fix(qwen 35): fix qwen35 thd attention mask dtype to bool

src/megatron/bridge/models/qwen_vl/modelling_qwen3_vl/model.py

@@ -400,7 +400,7 @@ def forward(
                 combined_embeddings = split_data_cp_rank(combined_embeddings, cp_size, 0, cp_rank)
             if packed_seq_params is not None:
                 if attention_mask is None:
-                    attention_mask = torch.ones_like(input_ids, dtype=torch.int32, device=input_ids.device)
+                    attention_mask = torch.ones_like(input_ids, dtype=torch.bool, device=input_ids.device)
                 input_ids_thd, _ = preprocess_packed_seqs(
                     input_ids, attention_mask, pre_process=True, pg_collection=self.pg_collection
                 )
@@ -456,7 +456,7 @@ def forward(
             # convert lm_input_ids to THD format so it matches position_ids.
             if packed_seq_params is not None:
                 if attention_mask is None:
-                    attention_mask = torch.ones_like(input_ids, dtype=torch.int32, device=input_ids.device)
+                    attention_mask = torch.ones_like(input_ids, dtype=torch.bool, device=input_ids.device)
                 lm_input_ids, _ = preprocess_packed_seqs(
                     input_ids, attention_mask, pre_process=True, pg_collection=self.pg_collection
                 )

src/megatron/bridge/models/qwen_vl/modelling_qwen3_vl/utils.py

@@ -668,6 +668,11 @@ def preprocess_packed_seqs(
     """
     batch_size = input_ids.shape[0]
 
+    # Ensure boolean dtype for correct advanced indexing (bool → mask select,
+    # int → fancy index which silently corrupts data when values are 0/1).
+    if attention_mask.dtype != torch.bool:
+        attention_mask = attention_mask.bool()
+
     seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
     if pg_collection is not None:
         tp_size = pg_collection.tp.size()