[CI] Fixing some AMD failures (#42879)

* fix qwen2 & qwen2_5_omni

* one more fix

* fix qwen2_5_vl

* fix

* fix some more failures on nvidia

Author: Abdennacer-Badaoui

src/transformers/models/qwen2_5_omni/modeling_qwen2_5_omni.py

@@ -2459,7 +2459,11 @@ def forward(
                 self.rope_deltas = rope_deltas
 
             else:
-                batch_size, seq_length, _ = inputs_embeds.shape
+                if inputs_embeds is not None:
+                    batch_size, seq_length, _ = inputs_embeds.shape
+                else:
+                    batch_size, seq_length = input_ids.shape
+
                 delta = (past_key_values_length + self.rope_deltas).to(input_ids.device)
                 position_ids = torch.arange(seq_length, device=input_ids.device)
                 position_ids = position_ids.view(1, -1).expand(batch_size, -1)

src/transformers/models/qwen2_5_omni/modular_qwen2_5_omni.py

@@ -2615,7 +2615,11 @@ def forward(
                 self.rope_deltas = rope_deltas
 
             else:
-                batch_size, seq_length, _ = inputs_embeds.shape
+                if inputs_embeds is not None:
+                    batch_size, seq_length, _ = inputs_embeds.shape
+                else:
+                    batch_size, seq_length = input_ids.shape
+
                 delta = (past_key_values_length + self.rope_deltas).to(input_ids.device)
                 position_ids = torch.arange(seq_length, device=input_ids.device)
                 position_ids = position_ids.view(1, -1).expand(batch_size, -1)

tests/models/qwen2/test_modeling_qwen2.py

@@ -78,11 +78,10 @@ def test_model_450m_logits(self):
         with torch.no_grad():
             out = model(input_ids).logits.float().cpu()
         # Expected mean on dim = -1
-        EXPECTED_MEAN = torch.tensor([[-1.9537, -1.6193, -1.4123, -1.4673, -1.8511, -1.9309, -1.9826, -2.1776]])
+        EXPECTED_MEAN = torch.tensor([[-2.2121, -1.6335, -1.4816, -1.5035, -1.9110, -1.8979, -1.9682, -2.1980]])
         torch.testing.assert_close(out.mean(-1), EXPECTED_MEAN, rtol=1e-2, atol=1e-2)
         # slicing logits[0, 0, 0:30]
-        EXPECTED_SLICE = torch.tensor([3.2025, 7.1265, 4.6058, 3.6423, 1.6357, 3.9265, 5.1883, 5.8760, 2.7942, 4.4823, 3.2571, 2.1063, 3.4275, 4.2028, 1.9767, 5.2115, 6.6756, 6.3999, 6.0483, 5.7378, 5.6660, 5.2298, 5.4103, 5.1248, 5.4376, 2.4570, 2.6107, 5.4039, 2.8077, 4.7777])  # fmt: skip
-        print(out[0, 0, :30])
+        EXPECTED_SLICE = torch.tensor([2.7344, 4.2812, 4.1562, 2.3906, 1.1875, 2.1562, 3.1719, 3.1406, 1.2891, 3.6094, 3.3125, 1.8203, 2.9219, 3.2344, 1.5938, 6.2500, 7.4062, 7.2188, 6.5938, 6.0312, 6.1562, 5.3750, 5.9688, 5.5938, 6.1250, 1.2656, 1.6016, 3.4062, 1.7891, 3.6406])  # fmt: skip
         torch.testing.assert_close(out[0, 0, :30], EXPECTED_SLICE, rtol=1e-4, atol=1e-4)
 
         del model
@@ -92,7 +91,7 @@ def test_model_450m_logits(self):
     @slow
     def test_model_450m_generation(self):
         EXPECTED_TEXT_COMPLETION = (
-            """My favourite condiment is 100% natural, organic and vegan. I love to use it in my cooking and I"""
+            """My favourite condiment is 100% natural, organic and vegan. I love to use it in my cooking, but"""
         )
         prompt = "My favourite condiment is "
         tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-0.5B", use_fast=False)
@@ -161,7 +160,7 @@ def test_model_450m_long_prompt_sdpa(self):
         gc.collect()
 
         EXPECTED_TEXT_COMPLETION = (
-            "My favourite condiment is 100% natural, organic and vegan. I love to use it in my cooking and I"
+            "My favourite condiment is 100% natural, organic and vegan. I love to use it in my cooking, but"
         )
         prompt = "My favourite condiment is "
         tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-0.5B", use_fast=False)
@@ -211,11 +210,8 @@ def test_export_static_cache(self):
         tokenizer = AutoTokenizer.from_pretrained(qwen_model, pad_token="</s>", padding_side="right")
 
         expected_text_completions = Expectations({
-            ("cuda", None): [
-                "My favourite condiment is 100% natural, organic, gluten free, vegan, and free from preservatives. I"
-            ],
             ("cuda", 8): [
-                "My favourite condiment is 100% natural, organic, gluten free, vegan, and vegetarian. I love to use"
+                "My favourite condiment is 100% natural, organic, gluten free, vegan, and free from preservatives. I"
             ],
             ("rocm", (9, 4)): [
                 "My favourite condiment is 100% natural, organic and vegan. I love to use it in my cooking, but"

tests/models/qwen2_5_omni/test_modeling_qwen2_5_omni.py

@@ -684,7 +684,7 @@ def test_small_model_integration_test(self):
 
         EXPECTED_DECODED_TEXT = Expectations({
             ("xpu", None): "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is glass shattering, and the dog is a Labrador Retriever.",
-            ("cuda", (8, 6)): "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is glass shattering, and the dog is a Labrador Retriever.",
+            ("cuda", (8, 6)): "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is a glass shattering. The dog in the picture is a Labrador Retriever.",
             ("rocm", (9, 4)): "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is glass shattering, and the dog is a Labrador Retriever.",
         }).get_expectation()  # fmt: skip
 
@@ -720,11 +720,11 @@ def test_small_model_integration_test_batch(self):
                     "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is of glass shattering, and the dog in the picture is a Labrador Retriever",
                 ],
                 ("cuda", 8): [
-                    "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is glass shattering, and the dog is a Labrador Retriever.",
-                    "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is glass shattering, and the dog is a Labrador Retriever.",
+                    "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is a glass shattering. The dog in the picture is a Labrador Retriever.",
+                    "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is a glass shattering. The dog in the picture is a Labrador Retriever.",
                 ],
                 ("rocm", (9, 4)): [
-                    "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is glass shattering, and the dog is a Labrador Retriever.",
+                    "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is a glass shattering. The dog in the picture is a Labrador Retriever.",
                     "system\nYou are a helpful assistant.\nuser\nWhat's that sound and what kind of dog is this?\nassistant\nThe sound is glass shattering, and the dog is a Labrador Retriever.",
                 ],
             }

tests/models/qwen2_5_vl/test_modeling_qwen2_5_vl.py

@@ -572,8 +572,8 @@ def test_small_model_integration_test_batch_different_resolutions(self):
                     "system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\n addCriterion\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and gentle nature, which is",
                 ],
                 ("cuda", (8, 6)): [
-                    'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and energetic nature, which is evident in',
-                    'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and energetic nature, which is evident in',
+                    'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\n addCriterion\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and gentle nature, which is',
+                    'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\n addCriterion\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and gentle nature, which is',
                 ],
                 ("rocm", None): [
                     'system\nYou are a helpful assistant.\nuser\nWhat kind of dog is this?\nassistant\nThe dog in the picture appears to be a Labrador Retriever. Labradors are known for their friendly and energetic nature, which is evident in',
@@ -706,7 +706,7 @@ def test_small_model_integration_test_with_video(self):
         output = model.generate(**inputs, max_new_tokens=30)
 
         EXPECTED_DECODED_TEXT = [
-            'system\nYou are a helpful assistant.\nuser\nWhat is shown in this video?\nassistant\nThe video shows an indoor tennis court with a person standing on one side, preparing to serve the ball. The individual is dressed in athletic attire, including',
+            'system\nYou are a helpful assistant.\nuser\nWhat is shown in this video?\nassistant\nThe video shows an indoor tennis court with a player standing on the baseline, preparing to serve. The player is wearing a white shirt and black shorts,',
         ]  # fmt: skip
         self.assertEqual(
             self.processor.batch_decode(output, skip_special_tokens=True),