feat: Prototype Module-Acceleration in Dynamo

- Add support for excluding entire Torch modules from tracing in Dynamo
using Torch custom operators
- Develop new dataclass to store required replacement functions and
operators in a streamlined way
- Add new registry to store mapping between replacement operators and
their corresponding dataclass
- Add documentation for easy additions of new module-level exclusion
operators
- Add robust testing and address recent review comments

Author: gs-olive

py/torch_tensorrt/__init__.py

@@ -94,7 +94,7 @@ def _find_lib(name, paths):
 
 from torch_tensorrt import fx
 
-if version.parse(torch.__version__) >= version.parse("2.dev"):
+if version.parse(torch.__version__) >= version.parse("2.1.dev"):
     from torch_tensorrt import dynamo
     from torch_tensorrt.dynamo import backend
 

py/torch_tensorrt/dynamo/backend/backends.py

@@ -8,6 +8,9 @@
 from torch_tensorrt.dynamo.backend.lowering._decompositions import (
     get_decompositions,
 )
+from torch_tensorrt.dynamo.backend.lowering._pre_aot_lowering import (
+    pre_aot_module_replacement,
+)
 from torch_tensorrt.dynamo.backend.lowering._partition import (
     partition,
     get_submod_inputs,
@@ -45,6 +48,13 @@ def aot_torch_tensorrt_aten_backend(
         settings=settings,
     )
 
+    logger.debug("Pre-module replacement graph:\n" + str(gm.graph))
+
+    # Perform Pre-AOT Lowering for Module-Level Replacement
+    gm = pre_aot_module_replacement(gm)
+
+    logger.debug("Post-module replacement graph:\n" + str(gm.graph))
+
     # Invoke AOTAutograd to translate operators to aten
     return aot_module_simplified(
         gm,
@@ -70,6 +80,8 @@ def _pretraced_backend(
         Compiled FX GraphModule
     """
     try:
+        logger.debug("Post-AOT Autograd graph:\n" + str(gm.graph))
+
         trt_compiled = _compile_module(
             gm,
             sample_inputs,

py/torch_tensorrt/dynamo/backend/lowering/__init__.py

@@ -1,7 +1,9 @@
-from torch_tensorrt.dynamo.backend.lowering._decompositions import (
+from ._decompositions import (
     get_decompositions,
 )
-from torch_tensorrt.dynamo.backend.lowering._partition import (
-    partition,
-    get_submod_inputs,
+from ._pre_aot_lowering import (
+    MODULE_SUBSTITUTION_REGISTRY,
+    module_substitution,
 )
+from ._partition import partition, get_submod_inputs, DEFAULT_SINGLE_NODE_PARTITIONS
+from .module_substitutions import *

py/torch_tensorrt/dynamo/backend/lowering/_partition.py

@@ -1,9 +1,10 @@
 import logging
-from typing import Dict, List, Optional, Sequence
+from typing import Dict, List, Optional, Sequence, Set
 
 import torch
 
 from torch_tensorrt.dynamo.backend._defaults import MIN_BLOCK_SIZE
+from torch_tensorrt.dynamo.backend.lowering import MODULE_SUBSTITUTION_REGISTRY
 from torch.fx.passes.infra.partitioner import CapabilityBasedPartitioner, Partition
 from torch.fx.graph_module import GraphModule
 from torch.fx.node import _get_qualified_name
@@ -14,6 +15,11 @@
 
 logger = logging.getLogger(__name__)
 
+DEFAULT_SINGLE_NODE_PARTITIONS: Set[str] = set(
+    _get_qualified_name(module.new_operator)
+    for module in MODULE_SUBSTITUTION_REGISTRY.values()
+)
+
 
 class TRTPartitioner(CapabilityBasedPartitioner):
     """Partitioner to split an FX graph into subgraphs based on operator support
@@ -35,7 +41,9 @@ def __init__(
         operator_support: OperatorSupport,
         *,
         non_compute_ops: Optional[Sequence[str]] = None,
-        allowed_single_node_partition_ops: Optional[Sequence[str]] = None,
+        allowed_single_node_partition_ops: Optional[
+            Sequence[str]
+        ] = DEFAULT_SINGLE_NODE_PARTITIONS,
         min_block_size=MIN_BLOCK_SIZE,
     ) -> None:
         super().__init__(

py/torch_tensorrt/dynamo/backend/lowering/_pre_aot_lowering.py

@@ -0,0 +1,121 @@
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, Type
+import torch
+import logging
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class ModuleReplacement:
+    """Class to store key functionality for module replacement"""
+
+    # torch.ops.___ name for replacement function for module
+    new_operator: torch._ops.OpOverload
+
+    # Function taking a containing graph, a submodule, and a 'call_module' node and returning
+    # a replacement node, with type 'call_function', or raising an Error if incompatibility is detected
+    # Note: subgraph_insertion_fn should NOT delete nodes or recompile the graph
+    subgraph_insertion_fn: Callable[
+        [torch.fx.GraphModule, torch.nn.Module, torch.fx.Node], torch.fx.Node
+    ]
+
+
+# Dictionary mapping module to ModuleReplacement instance
+MODULE_SUBSTITUTION_REGISTRY: Dict[Type[torch.nn.Module], ModuleReplacement] = dict()
+
+
+def module_substitution(
+    module_to_replace: Type[torch.nn.Module],
+    new_operator: torch._ops.OpOverload,
+    enabled: bool = True,
+) -> Callable[[Any], Any]:
+    """Decorator to register subgraph insertion functions
+
+    Args:
+        module_to_replace: nn.Module to replace
+        new_operator: Custom torch operator to replace with
+        enabled: Whether the substitution is enabled or disabled
+    Returns:
+        torch.fx.GraphModule
+    """
+
+    def register_substitution(subgraph_insertion_fn):
+        """Function for use if substitution is enabled"""
+        module_replacement = ModuleReplacement(
+            new_operator=new_operator, subgraph_insertion_fn=subgraph_insertion_fn
+        )
+        MODULE_SUBSTITUTION_REGISTRY[module_to_replace] = module_replacement
+        return subgraph_insertion_fn
+
+    def disable_substitution(subgraph_insertion_fn):
+        """Function for use if substitution is disabled"""
+        return subgraph_insertion_fn
+
+    return register_substitution if enabled else disable_substitution
+
+
+def pre_aot_module_replacement(gm: torch.fx.GraphModule):
+    """Perform module-level graph replacement prior to AOT tracing
+
+    Args:
+        gm: FX GraphModule to perform module replacement on
+    Returns:
+        torch.fx.GraphModule
+
+    """
+    # Ensure all parameters are in inference mode
+    for param in gm.parameters():
+        param.requires_grad = False
+
+    # Iterate over graph nodes, extracting module calls, to check for interceptions
+    for n in gm.graph.nodes:
+        if n.op == "call_module":
+            # Extract submodule from graph
+            submodule = gm.get_submodule(n.target)
+
+            # If submodule is a member of the substitution registry, replace it
+            if type(submodule) in MODULE_SUBSTITUTION_REGISTRY:
+
+                try:
+                    replacement = MODULE_SUBSTITUTION_REGISTRY[type(submodule)]
+                    op, insertion_fn = (
+                        replacement.new_operator,
+                        replacement.subgraph_insertion_fn,
+                    )
+                    logger.debug(
+                        f"Replacing module of type {type(submodule)} with {op}"
+                    )
+
+                    # Insert new node prior to older node
+                    with gm.graph.inserting_before(n):
+                        new_node = insertion_fn(gm, submodule, n)
+
+                    # If submodule is not a native torch.nn module, it must be manually excluded
+                    # from Dynamo tracing
+                    if not type(submodule).__module__.startswith("torch.nn"):
+                        torch._dynamo.allowed_functions._allowed_function_ids.add(
+                            id(type(submodule))
+                        )
+
+                    # Replace all original node uses and clean up graph
+                    n.replace_all_uses_with(new_node)
+                    gm.graph.eliminate_dead_code()
+                    gm.graph.lint()
+                    gm.recompile()
+
+                # A module replacement can fail in the event that the specific instance of the submodule cannot
+                # be replaced
+                except Exception:
+                    logger.debug(
+                        f"Encountered error while replacing {type(submodule)}",
+                        exc_info=True,
+                    )
+                    continue
+
+    # Perform cleanup and recompilation before returning module
+    gm.graph.eliminate_dead_code()
+    gm.graph.lint()
+    gm.recompile()
+    return gm

py/torch_tensorrt/dynamo/backend/lowering/module_substitutions/__init__.py

@@ -0,0 +1 @@
+from .maxpool1d import *

py/torch_tensorrt/dynamo/backend/lowering/module_substitutions/maxpool1d.py

@@ -0,0 +1,127 @@
+from typing import Dict, Tuple
+import torch
+from torch._custom_op.impl import custom_op
+from torch.fx.node import Argument, Target
+
+from torch_tensorrt.fx.converter_registry import tensorrt_converter
+from torch_tensorrt.fx.converters import acc_ops_converters
+from torch_tensorrt.fx.types import TRTNetwork, TRTTensor
+
+from torch_tensorrt.dynamo.backend.lowering import module_substitution
+
+
+# This file serves as an example and a tutorial for excluding custom modules from
+# torch.compile tracing. Each required step is labeled with a number indicating the
+# preferable implementation order.
+
+
+# 1. The Placeholder
+#
+# Specify the schema and namespace of the operator, as well as a placeholder function
+# representing the schema. The schema should be in torch JIT syntax, indicating input and output
+# types. The namespace, such as tensorrt, will cause the op to be registered as torch.ops.tensorrt.your_op
+# Then, create a placeholder function with no operations, but having the same schema and naming as that
+# used in the decorator
+@custom_op(
+    qualname="tensorrt::maxpool1d",
+    manual_schema="(Tensor x, int[1] kernel_size, int[1] stride, int[1] padding, int[1] dilation, bool ceil_mode) -> Tensor",
+)
+def maxpool1d(x, kernel_size, stride, padding, dilation, ceil_mode):
+    # Defines operator schema, name, namespace, and function header
+    ...
+
+
+# 2. The Generic Implementation
+#
+# Define the default implementation of the operator in torch syntax. This is used for autograd
+# and other tracing functionality. Generally, the torch.nn.functional analog of the operator to replace
+# is desirable. If the operator to replace is a custom module you've written, then add its Torch
+# implementation here. Note that the function header to the generic function can have specific arguments
+# as in the above placeholder
+@maxpool1d.impl("cpu")
+@maxpool1d.impl("cuda")
+@maxpool1d.impl_abstract()
+def maxpool1d_generic(
+    *args,
+    **kwargs,
+):
+    # Defines an implementation for AOT Autograd to use for shape analysis/propagation
+    return torch.nn.functional.max_pool1d(
+        *args,
+        **kwargs,
+    )
+
+
+# 3. The Module Substitution Function
+#
+# Define a function which can intercept a node of the kind to be replaced, extract
+# the relevant data from that node/submodule, and then re-package the information
+# for use by an accelerated implementation (to be implemented in step 4). This function
+# should use the operator defined in step 1 (for example torch.ops.tensorrt.maxpool1d).
+# It should refactor the args and kwargs as is needed by the accelerated implementation.
+#
+# If the submodule has weights or other Tensor fields which the accelerated implementation
+# needs, the function should insert the necessary nodes to access those weights. For example,
+# if the weight Tensor of a submodule is needed, one could write:
+#
+#       weights = gm.graph.get_attr(n.target + ".weight", torch.Tensor)
+#       bias = gm.graph.get_attr(n.target + ".bias", torch.Tensor)
+#       ...
+#       kwargs={"weight": weights,
+#               "bias": bias,
+#               ...
+#
+@module_substitution(torch.nn.MaxPool1d, torch.ops.tensorrt.maxpool1d)
+def maxpool1d_insertion_fn(
+    gm: torch.fx.GraphModule, submodule: torch.nn.Module, node: torch.fx.Node
+) -> torch.fx.Node:
+    # Defines insertion function for new node
+    new_node = gm.graph.call_function(
+        torch.ops.tensorrt.maxpool1d,
+        args=node.args,
+        kwargs={
+            "kernel_size": submodule.kernel_size,
+            "stride": submodule.stride,
+            "padding": submodule.padding,
+            "dilation": submodule.dilation,
+            "ceil_mode": submodule.ceil_mode,
+        },
+    )
+
+    return new_node
+
+
+# 4. The Accelerated Implementation
+#
+# Define an accelerated implementation of the operator, and register it as necessary.
+# This accelerated implementation should consume the args/kwargs specified in step 3.
+# One should expect that torch.compile will compress all kwargs into the args field in
+# the order specified in the schema written in step 1.
+@tensorrt_converter(torch.ops.tensorrt.maxpool1d.default)
+def tensorrt_maxpool1d(
+    network: TRTNetwork,
+    target: Target,
+    args: Tuple[Argument, ...],
+    kwargs: Dict[str, Argument],
+    name: str,
+) -> TRTTensor:
+    # Defines converter replacing the default operator for this function
+    kwargs_new = {
+        "input": args[0],
+        "kernel_size": args[1],
+        "stride": args[2],
+        "padding": args[3],
+        "dilation": args[4],
+        "ceil_mode": False if len(args) < 6 else args[5],
+    }
+
+    return acc_ops_converters.acc_ops_max_pool1d(
+        network, target, None, kwargs_new, name
+    )
+
+
+# 5. Add Imports
+#
+# Add your accelerated module file to the __init__.py in this directory, to ensure
+# all registrations are run. For instance, if the new module file is called new_mod.py,
+# one should add `from .new_mod import *` to the __init__.py

py/torch_tensorrt/dynamo/backend/test/test_pre_aot_lowering.py

@@ -0,0 +1,55 @@
+import torch
+from utils import lower_graph_testing
+from torch.testing._internal.common_utils import run_tests, TestCase
+from torch_tensorrt.dynamo import compile
+
+
+class TestMaxPool1D(TestCase):
+    def test_pre_aot_lowering_maxpool1d(self):
+        class MaxPool1D(torch.nn.Module):
+            def __init__(self, *args, **kwargs) -> None:
+                super().__init__(*args, **kwargs)
+                self.maxpool = torch.nn.MaxPool1d(2)
+
+            def forward(self, x):
+                return self.maxpool(x)
+
+        # Operations expected to be included in the traced graph after decompositions
+        expected_ops = {torch.ops.tensorrt.maxpool1d.default}
+
+        inputs = [
+            torch.rand(
+                9,
+                16,
+                2,
+            ).cuda(),
+        ]
+
+        fx_graph = torch.fx.symbolic_trace(MaxPool1D())
+        _, expected_ops_unseen = lower_graph_testing(
+            fx_graph, inputs, expected_ops=expected_ops, min_block_size=1
+        )
+
+        self.assertEquals(
+            len(expected_ops_unseen),
+            0,
+            f"The following expected ops were not encountered: {expected_ops_unseen}",
+        )
+
+        torch._dynamo.reset()
+
+        # Validate that the results between Torch and Torch-TRT are similar
+        optimized_model = compile(
+            fx_graph, inputs, min_block_size=1, pass_through_build_failures=True
+        )
+        optimized_model_results = optimized_model(*inputs).detach().cpu()
+        torch_model_results = fx_graph(*inputs).detach().cpu()
+
+        max_diff = torch.max(torch.abs(optimized_model_results - torch_model_results))
+        self.assertAlmostEqual(
+            max_diff, 0, f"Maxpool1d TRT outputs don't match with the original model."
+        )
+
+
+if __name__ == "__main__":
+    run_tests()