Changes done internally at Facebook (#1288)

bd46e8f292bf68fe6b87d2d5d206c89fda79a746 Shirong Wu <[email protected]> Disable group ln fuse pass
6ce1d3bc19d75b266e99355c96daeff7054dcbf8 Wei Wei <[email protected]> [fx2trt] set logging level to INFO at fx root
9d552dc3f69db9e4a249f80ef00803a9413e5d38 Wei Wei <[email protected]> [fx2trt] change OSS method lower_to_trt() to compile()


Co-authored-by: wwei6 <[email protected]>

Author: Wei

docsrc/tutorials/getting_started_with_fx_path.rst

@@ -34,7 +34,7 @@ Torch-TensorRT (FX Path) is in ``Beta`` phase and always recommended to work wit
 
 Converting a PyTorch Model to TensorRT Engine
 ---------------------------------------------
-In general, users are welcome to use the ``lower_to_trt()`` to finish the conversion from a model to tensorRT engine. It is a wrapper API that consists of the major steps needed to finish this converison. Please refer to ``lower_example.py`` file in ``examples/fx``.
+In general, users are welcome to use the ``compile()`` to finish the conversion from a model to tensorRT engine. It is a wrapper API that consists of the major steps needed to finish this converison. Please refer to ``lower_example.py`` file in ``examples/fx``.
 
 In this section, we will go through an example to illustrate the major steps that FX path uses. Users can refer to ``fx2trt_example.py`` file in ``examples/fx``.
 
@@ -60,9 +60,9 @@ symbolically traced variables cannot be used as inputs to control flow
 This means the model contains dynamic control flow. Please refer to section “Dynamic Control Flow” in `FX guide <https://pytorch.org/docs/stable/fx.html#dynamic-control-flow>`_.
 
 * **Step 2: Build TensorRT engine**
-There are `two different modes <https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#explicit-implicit-batch>`_ for how TensorRT handles batch dimension, explicit batch dimension and implicit batch dimension. This mode was used by early versions of TensorRT, and is now deprecated but continues to be supported for backwards compatibility. In explicit batch mode, all dimensions are explicit and can be dynamic, that is their length can change at execution time. Many new features, such as dynamic shapes and loops, are available only in this mode. User can still choose to use implicit batch mode when they set ``explicit_batch_dimension=False`` in ``lower_to_trt()``. We do not recommend to use it since it will lack of support in future TensorRT versions.
+There are `two different modes <https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#explicit-implicit-batch>`_ for how TensorRT handles batch dimension, explicit batch dimension and implicit batch dimension. This mode was used by early versions of TensorRT, and is now deprecated but continues to be supported for backwards compatibility. In explicit batch mode, all dimensions are explicit and can be dynamic, that is their length can change at execution time. Many new features, such as dynamic shapes and loops, are available only in this mode. User can still choose to use implicit batch mode when they set ``explicit_batch_dimension=False`` in ``compile()``. We do not recommend to use it since it will lack of support in future TensorRT versions.
 
-Explicit batch is the default mode and it must be set for dynamic shape. For most of vision task, user can choose to enable ``dynamic_batch`` in ``lower_to_trt()`` if they want to get the similar effects as implicit mode where only batch dimension changes. It has some requirements:
+Explicit batch is the default mode and it must be set for dynamic shape. For most of vision task, user can choose to enable ``dynamic_batch`` in ``compile()`` if they want to get the similar effects as implicit mode where only batch dimension changes. It has some requirements:
 1. Shapes of inputs, outputs and activations are fixed except batch dimension.
 2. Inputs, outputs and activations have batch dimension as the major dimension.
 3. All the operators in the model do not modify batch dimension (permute, transpose, split, etc.) or compute over batch dimension (sum, softmax, etc.).

examples/fx/lower_example.py

@@ -4,7 +4,7 @@
 
 import torch
 import torchvision
-from torch_tensorrt.fx.lower import lower_to_trt
+from torch_tensorrt.fx.lower import compile
 from torch_tensorrt.fx.utils import LowerPrecision
 
 
@@ -183,7 +183,7 @@ def run_configuration_benchmark(
         time = benchmark_torch_function(conf.batch_iter, lambda: module(*input))
     elif not conf.jit:
         # Run lowering eager mode benchmark
-        lowered_module = lower_to_trt(
+        lowered_module = compile(
             module,
             input,
             max_batch_size=conf.batch_size,

examples/fx/torchdynamo_example.py

@@ -5,7 +5,7 @@
 import torch
 import torchdynamo
 import torchvision
-from torch_tensorrt.fx.lower import lower_to_trt
+from torch_tensorrt.fx.lower import compile
 from torch_tensorrt.fx.utils import LowerPrecision
 from torchdynamo.optimizations import backends
 
@@ -197,7 +197,7 @@ def run_configuration_benchmark(
 
     if conf.trt:
         # Run lowering eager mode benchmark
-        lowered_module = lower_to_trt(
+        lowered_module = compile(
             module,
             input,
             max_batch_size=conf.batch_size,

py/torch_tensorrt/_compile.py

@@ -7,7 +7,7 @@
 from enum import Enum
 
 import torch_tensorrt.fx
-from torch_tensorrt.fx.lower import lower_to_trt
+import torch_tensorrt.fx.lower
 from torch_tensorrt.fx.utils import LowerPrecision
 
 
@@ -140,7 +140,7 @@ def compile(
         else:
             raise ValueError(f"Precision {enabled_precisions} not supported on FX")
 
-        return lower_to_trt(
+        return torch_tensorrt.fx.lower.compile(
             module,
             inputs,
             lower_precision=lower_precision,

py/torch_tensorrt/fx/__init__.py

@@ -1,4 +1,6 @@
 from .converters import *  # noqa: F403 F401
+import logging
+
 from .converter_registry import (  # noqa
     CONVERTERS,
     NO_EXPLICIT_BATCH_DIM_SUPPORT,
@@ -9,3 +11,5 @@
 from .input_tensor_spec import generate_input_specs, InputTensorSpec  # noqa
 from .lower_setting import LowerSetting  # noqa
 from .trt_module import TRTModule  # noqa
+
+logging.basicConfig(level=logging.INFO)

py/torch_tensorrt/fx/lower.py

@@ -25,7 +25,7 @@
 Input = Sequence[Any]
 
 
-def lower_to_trt(
+def compile(
     module: nn.Module,
     input,
     max_batch_size: int = 2048,
@@ -216,28 +216,32 @@ def create(
             )
         )
 
-    @decorate_method(validate_inference(atol=1e-1, rtol=1e-1))
     def __call__(
         self,
         module: nn.Module,
         inputs: Input,
         additional_inputs: Optional[Input] = None,
     ) -> nn.Module:
-        module.eval()
-
-        if (
-            self.lower_pass_manager_builder.lower_setting.lower_precision
-            == LowerPrecision.FP16
-        ):
-            module.half()
-            inputs = tuple(
-                x.half() if x is not None and x.dtype == torch.float32 else x
-                for x in inputs
+        lower_setting = self.lower_pass_manager_builder.lower_setting
+        atol = lower_setting.correctness_atol
+        rtol = lower_setting.correctness_rtol
+
+        @validate_inference(atol=atol, rtol=rtol)
+        def do_lower(module: nn.Module, inputs: Input) -> nn.Module:
+            module.eval()
+            if (
+                self.lower_pass_manager_builder.lower_setting.lower_precision
+                == LowerPrecision.FP16
+            ):
+                module.half()
+                inputs = tuple(
+                    x.half() if x is not None and x.dtype == torch.float32 else x
+                    for x in inputs
+                )
+            pm = self.lower_pass_manager_builder.build_trt_lower_pipeline(
+                inputs, additional_inputs
             )
-        pm = self.lower_pass_manager_builder.build_trt_lower_pipeline(
-            inputs, additional_inputs
-        )
-
-        lower_result = pm(module)
+            lower_result = pm(module)
+            return lower_result
 
-        return lower_result
+        return do_lower(module, inputs)

py/torch_tensorrt/fx/lower_setting.py

@@ -70,6 +70,8 @@ class LowerSetting(LowerSettingBasic):
     dynamic_batch: enable the dynamic shape in TRT with dim=-1 for the 1st dimension.
     tactic_sources: tactic sources for TensorRT kernel selection. Default to None,
     meaning all possible tactic sources.
+    correctness_atol: absolute tolerance for correctness check
+    correctness_rtol: relative tolerance for correctness check
     """
 
     input_specs: List[InputTensorSpec] = dc.field(default_factory=list)
@@ -90,3 +92,5 @@ class LowerSetting(LowerSettingBasic):
     opt_profile_replica: int = 1
     dynamic_batch: bool = True
     tactic_sources: Optional[int] = None
+    correctness_atol: float = 0.1
+    correctness_rtol: float = 0.1

py/torch_tensorrt/fx/test/converters/acc_op/test_type_as.py

@@ -1,5 +1,6 @@
-import torch
 import unittest
+
+import torch
 import torch_tensorrt.fx.tracer.acc_tracer.acc_ops as acc_ops
 from torch.testing._internal.common_utils import run_tests
 from torch_tensorrt.fx.tools.common_fx2trt import AccTestCase, InputTensorSpec

py/torch_tensorrt/fx/test/passes/test_graph_opts.py

@@ -8,8 +8,6 @@
 import torch_tensorrt.fx.tracer.acc_tracer.acc_tracer as acc_tracer
 from torch_tensorrt.fx.passes.graph_opts import common_subexpression_elimination
 
-_LOGGER: logging.Logger = logging.getLogger(__name__)
-
 
 _LOGGER: logging.Logger = logging.getLogger(__name__)
 

py/torch_tensorrt/fx/test/trt_lower/test_diagnostics.py

@@ -10,8 +10,6 @@
 
 import torch_tensorrt.fx.diagnostics as diag
 
-_LOGGER: logging.Logger = logging.getLogger(__name__)
-
 
 _LOGGER: logging.Logger = logging.getLogger(__name__)