Changes done internally at Facebook (#1178)

6703b98dff0695d91026f057b951dba1355825fa Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.prod
c822345d6d673e1653c2208435e34ab400bada3d Jason Park <jasonjk@fb.com> Add support for generic torch ops to be used in training.

Co-authored-by: wwei6 <wwei6@devvm5812.prn0.facebook.com>

Author: Wei

py/torch_tensorrt/fx/__init__.py

@@ -6,5 +6,6 @@
     tensorrt_converter,
 )
 from .fx2trt import TRTInterpreter, TRTInterpreterResult  # noqa
-from .input_tensor_spec import InputTensorSpec  # noqa
+from .input_tensor_spec import generate_input_specs, InputTensorSpec  # noqa
+from .lower_setting import LowerSetting  # noqa
 from .trt_module import TRTModule  # noqa

py/torch_tensorrt/fx/input_tensor_spec.py

@@ -1,11 +1,66 @@
-from typing import Iterable, List, NamedTuple, Sequence, Tuple
+from typing import Iterable, List, NamedTuple, Optional, Sequence, Tuple
 
 import torch
 
 from .types import Shape, ShapeRange
 from .utils import get_dynamic_dims
 
 
+def generate_input_specs(
+    inputs, lower_setting, additional_inputs=None, fixed_shape=False
+):
+    # AIT lower setting doesn't have explicit_batch_dimension field and
+    # we just return None.
+    if not hasattr(lower_setting, "explicit_batch_dimension"):
+        return None
+
+    if not lower_setting.explicit_batch_dimension or fixed_shape:
+        return InputTensorSpec.from_tensors(inputs)
+
+    # If we don't have additional inputs, we assume the first dimension
+    # is the dynamic batch dimension. Otherwise, we use the additional
+    # inputs to determine the batch dimension.
+    if additional_inputs is None:
+        return InputTensorSpec.from_tensors_with_dynamic_batch_size(
+            inputs,
+            (
+                0,
+                lower_setting.max_batch_size,
+                lower_setting.max_batch_size,
+            ),
+            lower_setting.opt_profile_replica,
+        )
+    else:
+        batch_dims = []
+
+        for i, j in zip(inputs, additional_inputs):
+            found_batch_dim = False
+
+            for idx, values in enumerate(zip(i.shape, j.shape)):
+                if values[0] != values[1]:
+                    assert (
+                        found_batch_dim is False
+                    ), f"We've already found a batch dim, {i.shape}, {j.shape}."
+                    batch_dims.append(idx)
+                    found_batch_dim = True
+
+            if not found_batch_dim:
+                raise RuntimeError(
+                    f"Failed to find batch dimension because shapes are the same, {i.shape}"
+                )
+
+        return InputTensorSpec.from_tensors_with_dynamic_batch_size(
+            inputs,
+            (
+                0,
+                lower_setting.max_batch_size,
+                lower_setting.max_batch_size,
+            ),
+            lower_setting.opt_profile_replica,
+            batch_dims,
+        )
+
+
 class InputTensorSpec(NamedTuple):
     """
     This class contains the information of a input tensor.
@@ -70,6 +125,7 @@ def from_tensors_with_dynamic_batch_size(
         tensors: Sequence[torch.Tensor],
         batch_size_range: Tuple[int, int, int],
         opt_profile_replica: int = 1,
+        batch_dims: Optional[List[int]] = None,
     ) -> List["InputTensorSpec"]:
         """
         Produce a list of InputTenosrSpec named tuples which would contain
@@ -83,20 +139,30 @@ def from_tensors_with_dynamic_batch_size(
                 the smallest batch size allowed. The second integer indiceates
                 the batch size that we'll optimize for. The third integer indicates
                 the largest batch size allowed.
+            opt_profile_replica (int): If dynamic shape is enabled, each execution
+                context requires a different optimization profile. This arg determines
+                how many optimization profile replicas we want to produce.
+            batch_dims (Optional[List[int]]): The batch dim might not be the leading dim
+                and allow user to specify the batch dims using this arg. Default we treat
+                dim 0 as the batch dim.
 
         Returns:
             A list of InputTensorSpec named tuples with dynamic ranges.
         """
+        if batch_dims is None:
+            batch_dims = [0] * len(tensors)
+
         input_specs = []
-        batch_size = tensors[0].size(0)
+        batch_size = tensors[0].size(batch_dims[0])
 
         for i, tensor in enumerate(tensors):
+            batch_dim = batch_dims[i]
             assert batch_size == tensor.size(
-                0
+                batch_dim
             ), f"The {i}th tensor (shape: {tensor.shape}) doesn't have the correct batch size: {batch_size}."
             shape = list(tensor.shape)
-            shape[0] = -1
-            shape_ranges: List[ShapeRange] = [tuple(tuple([bs] + shape[1:]) for bs in batch_size_range)] * opt_profile_replica  # type: ignore[list-item]
+            shape[batch_dim] = -1
+            shape_ranges: List[ShapeRange] = [tuple(tuple(shape[0:batch_dim] + [bs] + shape[batch_dim + 1 :]) for bs in batch_size_range)] * opt_profile_replica  # type: ignore[list-item]
             input_specs.append(
                 cls(tuple(shape), tensor.dtype, tensor.device, shape_ranges)
             )

py/torch_tensorrt/fx/lower.py

@@ -1,6 +1,6 @@
 import dataclasses as dc
 import logging
-from typing import Any, Callable, Sequence
+from typing import Any, Callable, Optional, Sequence
 
 # @manual=//deeplearning/trt/python:py_tensorrt
 import tensorrt as trt
@@ -10,15 +10,9 @@
 from torch.fx.passes.splitter_base import SplitResult
 
 from .fx2trt import TRTInterpreter, TRTInterpreterResult
-from .input_tensor_spec import InputTensorSpec
 from .lower_setting import LowerSetting
 from .passes.lower_pass_manager_builder import LowerPassManagerBuilder
-from .passes.pass_utils import (
-    chain_passes,
-    decorate_method,
-    PassFunc,
-    validate_inference,
-)
+from .passes.pass_utils import decorate_method, PassFunc, validate_inference
 from .tools.timing_cache_utils import TimingCacheManager
 from .tools.trt_splitter import TRTSplitter, TRTSplitterSetting
 
@@ -91,25 +85,8 @@ def create(cls, lower_setting):
         return LowerTrtInterpreter(lower_setting, timing_cache_manager)
 
     def __call__(self, mod, input, split_name) -> TRTInterpreterResult:
-        input_specs_val = (
-            self.lower_setting.input_specs
-            if self.lower_setting.input_specs
-            else (
-                InputTensorSpec.from_tensors_with_dynamic_batch_size(
-                    input,
-                    (
-                        0,
-                        self.lower_setting.max_batch_size,
-                        self.lower_setting.max_batch_size,
-                    ),
-                    self.lower_setting.opt_profile_replica,
-                )
-                if self.lower_setting.explicit_batch_dimension
-                and self.lower_setting.dynamic_batch
-                else InputTensorSpec.from_tensors(input)
-            )
-        )
-        logger.info(f"{split_name=} {input_specs_val=}")
+        assert self.lower_setting.input_specs, "Can't find input specs for lowering!"
+        logger.info(f"{split_name=} {self.lower_setting.input_specs=}")
 
         # Prepare algorithm selector and timing_cache for TRTInterpreter
         algo_selector = None
@@ -125,7 +102,7 @@ def __call__(self, mod, input, split_name) -> TRTInterpreterResult:
 
         interpreter = TRTInterpreter(
             mod,
-            input_specs=input_specs_val,
+            input_specs=self.lower_setting.input_specs,
             explicit_batch_dimension=self.lower_setting.explicit_batch_dimension,
             explicit_precision=self.lower_setting.explicit_precision,
             logger_level=trt.Logger.VERBOSE
@@ -242,6 +219,7 @@ def __call__(
         self,
         module: nn.Module,
         inputs: Input,
+        additional_inputs: Optional[Input] = None,
     ) -> nn.Module:
         module.eval()
 
@@ -254,7 +232,9 @@ def __call__(
                 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_lower_pipeline(inputs)
+        pm = self.lower_pass_manager_builder.build_lower_pipeline(
+            inputs, additional_inputs
+        )
 
         lower_result = pm(module)
 

py/torch_tensorrt/fx/passes/lower_pass_manager_builder.py

@@ -1,11 +1,13 @@
 from functools import partial, wraps
-from typing import Any, Callable, Sequence
+from typing import Any, Callable, Optional, Sequence
 
 import torch
 from torch import nn
 from torch.fx.passes.pass_manager import inplace_wrapper, PassManager
 from torch.fx.passes.shape_prop import ShapeProp
-from torch.fx.passes.splitter_base import SplitResult
+from torch.fx.passes.splitter_base import generate_inputs_for_submodules, SplitResult
+
+from ..input_tensor_spec import generate_input_specs
 
 from ..lower_setting import LowerSetting
 from ..observer import Observer
@@ -120,13 +122,33 @@ def _split_pass(self) -> PassManager:
 
     def _lower_pass(self) -> PassManager:
         def lower_func(split_result: SplitResult) -> nn.Module:
+            if (
+                hasattr(self.lower_setting, "explicit_batch_dimension")
+                and self.lower_setting.explicit_batch_dimension
+                and self._additional_input
+            ):
+                additional_submodule_inputs = generate_inputs_for_submodules(
+                    split_result.split_module,
+                    self._additional_input,
+                    list(split_result.submodule_inputs.keys()),
+                )
+            else:
+                additional_submodule_inputs = None
+
             for submod_name, submod_inputs in split_result.submodule_inputs.items():
                 submod = getattr(split_result.split_module, submod_name)
 
                 LOWER_SPLIT_PRE_OBSERVER.observe(submod_name, submod, submod_inputs)
 
                 # Only acc submodules will be lowered.
                 if not submod_name.startswith(split_result.non_acc_submodule_prefix):
+                    self.lower_setting.input_specs = generate_input_specs(
+                        submod_inputs,
+                        self.lower_setting,
+                        additional_submodule_inputs[submod_name]
+                        if additional_submodule_inputs
+                        else None,
+                    )
                     lowered_module = self._lower_func(
                         submod, submod_inputs, self.lower_setting, submod_name
                     )
@@ -139,8 +161,11 @@ def lower_func(split_result: SplitResult) -> nn.Module:
 
         return PassManager.build_from_passlist([lower_func])
 
-    def build_lower_pipeline(self, input: Input) -> PassManager:
+    def build_lower_pipeline(
+        self, input: Input, additional_input: Optional[Input] = None
+    ) -> PassManager:
         self._input = input
+        self._additional_input = additional_input
         passes = []
 
         passes.append(self._const_fold_pass())

py/torch_tensorrt/fx/passes/pass_utils.py

@@ -41,10 +41,13 @@ def _validate_inference(pass_: PassFunc) -> PassFunc:
 
         @wraps(pass_)
         def pass_with_validation(
-            module: fx.GraphModule, input: Input
+            module: fx.GraphModule,
+            input: Input,
+            *args,
+            **kwargs,
         ) -> fx.GraphModule:
             res0 = module(*input)
-            processed_module = pass_(module, input)
+            processed_module = pass_(module, input, *args, **kwargs)
             res1 = processed_module(*input)
 
             tensor_res_0 = _collect_tensors(res0)

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

@@ -2,7 +2,7 @@
 import torch_tensorrt.fx.tracer.acc_tracer.acc_ops as acc_ops
 from parameterized import parameterized
 from torch.testing._internal.common_utils import run_tests
-from torch_tensorrt.fx.tools.common_fx2trt import AccTestCase
+from torch_tensorrt.fx.tools.common_fx2trt import AccTestCase, InputTensorSpec
 
 # NOTE torch.prod will only accept one dim unlike other reduce ops which accept tuples
 
@@ -93,6 +93,26 @@ def forward(self, x):
             test_implicit_batch_dim=False,
         )
 
+    def test_prod_all_dims_with_dynamic_shape(
+        self,
+        op=torch.prod,
+    ):
+        class Prod(torch.nn.Module):
+            def forward(self, x):
+                return op(x)
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 1), (2, 3, 4, 5), (2, 3, 10, 10))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(
+            Prod(), input_specs, expected_ops={acc_ops.prod}
+        )
+
 
 if __name__ == "__main__":
     run_tests()

py/torch_tensorrt/fx/test/core/test_input_tensor_spec.py

@@ -4,7 +4,7 @@
 
 import torch
 from torch.testing._internal.common_utils import run_tests, TestCase
-from torch_tensorrt.fx import InputTensorSpec
+from torch_tensorrt.fx import generate_input_specs, InputTensorSpec, LowerSetting
 
 
 class TestTRTModule(TestCase):
@@ -47,6 +47,47 @@ def test_from_tensors_with_dynamic_batch_size(self):
                 self.assertEqual(batch_size, shape[0])
                 self.assertSequenceEqual(tensor.shape[1:], shape[1:])
 
+    def test_from_tensors_with_dynamic_batch_size_different_batch_dims(self):
+        tensors = [torch.randn(1, 2, 3), torch.randn(2, 1, 4)]
+        batch_size_range = [2, 3, 4]
+        specs = InputTensorSpec.from_tensors_with_dynamic_batch_size(
+            tensors, batch_size_range, batch_dims=[0, 1]
+        )
+        for i, spec_and_tensor in enumerate(zip(specs, tensors)):
+            spec, tensor = spec_and_tensor
+            self._validate_spec(spec, tensor, dynamic_dims=[i])
+
+            for batch_size, shape in zip(batch_size_range, spec.shape_ranges[0]):
+                self.assertEqual(batch_size, shape[i])
+                tensor_shape = list(tensor.shape)
+                tensor_shape[i] = batch_size
+                self.assertSequenceEqual(tensor_shape, shape)
+
+    def test_generate_input_specs(self):
+        lower_setting = LowerSetting(
+            explicit_batch_dimension=False, max_batch_size=256, opt_profile_replica=2
+        )
+
+        # Implicit batch dim.
+        inputs = [torch.randn(1, 2, 3)]
+        specs = generate_input_specs(inputs, lower_setting)
+        for spec, tensor in zip(specs, inputs):
+            self._validate_spec(spec, tensor)
+
+        # Explicit batch dim without additional inputs.
+        lower_setting.explicit_batch_dimension = True
+        specs = generate_input_specs(inputs, lower_setting)
+        for spec, tensor in zip(specs, inputs):
+            self._validate_spec(spec, tensor, dynamic_dims=[0])
+            self.assertEqual(len(spec.shape_ranges), lower_setting.opt_profile_replica)
+
+        # Explicit batch dim with additional inputs.
+        additional_inputs = [torch.randn(1, 1, 3)]
+        specs = generate_input_specs(inputs, lower_setting, additional_inputs)
+        for spec, tensor in zip(specs, inputs):
+            self._validate_spec(spec, tensor, dynamic_dims=[1])
+            self.assertEqual(len(spec.shape_ranges), lower_setting.opt_profile_replica)
+
 
 if __name__ == "__main__":
     run_tests()