[ci] Fixing some failing tests for important models (#43231)

* fix gpt2 tests

* use_safetensors=False for facebook/detr-resnet-50

* use @require_read_token decorator for LlamaTokenizationTest

* update an expected_output

* style

* fix an expectation

* remove @require_read_token decorator, smoother process is coming in #43233

Author: Abdennacer-Badaoui

tests/models/detr/test_modeling_detr.py

@@ -748,7 +748,9 @@ def default_image_processor(self):
         )
 
     def test_inference_no_head(self):
-        model = DetrModel.from_pretrained("facebook/detr-resnet-50", revision="no_timm").to(torch_device)
+        model = DetrModel.from_pretrained("facebook/detr-resnet-50", revision="no_timm", use_safetensors=False).to(
+            torch_device
+        )
 
         image_processor = self.default_image_processor
         image = prepare_img()

tests/models/gpt2/test_modeling_gpt2.py

@@ -461,6 +461,7 @@ def test_batch_generation(self):
         tokenizer = GPT2Tokenizer.from_pretrained("openai-community/gpt2")
 
         tokenizer.padding_side = "left"
+        max_length = 20
 
         # Define PAD Token = EOS Token = 50256
         tokenizer.pad_token = tokenizer.eos_token
@@ -485,24 +486,22 @@ def test_batch_generation(self):
         outputs = model.generate(
             input_ids=input_ids,
             attention_mask=inputs["attention_mask"].to(torch_device),
-            max_length=20,
+            max_length=max_length,
         )
 
         outputs_tt = model.generate(
             input_ids=input_ids,
             attention_mask=inputs["attention_mask"].to(torch_device),
             token_type_ids=token_type_ids,
-            max_length=20,
+            max_length=max_length,
         )
 
         inputs_non_padded = tokenizer(sentences[0], return_tensors="pt").input_ids.to(torch_device)
-        output_non_padded = model.generate(input_ids=inputs_non_padded, max_length=20)
+        output_non_padded = model.generate(input_ids=inputs_non_padded, max_length=max_length)
 
         num_paddings = inputs_non_padded.shape[-1] - inputs["attention_mask"][-1].long().sum().item()
         inputs_padded = tokenizer(sentences[1], return_tensors="pt").input_ids.to(torch_device)
-        output_padded = model.generate(
-            input_ids=inputs_padded, max_length=model.generation_config.max_length - num_paddings
-        )
+        output_padded = model.generate(input_ids=inputs_padded, max_length=max_length - num_paddings)
 
         batch_out_sentence = tokenizer.batch_decode(outputs, skip_special_tokens=True)
         batch_out_sentence_tt = tokenizer.batch_decode(outputs_tt, skip_special_tokens=True)
@@ -524,6 +523,7 @@ def test_batch_generation_2heads(self):
         tokenizer = GPT2Tokenizer.from_pretrained("openai-community/gpt2")
 
         tokenizer.padding_side = "left"
+        max_length = 20
 
         # This tokenizer has no pad token, so we have to set it in some way
         # Define PAD Token = EOS Token = 50256
@@ -549,24 +549,22 @@ def test_batch_generation_2heads(self):
         outputs = model.generate(
             input_ids=input_ids,
             attention_mask=inputs["attention_mask"].to(torch_device),
-            max_length=20,
+            max_length=max_length,
         )
 
         outputs_tt = model.generate(
             input_ids=input_ids,
             attention_mask=inputs["attention_mask"].to(torch_device),
             token_type_ids=token_type_ids,
-            max_length=20,
+            max_length=max_length,
         )
 
         inputs_non_padded = tokenizer(sentences[0], return_tensors="pt").input_ids.to(torch_device)
-        output_non_padded = model.generate(input_ids=inputs_non_padded, max_length=20)
+        output_non_padded = model.generate(input_ids=inputs_non_padded, max_length=max_length)
 
         num_paddings = inputs_non_padded.shape[-1] - inputs["attention_mask"][-1].long().sum().item()
         inputs_padded = tokenizer(sentences[1], return_tensors="pt").input_ids.to(torch_device)
-        output_padded = model.generate(
-            input_ids=inputs_padded, max_length=model.generation_config.max_length - num_paddings
-        )
+        output_padded = model.generate(input_ids=inputs_padded, max_length=max_length - num_paddings)
 
         batch_out_sentence = tokenizer.batch_decode(outputs, skip_special_tokens=True)
         batch_out_sentence_tt = tokenizer.batch_decode(outputs_tt, skip_special_tokens=True)

tests/models/llava/test_modeling_llava.py

@@ -585,7 +585,7 @@ def test_tokenizer_integration(self):
         fast_tokenizer.add_tokens("<image>", True)
 
         prompt = "<|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nWhat is shown in this image?<|im_end|><|im_start|>assistant\n"
-        EXPECTED_OUTPUT = ['<|im_start|>', 'system', '\n', 'Answer', '▁the', '▁questions', '.', '<|im_end|>', '<|im_start|>', 'user', '\n', '<image>', '\n', 'What', '▁is', '▁shown', '▁in', '▁this', '▁image', '?', '<|im_end|>', '<|im_start|>', 'ass', 'istant', '\n']  # fmt: skip
+        EXPECTED_OUTPUT = ['<|im_start|>', 'sy', 'st', 'em', '\n', 'An', 'sw', 'er', ' ', 'the', ' ', 'qu', 'est', 'ions', '.', '<|im_end|>', '<|im_start|>', 'us', 'er', '\n', '<image>', '\n', 'What', ' ', 'is', ' ', 'sh', 'own', ' ', 'in', ' ', 'th', 'is', ' ', 'im', 'age', '?', '<|im_end|>', '<|im_start|>', 'ass', 'ist', 'ant', '\n']  # fmt: skip
         self.assertEqual(slow_tokenizer.tokenize(prompt), EXPECTED_OUTPUT)
         self.assertEqual(fast_tokenizer.tokenize(prompt), EXPECTED_OUTPUT)
 

tests/models/qwen2_5_omni/test_modeling_qwen2_5_omni.py

@@ -723,7 +723,7 @@ 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 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 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.",
                     "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.",
                 ],
             }