Support compilation via Torchdynamo, AOT Autograd, NVFuser (#17308)

* Support compilation via Torchdynamo, AOT Autograd, NVFuser

* Address comments

* Lint

* Stas comments - missing quality test

* Lintere

* Quality test

* Doc lint

* Reset CUDA peak mem

* Add CustomTrainer

* require a single gpu

Co-authored-by: Stas Bekman <stas@stason.org>

src/transformers/testing_utils.py

@@ -70,6 +70,7 @@ from .utils import (
     is_torch_tf32_available,
     is_torch_tpu_available,
     is_torchaudio_available,
+    is_torchdynamo_available,
     is_vision_available,
 )
 
@@ -464,6 +465,11 @@ else:
     jax_device = None
 
 
+def require_torchdynamo(test_case):
+    """Decorator marking a test that requires TorchDynamo"""
+    return unittest.skipUnless(is_torchdynamo_available(), "test requires TorchDynamo")(test_case)
+
+
 def require_torch_gpu(test_case):
     """Decorator marking a test that requires CUDA and PyTorch."""
     return unittest.skipUnless(torch_device == "cuda", "test requires CUDA")(test_case)

src/transformers/trainer.py

@@ -139,8 +139,10 @@ from .utils import (
     is_sagemaker_dp_enabled,
     is_sagemaker_mp_enabled,
     is_torch_tpu_available,
+    is_torchdynamo_available,
     logging,
 )
+from .utils.generic import ContextManagers
 
 
 _is_torch_generator_available = False
@@ -2172,6 +2174,32 @@ class Trainer:
 
         return inputs
 
+    def compute_loss_context_manager(self):
+        """
+        A helper wrapper to group together context managers.
+        """
+        return ContextManagers(
+            [
+                self.torchdynamo_smart_context_manager(),
+                self.autocast_smart_context_manager(),
+            ]
+        )
+
+    def torchdynamo_smart_context_manager(self):
+        """
+        A helper wrapper that creates an appropriate context manager for `torchdynamo`.
+        """
+        ctx_manager = contextlib.nullcontext()
+        if is_torchdynamo_available():
+            import torchdynamo
+            from torchdynamo.optimizations.training import aot_autograd_speedup_strategy
+
+            if self.args.torchdynamo == "eager":
+                ctx_manager = torchdynamo.optimize("eager")
+            elif self.args.torchdynamo == "nvfuser":
+                ctx_manager = torchdynamo.optimize(aot_autograd_speedup_strategy)
+        return ctx_manager
+
     def autocast_smart_context_manager(self):
         """
         A helper wrapper that creates an appropriate context manager for `autocast` while feeding it the desired
@@ -2213,7 +2241,7 @@ class Trainer:
             loss_mb = smp_forward_backward(model, inputs, self.args.gradient_accumulation_steps, scaler=scaler)
             return loss_mb.reduce_mean().detach().to(self.args.device)
 
-        with self.autocast_smart_context_manager():
+        with self.compute_loss_context_manager():
             loss = self.compute_loss(model, inputs)
 
         if self.args.n_gpu > 1:
@@ -2907,7 +2935,7 @@ class Trainer:
                     logits = smp_nested_concat(logits_mb)
             else:
                 if has_labels:
-                    with self.autocast_smart_context_manager():
+                    with self.compute_loss_context_manager():
                         loss, outputs = self.compute_loss(model, inputs, return_outputs=True)
                     loss = loss.mean().detach()
 
@@ -2917,7 +2945,7 @@ class Trainer:
                         logits = outputs[1:]
                 else:
                     loss = None
-                    with self.autocast_smart_context_manager():
+                    with self.compute_loss_context_manager():
                         outputs = model(**inputs)
                     if isinstance(outputs, dict):
                         logits = tuple(v for k, v in outputs.items() if k not in ignore_keys)

src/transformers/trainer_seq2seq.py

@@ -183,7 +183,7 @@ class Seq2SeqTrainer(Trainer):
             generated_tokens = self._pad_tensors_to_max_len(generated_tokens, gen_kwargs["max_length"])
 
         with torch.no_grad():
-            with self.autocast_smart_context_manager():
+            with self.compute_loss_context_manager():
                 outputs = model(**inputs)
             if has_labels:
                 if self.label_smoother is not None:

src/transformers/training_args.py

@@ -450,6 +450,9 @@ class TrainingArguments:
         full_determinism (`bool`, *optional*, defaults to `False`)
             If `True`, [`enable_full_determinism`] is called instead of [`set_seed`] to ensure reproducible results in
             distributed training
+        torchdynamo (`str`, *optional*):
+            The token that is used to set the backend compiler for TorchDynamo. Possible choices are ["eager",
+            "nvfuser]. This is an experimental API and subject to change.
     """
 
     output_dir: str = field(
@@ -881,6 +884,20 @@ class TrainingArguments:
             )
         },
     )
+    torchdynamo: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": (
+                "Sets up the backend compiler for TorchDynamo. TorchDynamo is a Python level JIT compiler designed to"
+                " make unmodified PyTorch programs faster. TorchDynamo dynamically modifies the Python bytecode right"
+                " before its executed. It rewrites Python bytecode to extract sequences of PyTorch operations"
+                " and lifts them up into Fx graph. We can then pass these Fx graphs to other backend compilers. There"
+                " are two options - eager and nvfuser. Eager defaults to pytorch eager and is useful for debugging."
+                " nvfuser path uses AOT Autograd and nvfuser compiler to optimize the models."
+            ),
+            "choices": ["eager", "nvfuser"],
+        },
+    )
 
     def __post_init__(self):
         # Handle --use_env option in torch.distributed.launch (local_rank not passed as an arg then).

src/transformers/utils/__init__.py

@@ -130,6 +130,7 @@ from .import_utils import (
     is_torch_tf32_available,
     is_torch_tpu_available,
     is_torchaudio_available,
+    is_torchdynamo_available,
     is_training_run_on_sagemaker,
     is_vision_available,
     requires_backends,

src/transformers/utils/import_utils.py

@@ -376,6 +376,10 @@ def is_torch_tpu_available():
     return importlib.util.find_spec("torch_xla.core.xla_model") is not None
 
 
+def is_torchdynamo_available():
+    return importlib.util.find_spec("torchdynamo") is not None
+
+
 def is_datasets_available():
     return _datasets_available
 

tests/trainer/test_trainer.py

@@ -62,6 +62,7 @@ from transformers.testing_utils import (
     require_torch_non_multi_gpu,
     require_torch_tf32,
     require_torch_up_to_2_gpus,
+    require_torchdynamo,
     require_wandb,
     slow,
 )
@@ -1594,6 +1595,100 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
         # perfect world: fp32_init/2 == fp16_eval
         self.assertAlmostEqual(fp16_eval, fp32_init / 2, delta=5_000)
 
+    @require_torch_non_multi_gpu
+    @require_torchdynamo
+    def test_torchdynamo_full_eval(self):
+        # torchdynamo at the moment doesn't support DP/DDP, therefore require a single gpu
+        n_gpus = get_gpu_count()
+
+        bs = 8
+        eval_len = 16 * n_gpus
+        # make the params are somewhat big so that there will be enough RAM consumed to be able to
+        # measure things. We should get about 64KB for a+b in fp32
+        a = torch.ones(1000, bs) + 0.001
+        b = torch.ones(1000, bs) - 0.001
+
+        # 1. Default - without TorchDynamo
+        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len)
+        metrics = trainer.evaluate()
+        original_eval_loss = metrics["eval_loss"]
+        del trainer
+
+        # 2. TorchDynamo eager
+        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, torchdynamo="eager")
+        metrics = trainer.evaluate()
+        self.assertAlmostEqual(metrics["eval_loss"], original_eval_loss)
+        del trainer
+
+        # 3. TorchDynamo nvfuser
+        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, torchdynamo="nvfuser")
+        metrics = trainer.evaluate()
+        self.assertAlmostEqual(metrics["eval_loss"], original_eval_loss)
+
+    @require_torch_non_multi_gpu
+    @require_torchdynamo
+    def test_torchdynamo_memory(self):
+        # torchdynamo at the moment doesn't support DP/DDP, therefore require a single gpu
+        class CustomTrainer(Trainer):
+            def compute_loss(self, model, inputs, return_outputs=False):
+                x = inputs["x"]
+                output = model(x)
+                if self.args.n_gpu == 1:
+                    return output.mean()
+                return output
+
+        class MyModule(torch.nn.Module):
+            """Simple module that does aggressive fusion"""
+
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x):
+                for _ in range(20):
+                    x = torch.nn.functional.relu(x)
+                return x
+
+        mod = MyModule()
+
+        # 1. Default - without TorchDynamo
+        a = torch.ones(1024, 1024, device="cuda", requires_grad=True)
+        a.grad = None
+        trainer = CustomTrainer(model=mod)
+        # warmup
+        for _ in range(10):
+            orig_loss = trainer.training_step(mod, {"x": a})
+
+        torch.cuda.reset_peak_memory_stats()
+        orig_loss = trainer.training_step(mod, {"x": a})
+        orig_peak_mem = torch.cuda.max_memory_allocated()
+        del trainer
+
+        # Reset the peak for another measurement
+        gc.collect()
+        torch.cuda.empty_cache()
+        torch.cuda.reset_peak_memory_stats()
+
+        # 2. TorchDynamo nvfuser
+        a = torch.ones(1024, 1024, device="cuda", requires_grad=True)
+        a.grad = None
+        args = TrainingArguments(output_dir="None", torchdynamo="nvfuser")
+        trainer = CustomTrainer(model=mod, args=args)
+        # warmup
+        for _ in range(10):
+            loss = trainer.training_step(mod, {"x": a})
+
+        torch.cuda.reset_peak_memory_stats()
+        loss = trainer.training_step(mod, {"x": a})
+        peak_mem = torch.cuda.max_memory_allocated()
+        del trainer
+
+        # Functional check
+        self.assertAlmostEqual(loss, orig_loss)
+
+        # AOT Autograd recomputaion and nvfuser recomputation optimization
+        # aggressively fuses the operations and reduce the memory footprint.
+        self.assertGreater(orig_peak_mem, peak_mem * 2)
+
     @require_torch_gpu
     @require_torch_bf16
     def test_bf16_full_eval(self):