feat: Improve Dynamo partitioning System Performance on Large Models (pytorch#2175)

Author: gs-olive

py/torch_tensorrt/dynamo/_defaults.py

@@ -10,3 +10,4 @@
 OPTIMIZATION_LEVEL = None
 TRUNCATE_LONG_AND_DOUBLE = False
 USE_PYTHON_RUNTIME = False
+USE_FAST_PARTITIONER = True

py/torch_tensorrt/dynamo/_settings.py

@@ -10,6 +10,7 @@
     PASS_THROUGH_BUILD_FAILURES,
     PRECISION,
     TRUNCATE_LONG_AND_DOUBLE,
+    USE_FAST_PARTITIONER,
     USE_PYTHON_RUNTIME,
     VERSION_COMPATIBLE,
     WORKSPACE_SIZE,
@@ -29,3 +30,4 @@ class CompilationSettings:
     optimization_level: Optional[int] = OPTIMIZATION_LEVEL
     use_python_runtime: Optional[bool] = USE_PYTHON_RUNTIME
     truncate_long_and_double: bool = TRUNCATE_LONG_AND_DOUBLE
+    use_fast_partitioner: bool = USE_FAST_PARTITIONER

py/torch_tensorrt/dynamo/backend/backends.py

@@ -7,13 +7,12 @@
 import torch
 import torch._dynamo as td
 from torch._functorch.aot_autograd import aot_module_simplified, make_boxed_compiler
-from torch_tensorrt.dynamo import CompilationSettings
+from torch_tensorrt.dynamo import CompilationSettings, partitioning
 from torch_tensorrt.dynamo.conversion import (
     convert_module,
     repair_long_or_double_inputs,
 )
 from torch_tensorrt.dynamo.lowering._decompositions import get_decompositions
-from torch_tensorrt.dynamo.lowering._partition import get_submod_inputs, partition
 from torch_tensorrt.dynamo.lowering._pre_aot_lowering import pre_aot_substitutions
 from torch_tensorrt.dynamo.utils import parse_dynamo_kwargs
 
@@ -111,24 +110,68 @@ def _compile_module(
     Returns:
         Compiled FX GraphModule
     """
-    # Partition module into components that can be TRT-accelerated
-    partitioned_module = partition(
-        gm,
-        verbose=settings.debug,
-        min_block_size=settings.min_block_size,
-        torch_executed_ops=settings.torch_executed_ops,
+    # Check the number of supported operations in the graph
+    num_supported_ops, total_ops = partitioning.get_graph_converter_support(
+        gm, settings.debug, settings.torch_executed_ops
     )
 
+    # If the number of supported operations is 0 or less than the block size, skip the subgraph
+    # TODO: Add condition to second expression below when require_full_compilation is added
+    if num_supported_ops == 0 or (num_supported_ops < settings.min_block_size):
+        logger.warning(
+            f"{num_supported_ops} supported operations detected in subgraph containing {total_ops} computational nodes. "
+            f"Skipping this subgraph, since min_block_size was detected to be {settings.min_block_size}"
+        )
+        return gm
+    else:
+        logger.debug(
+            f"Detected support for {num_supported_ops} operators out of {total_ops} in subgraph."
+        )
+
+    # Partition module into components that can be TRT-accelerated
+    fast_partitioner_failed = False
+
+    # If specified, try using the fast partitioner and fall back to the global one on failure
+    if settings.use_fast_partitioner:
+        try:
+            partitioned_module = partitioning.fast_partition(
+                gm,
+                verbose=settings.debug,
+                min_block_size=settings.min_block_size,
+                torch_executed_ops=settings.torch_executed_ops,
+            )
+        except torch.fx.passes.splitter_base.FxNetSplitterInternalError:
+            logger.error(
+                "Partitioning failed on the subgraph with fast partition. See trace above. "
+                + "Retrying with global partition.",
+                exc_info=True,
+            )
+
+            fast_partitioner_failed = True
+            settings.use_fast_partitioner = False
+
+    if not settings.use_fast_partitioner:
+        partitioned_module = partitioning.global_partition(
+            gm,
+            verbose=settings.debug,
+            min_block_size=settings.min_block_size,
+            torch_executed_ops=settings.torch_executed_ops,
+        )
+
     # Store TRT replicas of Torch subgraphs
     trt_modules = {}
 
     # Iterate over all components that can be accelerated
     # Generate the corresponding TRT Module for those
     for name, _ in partitioned_module.named_children():
+        # Criteria for a module to be convertible to TRT
+        if settings.use_fast_partitioner and "_run_on_acc" not in name:
+            continue
+
         submodule = getattr(partitioned_module, name)
 
         # Get submodule inputs
-        submodule_inputs = get_submod_inputs(
+        submodule_inputs = partitioning.get_submod_inputs(
             partitioned_module, submodule, sample_inputs
         )
 
@@ -153,4 +196,8 @@ def _compile_module(
     for name, trt_mod in trt_modules.items():
         setattr(partitioned_module, name, trt_mod)
 
+    # Reset settings object to user specification after fallback to global partitioning mode
+    if fast_partitioner_failed:
+        settings.use_fast_partitioner = True
+
     return partitioned_module

py/torch_tensorrt/dynamo/compile.py

@@ -7,7 +7,6 @@
 import torch
 import torch_tensorrt
 from torch.fx.passes.pass_manager import PassManager
-from torch.fx.passes.splitter_base import SplitResult
 from torch_tensorrt._Device import Device
 from torch_tensorrt._enums import (  # TODO: Should probabably be the TRT EngineCapability Enum
     EngineCapability,
@@ -21,18 +20,17 @@
     PASS_THROUGH_BUILD_FAILURES,
     PRECISION,
     TRUNCATE_LONG_AND_DOUBLE,
+    USE_FAST_PARTITIONER,
     USE_PYTHON_RUNTIME,
     VERSION_COMPATIBLE,
     WORKSPACE_SIZE,
 )
 from torch_tensorrt.dynamo.backend.backends import _compile_module
-from torch_tensorrt.dynamo.conversion import convert_module
 from torch_tensorrt.dynamo.lowering._fusers import (
     fuse_permute_linear,
     fuse_permute_matmul,
 )
 from torch_tensorrt.dynamo.utils import prepare_device, prepare_inputs
-from torch_tensorrt.fx.tools.trt_splitter import TRTSplitter, TRTSplitterSetting
 
 logger = logging.getLogger(__name__)
 
@@ -64,6 +62,7 @@ def compile(
     version_compatible: bool = VERSION_COMPATIBLE,
     optimization_level: Optional[int] = OPTIMIZATION_LEVEL,
     use_python_runtime: bool = USE_PYTHON_RUNTIME,
+    use_fast_partitioner: bool = USE_FAST_PARTITIONER,
     **kwargs: Any,
 ) -> torch.fx.GraphModule:
     if debug:
@@ -75,7 +74,7 @@ def compile(
         "The Dynamo backend is an experimental feature, for which only the "
         + "following arguments are supported: "
         + "{enabled_precisions, debug, workspace_size, min_block_size, "
-        + "torch_executed_ops, pass_through_build_failures}"
+        + "torch_executed_ops, pass_through_build_failures, use_fast_partitioner}"
     )
 
     if not isinstance(inputs, collections.abc.Sequence):
@@ -115,55 +114,12 @@ def compile(
         "optimization_level": optimization_level,
         "use_python_runtime": use_python_runtime,
         "truncate_long_and_double": truncate_long_and_double,
+        "use_fast_partitioner": use_fast_partitioner,
     }
 
     settings = CompilationSettings(**compilation_options)
-    if kwargs.get("use_capability_partitioner", None):
-        model = lower_model(gm, torch_inputs)
-        return _compile_module(model, torch_inputs, settings)
-    else:
-        split_result = lower_model_using_trt_splitter(gm, torch_inputs)
-        trt_module = _compile_graph(split_result, torch_inputs, settings)
-
-        return trt_module
 
-
-def _compile_graph(
-    split_result: SplitResult,
-    inputs: Any,
-    settings: CompilationSettings = CompilationSettings(),
-    **kwargs: Any,
-) -> torch.fx.GraphModule:
-    for submod_name, submod_inputs in split_result.submodule_inputs.items():
-        submod = getattr(split_result.split_module, submod_name)
-        # Only acc submodules will be lowered.
-        if not submod_name.startswith(split_result.non_acc_submodule_prefix):
-            # Create TRT Module from submodule
-            trt_mod = convert_module(
-                submod,
-                submod_inputs,
-                settings=settings,
-                name=submod_name,
-            )
-            setattr(split_result.split_module, submod_name, trt_mod)
-
-    return split_result.split_module
-
-
-def lower_model_using_trt_splitter(
-    model: torch.nn.Module, inputs: Any, **kwargs: Any
-) -> SplitResult:
-    # Perform basic lowering
-    model = lower_model(model, inputs)
-    splitter_setting = TRTSplitterSetting()
-    splitter_setting.use_implicit_batch_dim = False
-    splitter_setting.min_acc_module_size = 1
-    splitter_setting.use_experimental_rt = False
-    splitter = TRTSplitter(model, inputs, settings=splitter_setting)
-    splitter.node_support_preview()
-    split_result = splitter.generate_split_results()
-
-    return split_result
+    return _compile_module(gm, torch_inputs, settings)
 
 
 def lower_model(

py/torch_tensorrt/dynamo/conversion/converter_registry.py

@@ -349,8 +349,8 @@ def unique_targets(self) -> Set[Target]:
         """Returns the set of unique converter targets stored across all registries"""
         return set.union(*[set(registry.keys()) for registry in self.registries])
 
-    # TODO: Make this a static method since it does not need state
-    def qualified_name_or_str(self, target: Target) -> str:
+    @staticmethod
+    def qualified_name_or_str(target: Target) -> str:
         """Returns string representation of an FX Node target"""
         if isinstance(target, str):
             return target

py/torch_tensorrt/dynamo/lowering/__init__.py

@@ -1,10 +1,5 @@
 from ._decompositions import get_decompositions  # noqa: F401
-from ._fusers import *  # noqa: F403
-from ._partition import (  # noqa: F401
-    DEFAULT_SINGLE_NODE_PARTITIONS,
-    get_submod_inputs,
-    partition,
-)
+from ._fusers import *  # noqa: F401
 from ._pre_aot_lowering import SUBSTITUTION_REGISTRY  # noqa: F401
 from ._pre_aot_lowering import register_substitution  # noqa: F401
-from .substitutions import *  # noqa: F403
+from .substitutions import *  # noqa: F401

py/torch_tensorrt/dynamo/partitioning/__init__.py

@@ -0,0 +1,3 @@
+from ._adjacency_partitioner import partition as fast_partition
+from ._global_partitioner import partition as global_partition
+from .common import get_graph_converter_support, get_submod_inputs