fix rocm expectations for intervl

Author: Abdennacer-Badaoui

tests/models/internvl/test_modeling_internvl.py

@@ -430,7 +430,14 @@ def test_qwen2_small_model_integration_batched_generate_multi_image(self):
         # Check first output
         decoded_output = processor.decode(output[0], skip_special_tokens=True)
         # Batching seems to alter the output slightly, but it is also the case in the original implementation. This seems to be expected: https://github.com/huggingface/transformers/issues/23017#issuecomment-1649630232
-        expected_output = "user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature's peace."  # fmt: skip
+        expected_outputs = Expectations(
+            {
+                ("xpu", 3): 'user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature\'s peace.',
+                ("cuda", 7): 'user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature\'s peace.',
+                ("rocm", (9, 4)): 'user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature\'s embrace.',
+            }
+        )  # fmt: skip
+        expected_output = expected_outputs.get_expectation()
         self.assertEqual(
             decoded_output,
             expected_output,
@@ -443,6 +450,7 @@ def test_qwen2_small_model_integration_batched_generate_multi_image(self):
             {
                 ("xpu", 3): "user\n\nWhat are the differences between these two images?\nassistant\nThe images show the Statue of Liberty and the Golden Gate Bridge from different angles. Here are the differences:\n\n1. **Foreground",
                 ("cuda", 7): "user\n\nWhat are the differences between these two images?\nassistant\nThe images show the Statue of Liberty and the Golden Gate Bridge from different angles. Here are the differences:\n\n1. **Foreground",
+                ("rocm", (9, 4)): "user\n\nWhat are the differences between these two images?\nassistant\nThe images show the Statue of Liberty and the Golden Gate Bridge from different angles. Here are the main differences:\n\n1. **",
             }
         )  # fmt: skip
         expected_output = expected_outputs.get_expectation()
@@ -562,11 +570,13 @@ def test_qwen2_small_model_integration_interleaved_images_videos(self):
         output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
 
         decoded_output = processor.decode(output[0], skip_special_tokens=True)
+        print("heeeere :", decoded_output)
         # Batching seems to alter the output slightly, but it is also the case in the original implementation. This seems to be expected: https://github.com/huggingface/transformers/issues/23017#issuecomment-1649630232
         expected_outputs = Expectations(
             {
                 ("xpu", 3): "user\n\n\nWhat are the differences between these two images?\nassistant\nThe images depict two distinct scenes:\n\n1. **Left Image:**\n   - The Statue of Liberty is prominently featured on an",
                 ("cuda", 7): 'user\n\n\nWhat are the differences between these two images?\nassistant\nThe images depict two distinct scenes:\n\n1. **Left Image:**\n   - The Statue of Liberty is prominently featured on an',
+                ("rocm", (9, 4)): 'user\n\n\nWhat are the differences between these two images?\nassistant\nThe images depict two distinct scenes:\n\n1. **Left Image:**\n   - This image features the Statue of Liberty on Liberty',
             }
         )  # fmt: skip
         expected_output = expected_outputs.get_expectation()
@@ -582,6 +592,7 @@ def test_qwen2_small_model_integration_interleaved_images_videos(self):
             {
                 ("xpu", 3): "user\nFrame1: \nFrame2: \nFrame3: \nFrame4: \nFrame5: \nFrame6: \nFrame7: \nFrame8: \nWhat type of shot is the man performing?\nassistant\nA forehand shot",
                 ("cuda", 7): 'user\nFrame1: \nFrame2: \nFrame3: \nFrame4: \nFrame5: \nFrame6: \nFrame7: \nFrame8: \nWhat type of shot is the man performing?\nassistant\nA forehand shot',
+                ("rocm", (9, 4)): 'user\nFrame1: \nFrame2: \nFrame3: \nFrame4: \nFrame5: \nFrame6: \nFrame7: \nFrame8: \nWhat type of shot is the man performing?\nassistant\nA forehand shot',
             }
         )  # fmt: skip
         expected_output = expected_outputs.get_expectation()
@@ -593,9 +604,14 @@ def test_qwen2_small_model_integration_interleaved_images_videos(self):
 
         # Check third output
         decoded_output = processor.decode(output[2], skip_special_tokens=True)
-        expected_output = (
-            "user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature's peace."
-        )
+        expected_outputs = Expectations(
+            {
+                ("xpu", 3): 'user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature\'s peace.',
+                ("cuda", 7): 'user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature\'s peace.',
+                ("rocm", (9, 4)): 'user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature\'s embrace.',
+            }
+        )  # fmt: skip
+        expected_output = expected_outputs.get_expectation()
         self.assertEqual(
             decoded_output,
             expected_output,
@@ -658,7 +674,7 @@ def test_llama_small_model_integration_forward(self):
                 ("xpu", 3): [-9.8828,  -0.4954,   1.4561, -10.3438, -10.3438],
                 ("cuda", 7): [-9.8750,  -0.4861,   1.4648, -10.3359, -10.3359],
                 ("cuda", 8): [-9.8906,  -0.4995,   1.4473, -10.3359, -10.3438],
-                ("rocm", (9, 4)): [ -9.8828,  -0.5005,   1.4697, -10.3438, -10.3438],
+                ("rocm", (9, 4)): [ -9.8672,  -0.4888,   1.4648, -10.3281, -10.3281],
                 ("rocm", (9, 5)): [ -9.8906,  -0.4976,   1.4502, -10.3359, -10.3438],
             }
         )  # fmt: skip
@@ -934,7 +950,7 @@ def test_llama_small_model_integration_interleaved_images_videos(self):
                 ("xpu", 3): "user\n\n\nWhat are the difference between these two images?\nassistant\nI apologize for the confusion in my previous response. Upon closer inspection, the differences between the two images are:\n\n1. **",
                 ("cuda", 7): 'user\n\n\nWhat are the difference between these two images?\nassistant\nI apologize for the confusion in my previous response. Upon closer inspection, the differences between the two images are:\n\n1. **',
                 ("cuda", 8): 'user\n\n\nWhat are the difference between these two images?\nassistant\nI apologize for the confusion in my previous response. After re-examining the images, I can see that there are no',
-                ("rocm", (9, 4)): 'user\n\n\nWhat are the difference between these two images?\nassistant\nI apologize for the confusion in my previous response. Upon closer inspection, the differences between the two images are:\n\n1. **',
+                ("rocm", (9, 4)): 'user\n\n\nWhat are the difference between these two images?\nassistant\nI apologize for the confusion in my previous response. After re-examining the images, I can see that there are no',
                 ("rocm", (9, 5)): 'user\n\n\nWhat are the difference between these two images?\nassistant\nI apologize for the confusion in my previous response. After re-examining the images, I can see that there are no',
             }
         )  # fmt: skip