Fix attn mask ignore logic in training-time trace (huggingface#32613)

* fix attn mask logic for training-time trace

* add test

* fix

* fix

* fix

* fix

* fix

* format

* [run-slow] llama

* avoid accelearate

* [run-slow] llama

src/transformers/cache_utils.py

@@ -1284,7 +1284,6 @@ def __init__(
         max_batch_size: Optional[int] = None,
         layer_device_map: Optional[Dict[int, Union[str, torch.device, int]]] = None,
     ) -> None:
-        super().__init__()
         if not hasattr(config, "sliding_window") or config.sliding_window is None:
             raise ValueError(
                 "Setting `cache_implementation` to 'sliding_window' requires the model config supporting "

src/transformers/modeling_attn_mask_utils.py

@@ -281,7 +281,7 @@ def _ignore_causal_mask_sdpa(
         elif sliding_window is None or key_value_length < sliding_window:
             if len(attention_mask.shape) == 4:
                 return False
-            elif (is_training or not is_tracing) and torch.all(attention_mask == 1):
+            elif not is_tracing and torch.all(attention_mask == 1):
                 if query_length == 1 or key_value_length == query_length:
                     # For query_length == 1, causal attention and bi-directional attention are the same.
                     ignore_causal_mask = True

tests/models/gemma/test_modeling_gemma.py

@@ -321,6 +321,9 @@ class GemmaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
     # used in `test_torch_compile`
     _torch_compile_test_ckpt = "google/gemma-2b"
 
+    # used in `test_torch_compile_for_training`
+    _torch_compile_train_cls = GemmaForCausalLM if is_torch_available() else None
+
     # TODO (ydshieh): Check this. See https://app.circleci.com/pipelines/github/huggingface/transformers/79245/workflows/9490ef58-79c2-410d-8f51-e3495156cf9c/jobs/1012146
     def is_pipeline_test_to_skip(
         self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
@@ -808,7 +811,7 @@ def test_compile_static_cache(self):
 
         prompts = ["Hello I am doing", "Hi today"]
         tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b", pad_token="</s>", padding_side="right")
-        model = GemmaForCausalLM.from_pretrained("google/gemma-2b", device_map="sequential", torch_dtype=torch.float16)
+        model = GemmaForCausalLM.from_pretrained("google/gemma-2b", device_map=torch_device, torch_dtype=torch.float16)
         inputs = tokenizer(prompts, return_tensors="pt", padding=True).to(model.device)
 
         # Dynamic Cache

tests/models/llama/test_modeling_llama.py

@@ -319,6 +319,9 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
     # used in `test_torch_compile`
     _torch_compile_test_ckpt = "meta-llama/Llama-2-7b-hf"
 
+    # used in `test_torch_compile_for_training`
+    _torch_compile_train_cls = LlamaForCausalLM if is_torch_available() else None
+
     def setUp(self):
         self.model_tester = LlamaModelTester(self)
         self.config_tester = ConfigTester(self, config_class=LlamaConfig, hidden_size=37)
@@ -874,7 +877,7 @@ def test_compile_static_cache(self):
         ]
         tokenizer = LlamaTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf", pad_token="</s>", padding_side="right")
         model = LlamaForCausalLM.from_pretrained(
-            "meta-llama/Llama-2-7b-hf", device_map="sequential", torch_dtype=torch.float16
+            "meta-llama/Llama-2-7b-hf", device_map=torch_device, torch_dtype=torch.float16
         )
         inputs = tokenizer(prompts, return_tensors="pt", padding=True).to(model.device)
 

tests/models/mistral/test_modeling_mistral.py

@@ -677,7 +677,7 @@ def test_compile_static_cache(self):
         tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1", use_fast=False)
         tokenizer.pad_token = tokenizer.eos_token
         model = MistralForCausalLM.from_pretrained(
-            "mistralai/Mistral-7B-v0.1", device_map="sequential", torch_dtype=torch.float16
+            "mistralai/Mistral-7B-v0.1", device_map=torch_device, torch_dtype=torch.float16
         )
         inputs = tokenizer(prompts, return_tensors="pt", padding=True).to(model.device)
 

tests/models/nemotron/test_modeling_nemotron.py

@@ -94,6 +94,8 @@ class NemotronModelTest(GemmaModelTest):
 
     # used in `test_torch_compile`
     _torch_compile_test_ckpt = "nvidia/nemotron-3-8b-base-4k-hf"
+    # used in `test_torch_compile_for_training`
+    _torch_compile_train_cls = NemotronForCausalLM if is_torch_available() else None
 
     def setUp(self):
         self.model_tester = NemotronModelTester(self)

tests/test_modeling_common.py

@@ -4937,6 +4937,49 @@ def test_torch_compile(self):
         for i in range(n_iter):
             _ = model.generate(**input_ids, do_sample=False)
 
+    @slow
+    @require_torch_gpu
+    def test_torch_compile_for_training(self):
+        if version.parse(torch.__version__) < version.parse("2.3"):
+            self.skipTest(reason="This test requires torch >= 2.3 to run.")
+
+        if not hasattr(self, "_torch_compile_train_cls"):
+            self.skipTest(f"{self.__class__.__name__} doesn't have the attribute `_torch_compile_train_cls`.")
+
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+        cls = self._torch_compile_train_cls
+        model = cls(config).to(torch_device)
+
+        inputs = {
+            "input_ids": torch.randint(low=1, high=model.config.vocab_size, size=(2, 10), device=torch_device),
+            "attention_mask": torch.tensor(
+                [[1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1]],
+                dtype=torch.int64,
+                device=torch_device,
+            ),
+            "position_ids": torch.arange(0, 10, device=torch_device).unsqueeze(0),
+            "labels": torch.randint(low=1, high=model.config.vocab_size, size=(2, 10), device=torch_device),
+        }
+
+        # eager backward
+        set_seed(42)
+        loss = model(**inputs).loss
+        loss.backward()
+
+        params = {name: param.grad.clone().detach().cpu() for name, param in model.named_parameters()}
+        model.zero_grad()
+        del loss
+
+        model = torch.compile(model, fullgraph=True, mode="reduce-overhead")
+        # forward compilation
+        set_seed(42)
+        loss = model(**inputs).loss
+        # backward compilation
+        loss.backward()
+        # check grad matches
+        for name, param in model._orig_mod.named_parameters():
+            torch.testing.assert_close(param.grad.detach().cpu(), params[name], rtol=1e-4, atol=1e-4)
+
     @slow
     @require_torch_gpu  # Testing cuda graphs.
     @require_read_token