Merge branch 'main' into transformers_future

Author: regisss

optimum/habana/transformers/models/mllama/modeling_mllama.py

@@ -1151,8 +1151,11 @@ def _update_model_kwargs_for_generation(self, outputs, model_kwargs, is_encoder_
         # add cross-attn mask for new token
         if cross_attention_mask_prev is not None:
             token_idx = model_kwargs.get("token_idx", None)
+            token_idx_cpu = model_kwargs.get(
+                "token_idx_cpu", None
+            )  # returns an integer so following slicing ops happen using int instead of tensor
             if token_idx is not None:
-                mask = cross_attention_mask_prev[:, token_idx - 2 : token_idx - 1, ...]
+                mask = cross_attention_mask_prev[:, token_idx_cpu - 2 : token_idx_cpu - 1, ...]
                 cross_attention_mask_prev.index_copy_(1, token_idx - 1, mask)
                 model_kwargs["cross_attention_mask"] = cross_attention_mask_prev
             else:

tests/baselines/fixture/tests/test_examples.json

@@ -655,6 +655,20 @@
       "train_samples_per_second": 2845.068
     }
   },
+  "tests/test_examples.py::MultiCardVisionLanguageExampleTester::test_run_clip_clip-roberta_multi_card": {
+    "gaudi1": {
+      "train_runtime": 314.7726,
+      "train_samples_per_second": 2560.999
+    },
+    "gaudi2": {
+      "train_runtime": 59.5,
+      "train_samples_per_second": 14124
+    },
+    "gaudi3": {
+      "train_runtime": 59.5,
+      "train_samples_per_second": 14124
+    }
+  },
   "tests/test_examples.py::QuestionAnsweringExampleTester::test_run_qa_roberta-large_single_card": {
     "gaudi1": {
       "eval_f1": 94.2959,