx = image_group_ids[batch_idx, safe_kv_idx]
         image_group_ids_at_kv_idx = torch.where(kv_idx < image_group_ids.shape[1], image_group_ids_at_kv_idx, -1)
 
-        is_image_block = (token_type_ids[batch_idx, q_idx] == 1) & (token_type_ids_at_kv_idx == 1)
-        same_image_block = image_group_ids[batch_idx, q_idx] == image_group_ids_at_kv_idx
+        is_image_block = (token_type_ids_at_q_idx == 1) & (token_type_ids_at_kv_idx == 1)
+        same_image_block = image_group_ids_at_q_idx == image_group_ids_at_kv_idx
 
         # This is bidirectional attention whenever we are dealing with image tokens
         return is_image_block & same_image_block