Merge pull request #1979 from pytorch/dynamo_module_level_acceleration

feat: Module-Acceleration in Dynamo [5 / x]

Author: gs-olive

py/torch_tensorrt/__init__.py

@@ -95,7 +95,7 @@ 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_substitutions,
+)
 from torch_tensorrt.dynamo.backend.lowering._partition import (
     partition,
     get_submod_inputs,
@@ -41,6 +44,9 @@ def aot_torch_tensorrt_aten_backend(
         settings=settings,
     )
 
+    # Perform Pre-AOT Lowering for Module-Level Replacement
+    gm = pre_aot_substitutions(gm)
+
     # Invoke AOTAutograd to translate operators to aten
     return aot_module_simplified(
         gm,
@@ -65,6 +71,8 @@ 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 (
+    SUBSTITUTION_REGISTRY,
+    register_substitution,
 )
+from ._partition import partition, get_submod_inputs, DEFAULT_SINGLE_NODE_PARTITIONS
+from .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 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(to_replace.new_operator)
+    for to_replace in 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 +1,139 @@
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, Optional, Type, Union
+import torch
+import logging
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class Substitution:
+    """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 node, and optionally a submodule (if replacing a module)
+    # 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.fx.Node, Optional[torch.nn.Module]], torch.fx.Node
+    ]
+
+
+# Dictionary mapping module to Substitution instance
+SUBSTITUTION_REGISTRY: Dict[
+    Union[Type[torch.nn.Module], Callable], Substitution
+] = dict()
+
+
+def register_substitution(
+    module_or_function_to_replace: Union[Type[torch.nn.Module], Callable],
+    new_operator: torch._ops.OpOverload,
+    enabled: bool = True,
+) -> Callable[[Any], Any]:
+    """Decorator to register subgraph insertion functions
+
+    Args:
+        module_or_function_to_replace: nn.Module or node target Callable to replace
+        new_operator: Custom torch operator to replace with
+        enabled: Whether the substitution is enabled or disabled
+    Returns:
+        torch.fx.GraphModule
+    """
+
+    def enable_substitution(subgraph_insertion_fn):
+        """Function for use if substitution is enabled"""
+        replacement = Substitution(
+            new_operator=new_operator, subgraph_insertion_fn=subgraph_insertion_fn
+        )
+        SUBSTITUTION_REGISTRY[module_or_function_to_replace] = replacement
+        return subgraph_insertion_fn
+
+    def disable_substitution(subgraph_insertion_fn):
+        """Function for use if substitution is disabled"""
+        return subgraph_insertion_fn
+
+    return enable_substitution if enabled else disable_substitution
+
+
+def pre_aot_substitutions(gm: torch.fx.GraphModule):
+    """Perform graph substitutions prior to AOT tracing
+
+    Args:
+        gm: FX GraphModule to perform substitution on
+    Returns:
+        torch.fx.GraphModule
+
+    """
+    logger.debug("Pre-module replacement graph:\n" + str(gm.graph))
+
+    # 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:
+        exists_in_registry = False
+        to_replace = None
+
+        if n.op == "call_module":
+            # Extract submodule from graph, validate in registry
+            submodule = gm.get_submodule(n.target)
+            to_replace = type(submodule)
+            exists_in_registry = to_replace in SUBSTITUTION_REGISTRY
+        elif n.op == "call_function":
+            # Extract function from graph, validate in registry
+            to_replace = n.target
+            exists_in_registry = n.target in SUBSTITUTION_REGISTRY
+
+        # If submodule/function is a member of the substitution registry, replace it
+        if exists_in_registry:
+            try:
+                replacement = SUBSTITUTION_REGISTRY[to_replace]
+                op, insertion_fn = (
+                    replacement.new_operator,
+                    replacement.subgraph_insertion_fn,
+                )
+                logger.debug(f"Replacing node of type {to_replace} with {op}")
+
+                # Insert new node prior to older node
+                with gm.graph.inserting_before(n):
+                    new_node = insertion_fn(
+                        gm, n, submodule if n.op == "call_module" else None
+                    )
+
+                # If submodule is not a native torch.nn module, it must be manually excluded
+                # from Dynamo tracing
+                if n.op == "call_module" and not type(submodule).__module__.startswith(
+                    "torch.nn"
+                ):
+                    torch._dynamo.allowed_functions._allowed_function_ids.add(
+                        id(to_replace)
+                    )
+
+                # 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 replacement can fail in the event that the specific instance of the submodule/function
+            # cannot be replaced
+            except Exception:
+                logger.debug(
+                    f"Encountered error while replacing {to_replace}",
+                    exc_info=True,
+                )
+                continue
+
+    # Perform cleanup and recompilation before returning module
+    gm.graph.eliminate_dead_code()
+    gm.graph.lint()
+    gm.recompile()
+
+    logger.debug("Post-module replacement graph:\n" + str(gm.graph))
+
+    return gm

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

@@ -0,0 +1,2 @@
+from .maxpool1d import *
+from .einsum import *

py/torch_tensorrt/dynamo/backend/lowering/substitutions/einsum.py

@@ -0,0 +1,80 @@
+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.converter_utils import set_layer_name
+from torch_tensorrt.fx.types import TRTNetwork, TRTTensor
+
+from torch_tensorrt.dynamo.backend.lowering import register_substitution
+
+
+@custom_op(
+    qualname="tensorrt::einsum",
+    manual_schema="(str equation, Tensor[] tensors) -> Tensor",
+)
+def einsum(equation, tensors):
+    # Defines operator schema, name, namespace, and function header
+    ...
+
+
+@einsum.impl("cpu")
+@einsum.impl("cuda")
+@einsum.impl_abstract()
+def einsum_generic(
+    *args,
+    **kwargs,
+):
+    # Defines a converter implementation for AOT Autograd to use for shape analysis/propagation
+    return torch.einsum(
+        *args,
+        **kwargs,
+    )
+
+
+@tensorrt_converter(torch.ops.tensorrt.einsum.default)
+def aten_ops_einsum(
+    network: TRTNetwork,
+    target: Target,
+    args: Tuple[Argument, ...],
+    kwargs: Dict[str, Argument],
+    name: str,
+) -> TRTTensor:
+    # Defines converter replacing the default operator for this function
+    for input_trt in args[1]:
+        if not isinstance(input_trt, TRTTensor):
+            raise RuntimeError(f"Einsum received non-TRTTensor input: {input_trt}")
+
+    einsum_layer = network.add_einsum(inputs=args[1], equation=args[0])
+
+    set_layer_name(einsum_layer, target, name)
+    return einsum_layer.get_output(0)
+
+
+@register_substitution(torch.einsum, torch.ops.tensorrt.einsum)
+def einsum_insertion_fn(
+    gm: torch.fx.GraphModule,
+    node: torch.fx.Node,
+    _unused: None = None,
+) -> torch.fx.Node:
+    equation = node.args[0]
+
+    # Ensure inputs is a list of (Tensor) arguments
+    if isinstance(node.args[1], (tuple, list)):
+        inputs = node.args[1]
+    else:
+        inputs = node.args[1:]
+
+    assert (
+        1 <= len(inputs) <= 2
+    ), f"TRT Einsum currently only supports 1 or 2 Tensors, got {len(inputs)} Tensors"
+
+    # Ensure the input is formatted as an equation and
+    new_node = gm.graph.call_function(
+        torch.ops.tensorrt.einsum,
+        args=(equation, inputs),
+        kwargs=node.kwargs,
+    )
+
+    return new_node