update

src/transformers/models/llava_next_video/modeling_llava_next_video.py

@@ -208,7 +208,7 @@ def __init__(self, config):
         mode = config.spatial_pool_mode
         stride = config.spatial_pool_stride
         out_channels = getattr(config, "spatial_pool_out_channels", config.vision_config.hidden_size)
-        self.image_size = config.vision_config.image_size // config.vision_config.patch_size**2
+        self.image_size = (config.vision_config.image_size // config.vision_config.patch_size) ** 2
 
         if mode == "average":
             self.pool = nn.AvgPool2d(kernel_size=stride, stride=stride)

src/transformers/models/unispeech/modeling_unispeech.py

@@ -1577,7 +1577,7 @@ def forward(
         quantized_features, codevector_perplexity = self.quantizer(extract_features)
 
         # project quantized features twice
-        quantized_features = self.project_q(quantized_features)
+        quantized_features = self.project_q(quantized_features.to(self.project_q.weight.dtype))
         quantized_features = self.project_hid(quantized_features)
 
         prob_replace_matrix = torch.empty(transformer_features.size(0), transformer_features.size(1)).fill_(

tests/models/albert/test_modeling_albert.py

@@ -16,11 +16,12 @@
 
 import unittest
 
+from parameterized import parameterized
 from packaging import version
 
 from transformers import AlbertConfig, AutoTokenizer, is_torch_available
 from transformers.models.auto import get_values
-from transformers.testing_utils import require_torch, slow, torch_device
+from transformers.testing_utils import require_torch_sdpa, require_torch, slow, torch_device
 
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
@@ -288,6 +289,12 @@ def setUp(self):
         self.model_tester = AlbertModelTester(self)
         self.config_tester = ConfigTester(self, config_class=AlbertConfig, hidden_size=37)
 
+    @parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
+    @require_torch_sdpa
+    @unittest.skip("Albert requires `head_mask` which is currently not done in this test.")
+    def test_eager_matches_sdpa_inference(self):
+        pass
+
     def test_config(self):
         self.config_tester.run_common_tests()
 

tests/models/granite/test_modeling_granite.py

@@ -445,15 +445,6 @@ def test_use_flash_attention_2_true(self):
                 if not has_flash:
                     raise ValueError("The flash model should have flash attention layers")
 
-    @parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
-    @require_torch_sdpa
-    @slow
-    def test_eager_matches_sdpa_inference(self, torch_dtype: str):
-        """
-        skipping the test since mup is very flaky and gets consistently different outputs
-        """
-        self.skipTest("skipping the test since mup is very flaky and gets consistently different outputs")
-
 
 @require_torch_gpu
 class GraniteIntegrationTest(unittest.TestCase):

tests/models/granitemoe/test_modeling_granitemoe.py

@@ -444,15 +444,6 @@ def test_use_flash_attention_2_true(self):
                 if not has_flash:
                     raise ValueError("The flash model should have flash attention layers")
 
-    @parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
-    @require_torch_sdpa
-    @slow
-    def test_eager_matches_sdpa_inference(self, torch_dtype: str):
-        """
-        skipping the test since mup is very flaky and gets consistently different outputs
-        """
-        self.skipTest("skipping the test since mup is very flaky and gets consistently different outputs")
-
 
 @require_torch_gpu
 class GraniteMoeIntegrationTest(unittest.TestCase):

tests/models/idefics/test_modeling_idefics.py

@@ -134,7 +134,7 @@ def __init__(
             num_attention_heads=self.vision_num_attention_heads,
             num_hidden_layers=self.vision_num_hidden_layers,
             intermediate_size=self.vision_intermediate_size,
-        )
+        ).to_dict()
 
         self.perceiver_qk_layer_norms_perceiver = perceiver_qk_layer_norms_perceiver
         self.perceiver_resampler_depth = perceiver_resampler_depth
@@ -316,7 +316,6 @@ def prepare_pixel_values(self):
         return floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
 
     @require_torch_sdpa
-    @slow
     @parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
     def test_eager_matches_sdpa_inference(self, torch_dtype: str):
         self.skipTest(reason="Idefics has a hard requirement on SDPA, skipping this test")
@@ -349,6 +348,12 @@ def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
 
         return inputs_dict
 
+    @parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
+    @require_torch_sdpa
+    @unittest.skip("Idefics requires both text and image inputs which is currently not done in this test.")
+    def test_eager_matches_sdpa_inference(self):
+        pass
+
     def test_model_outputs_equivalence(self):
         try:
             orig = self.all_model_classes
@@ -598,6 +603,12 @@ def setUp(self):
         )
         self.config_tester = ConfigTester(self, config_class=IdeficsConfig, hidden_size=37)
 
+    @parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
+    @require_torch_sdpa
+    @unittest.skip("Idefics requires both text and image inputs which is currently not done in this test.")
+    def test_eager_matches_sdpa_inference(self, torch_dtype):
+        pass
+
     @pytest.mark.generate
     def test_left_padding_compatibility(self):
         """Overwrite because IDEFICS needs image attention mask to be also padded"""

tests/models/llava_next/test_modeling_llava_next.py

@@ -90,7 +90,7 @@ def __init__(
         },
         is_training=True,
         vision_config={
-            "image_size": 16,
+            "image_size": 8,
             "patch_size": 4,
             "num_channels": 3,
             "is_training": True,
@@ -123,10 +123,10 @@ def __init__(
         self.batch_size = 3
         self.num_channels = 3
         self.image_size = 30
-        self.encoder_seq_length = 95
-        self.image_grid_pinpoints = [[32, 32]]
-        self.num_image_tokens = 88
+        self.image_grid_pinpoints = [[16, 16]]
+        self.num_image_tokens = 24
         self.seq_length = seq_length + self.num_image_tokens
+        self.encoder_seq_length = self.seq_length
 
     def get_config(self):
         return LlavaNextConfig(

tests/models/llava_next_video/test_modeling_llava_next_video.py

@@ -91,7 +91,7 @@ def __init__(
         },
         is_training=True,
         vision_config={
-            "image_size": 16,
+            "image_size": 8,
             "patch_size": 4,
             "num_channels": 3,
             "is_training": True,
@@ -125,11 +125,12 @@ def __init__(
         self.batch_size = 3
         self.num_channels = 3
         self.image_size = 30
-        self.encoder_seq_length = 127
-        self.image_grid_pinpoints = [[32, 32]]
-        self.num_image_tokens = 88
-        self.num_video_tokens = 32
+
+        self.image_grid_pinpoints = [[16, 16]]
+        self.num_image_tokens = 24
+        self.num_video_tokens = 8
         self.seq_length = seq_length + self.num_image_tokens + self.num_video_tokens
+        self.encoder_seq_length = self.seq_length
 
     def get_config(self):
         return LlavaNextVideoConfig(

tests/models/mimi/test_modeling_mimi.py

@@ -411,8 +411,10 @@ def test_identity_shortcut(self):
 
     @parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
     @require_torch_sdpa
-    @slow
     def test_eager_matches_sdpa_inference(self, torch_dtype: str):
+        if torch_dtype == "float16" and torch_device == "cpu":
+            self.skipTest("`replication_pad1d` not implemented for 'Half")
+
         if not self.has_attentions:
             self.skipTest(reason="Model architecture does not support attentions")
 
@@ -513,7 +515,7 @@ def get_mean_reldiff(failcase, x, ref, atol, rtol):
                             can_output_attn = "output_attentions" in inspect.signature(model_sdpa.forward).parameters
                             if not (self.has_attentions and can_output_attn) and output_attentions:
                                 continue
-                            for batch_size in [1, 5]:
+                            for batch_size in [7]:
                                 dummy_input = inputs_dict[model.main_input_name]
 
                                 if dummy_input.dtype in [torch.float32, torch.bfloat16, torch.float16]:
@@ -564,11 +566,11 @@ def get_mean_reldiff(failcase, x, ref, atol, rtol):
 
                                     dummy_attention_mask[:] = 1
                                     if padding_side == "left":
-                                        dummy_attention_mask[-1, :-1] = 1
-                                        dummy_attention_mask[-1, -4:] = 0
+                                        dummy_attention_mask[-1, :2] = 0
+                                        dummy_attention_mask[-1, 2:] = 1
                                     elif padding_side == "right":
-                                        dummy_attention_mask[-1, 1:] = 1
-                                        dummy_attention_mask[-1, :3] = 0
+                                        dummy_attention_mask[-1, -2:] = 0
+                                        dummy_attention_mask[-1, :-2] = 1
 
                                 for enable_kernels in [False, True]:
                                     failcase = f"padding_side={padding_side}, use_mask={use_mask}, batch_size={batch_size}, enable_kernels={enable_kernels}"
@@ -655,52 +657,32 @@ def get_mean_reldiff(failcase, x, ref, atol, rtol):
 
                                     # Masked tokens output slightly deviates - we don't mind that.
                                     if use_mask:
+                                        _logits_sdpa = torch.zeros_like(input=logits_sdpa)
+                                        _logits_eager = torch.zeros_like(input=logits_eager)
+
+                                        _logits_sdpa[:-1] = logits_sdpa[:-1]
+                                        _logits_eager[:-1] = logits_eager[:-1]
+
                                         if padding_side == "left":
-                                            sub_sdpa = logits_sdpa[:-1]
-                                            sub_eager = logits_eager[:-1]
-                                            if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
-                                                fail_cases.append(
-                                                    get_mean_reldiff(failcase, sub_sdpa, sub_eager, atol, rtol)
-                                                )
-
-                                            sub_sdpa = logits_sdpa[-1, :-4]
-                                            sub_eager = logits_eager[-1, :-4]
-                                            if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
-                                                fail_cases.append(
-                                                    get_mean_reldiff(failcase, sub_sdpa, sub_eager, atol, rtol)
-                                                )
-
-                                            # Testing the padding tokens is not really meaningful but anyway
-                                            # sub_sdpa = logits_sdpa[-1, -4:]
-                                            # sub_eager = logits_eager[-1, -4:]
-                                            # if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
-                                            #     fail_cases.append(get_mean_reldiff(failcase, sub_sdpa, sub_eager, 4e-2, 4e-2))
-                                        elif padding_side == "right":
-                                            sub_sdpa = logits_sdpa[:-1]
-                                            sub_eager = logits_eager[:-1]
-                                            if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
-                                                fail_cases.append(
-                                                    get_mean_reldiff(failcase, sub_sdpa, sub_eager, atol, rtol)
-                                                )
-
-                                            sub_sdpa = logits_sdpa[-1, 3:]
-                                            sub_eager = logits_eager[-1, 3:]
-                                            if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
-                                                fail_cases.append(
-                                                    get_mean_reldiff(failcase, sub_sdpa, sub_eager, atol, rtol)
-                                                )
-
-                                            # Testing the padding tokens is not really meaningful but anyway
-                                            # sub_sdpa = logits_sdpa[-1, :3]
-                                            # sub_eager = logits_eager[-1, :3]
-                                            # if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
-                                            #     fail_cases.append(get_mean_reldiff(failcase, sub_sdpa, sub_eager, 4e-2, 4e-2))
+                                            _logits_sdpa[-1:, 2:] = logits_sdpa[-1:, 2:]
+                                            _logits_eager[-1:, 2:] = logits_eager[-1:, 2:]
 
-                                    else:
-                                        if not torch.allclose(logits_sdpa, logits_eager, atol=atol, rtol=rtol):
-                                            fail_cases.append(
-                                                get_mean_reldiff(failcase, logits_sdpa, logits_eager, atol, rtol)
-                                            )
+                                        elif padding_side == "right":
+                                            _logits_sdpa[-1:, 2:] = logits_sdpa[-1:, :-2]
+                                            _logits_eager[-1:, 2:] = logits_eager[-1:, :-2]
+
+                                        logits_sdpa = _logits_sdpa
+                                        logits_eager = _logits_eager
+
+                                    results = [
+                                        torch.allclose(_logits_sdpa, _logits_eager, atol=atol, rtol=rtol)
+                                        for (_logits_sdpa, _logits_eager) in zip(logits_sdpa, logits_eager)
+                                    ]
+                                    # If 80% batch elements have matched results, it's fine
+                                    if np.mean(results) < 0.8:
+                                        fail_cases.append(
+                                            get_mean_reldiff(failcase, logits_sdpa, logits_eager, atol, rtol)
+                                        )
 
                 self.assertTrue(len(fail_cases) == 0, "\n".join(fail_cases))
 

tests/models/mllama/test_modeling_mllama.py

@@ -353,13 +353,6 @@ def _check_attentions_for_generate(
 
             self.assertListEqual([layer_attention.shape for layer_attention in iter_attentions], expected_shapes)
 
-    @require_torch_sdpa
-    @slow
-    @is_flaky()
-    def test_eager_matches_sdpa_inference_1_bfloat16(self):
-        # A workaround to override parametrized test with flaky decorator
-        super().test_eager_matches_sdpa_inference_1_bfloat16()
-
     @unittest.skip("For some unknown reasons the tests fails in CrossAttention layer when doing torch.sdpa(). ")
     def test_sdpa_can_compile_dynamic(self):
         pass