fix: Unify export/compile compilation utilities

- Update argument naming for compatibility with `fx_ts_compat` utilities
- Add support for new TRT 8.6 utilities, including auxiliary streams,
version compatibility, and optimization levels
- Add support for TRTModuleNext use during compilation with Dynamo
compile
- Improve documentation of features and version checking for TRT feature
compatibility
- Add test cases for new `TRTModuleNext` functionality and for TRT
custom options functionality

Author: gs-olive

examples/dynamo/dynamo_compile_advanced_usage.py

@@ -66,6 +66,8 @@ def forward(self, x: torch.Tensor, y: torch.Tensor):
     debug=True,
     min_block_size=2,
     torch_executed_ops={},
+    optimization_level=4,
+    use_experimental_rt=True,
 )
 
 # Run the model on an input to cause compilation, as so:

py/torch_tensorrt/dynamo/backend/__init__.py

@@ -4,7 +4,7 @@
 import torch_tensorrt
 from functools import partial
 
-from typing import Any, Sequence
+from typing import Any, Optional, Sequence
 from torch_tensorrt import EngineCapability, Device
 from torch_tensorrt.fx.utils import LowerPrecision
 
@@ -14,9 +14,13 @@
 from torch_tensorrt.dynamo.backend._defaults import (
     PRECISION,
     DEBUG,
-    MAX_WORKSPACE_SIZE,
+    WORKSPACE_SIZE,
     MIN_BLOCK_SIZE,
     PASS_THROUGH_BUILD_FAILURES,
+    MAX_AUX_STREAMS,
+    VERSION_COMPATIBLE,
+    OPTIMIZATION_LEVEL,
+    USE_EXPERIMENTAL_RT,
 )
 
 
@@ -35,7 +39,7 @@ def compile(
     debug=DEBUG,
     capability=EngineCapability.default,
     num_avg_timing_iters=1,
-    workspace_size=MAX_WORKSPACE_SIZE,
+    workspace_size=WORKSPACE_SIZE,
     dla_sram_size=1048576,
     dla_local_dram_size=1073741824,
     dla_global_dram_size=536870912,
@@ -45,6 +49,10 @@ def compile(
     min_block_size=MIN_BLOCK_SIZE,
     torch_executed_ops=[],
     torch_executed_modules=[],
+    max_aux_streams=MAX_AUX_STREAMS,
+    version_compatible=VERSION_COMPATIBLE,
+    optimization_level=OPTIMIZATION_LEVEL,
+    use_experimental_rt=USE_EXPERIMENTAL_RT,
     **kwargs,
 ):
     if debug:
@@ -86,6 +94,10 @@ def compile(
         workspace_size=workspace_size,
         min_block_size=min_block_size,
         torch_executed_ops=torch_executed_ops,
+        max_aux_streams=max_aux_streams,
+        version_compatible=version_compatible,
+        optimization_level=optimization_level,
+        use_experimental_rt=use_experimental_rt,
         **kwargs,
     )
 
@@ -105,19 +117,30 @@ def compile(
 def create_backend(
     precision: LowerPrecision = PRECISION,
     debug: bool = DEBUG,
-    workspace_size: int = MAX_WORKSPACE_SIZE,
+    workspace_size: int = WORKSPACE_SIZE,
     min_block_size: int = MIN_BLOCK_SIZE,
     torch_executed_ops: Sequence[str] = set(),
     pass_through_build_failures: bool = PASS_THROUGH_BUILD_FAILURES,
+    max_aux_streams: Optional[int] = MAX_AUX_STREAMS,
+    version_compatible: bool = VERSION_COMPATIBLE,
+    optimization_level: Optional[int] = OPTIMIZATION_LEVEL,
+    use_experimental_rt: bool = USE_EXPERIMENTAL_RT,
     **kwargs,
 ):
     """Create torch.compile backend given specified arguments
 
     Args:
         precision:
         debug: Whether to print out verbose debugging information
-        workspace_size: Maximum workspace TRT is allowed to use for the module
-        precision: Model Layer precision
+        workspace_size: Workspace TRT is allowed to use for the module (0 is default)
+        min_block_size: Minimum number of operators per TRT-Engine Block
+        torch_executed_ops: Sequence of operations to run in Torch, regardless of converter coverage
+        pass_through_build_failures: Whether to fail on TRT engine build errors (True) or not (False)
+        max_aux_streams: Maximum number of allowed auxiliary TRT streams for each engine
+        version_compatible: Provide version forward-compatibility for engine plan files
+        optimization_level: Builder optimization 0-5, higher levels imply longer build time,
+            searching for more optimization options. TRT defaults to 3
+        use_experimental_rt: Whether to use the new experimental TRTModuleNext for TRT engines
     Returns:
         Backend for torch.compile
     """
@@ -131,6 +154,10 @@ def create_backend(
         min_block_size=min_block_size,
         torch_executed_ops=torch_executed_ops,
         pass_through_build_failures=pass_through_build_failures,
+        max_aux_streams=max_aux_streams,
+        version_compatible=version_compatible,
+        optimization_level=optimization_level,
+        use_experimental_rt=use_experimental_rt,
     )
 
     return partial(

py/torch_tensorrt/dynamo/backend/_defaults.py

@@ -3,6 +3,10 @@
 
 PRECISION = LowerPrecision.FP32
 DEBUG = False
-MAX_WORKSPACE_SIZE = 20 << 30
+WORKSPACE_SIZE = 0
 MIN_BLOCK_SIZE = 5
 PASS_THROUGH_BUILD_FAILURES = False
+MAX_AUX_STREAMS = None
+VERSION_COMPATIBLE = False
+OPTIMIZATION_LEVEL = None
+USE_EXPERIMENTAL_RT = False

py/torch_tensorrt/dynamo/backend/_settings.py

@@ -1,21 +1,29 @@
 from dataclasses import dataclass, field
-from typing import Sequence
+from typing import Optional, Sequence
 
 from torch_tensorrt.fx.utils import LowerPrecision
 from torch_tensorrt.dynamo.backend._defaults import (
     PRECISION,
     DEBUG,
-    MAX_WORKSPACE_SIZE,
+    WORKSPACE_SIZE,
     MIN_BLOCK_SIZE,
     PASS_THROUGH_BUILD_FAILURES,
+    MAX_AUX_STREAMS,
+    VERSION_COMPATIBLE,
+    OPTIMIZATION_LEVEL,
+    USE_EXPERIMENTAL_RT,
 )
 
 
 @dataclass(frozen=True)
 class CompilationSettings:
     precision: LowerPrecision = PRECISION
     debug: bool = DEBUG
-    workspace_size: int = MAX_WORKSPACE_SIZE
+    workspace_size: int = WORKSPACE_SIZE
     min_block_size: int = MIN_BLOCK_SIZE
     torch_executed_ops: Sequence[str] = field(default_factory=set)
     pass_through_build_failures: bool = PASS_THROUGH_BUILD_FAILURES
+    max_aux_streams: Optional[int] = MAX_AUX_STREAMS
+    version_compatible: bool = VERSION_COMPATIBLE
+    optimization_level: Optional[int] = OPTIMIZATION_LEVEL
+    use_experimental_rt: bool = USE_EXPERIMENTAL_RT

py/torch_tensorrt/dynamo/backend/backends.py

@@ -135,6 +135,7 @@ def _compile_module(
             submodule,
             submodule_inputs,
             settings=settings,
+            name=name,
         )
 
         # Replace FX Module with TRT Module

py/torch_tensorrt/dynamo/backend/conversion.py

@@ -1,9 +1,10 @@
 from typing import Sequence, Union
 import torch
+import io
 from torch_tensorrt.fx.trt_module import TRTModule
 from torch_tensorrt import TRTModuleNext
 from torch_tensorrt.dynamo.backend._settings import CompilationSettings
-from torch_tensorrt.fx.fx2trt import (
+from torch_tensorrt.dynamo.fx_ts_compat.fx2trt import (
     InputTensorSpec,
     TRTInterpreter,
 )
@@ -15,30 +16,50 @@ def convert_module(
     module: torch.fx.GraphModule,
     inputs: Sequence[torch.Tensor],
     settings: CompilationSettings = CompilationSettings(),
+    name: str = "",
 ) -> Union[TRTModuleNext, TRTModule]:
     """Convert an FX module to a TRT module
     Args:
         module: FX GraphModule to convert
         inputs: Sequence of Tensors representing inputs to the module
         settings: Compilation settings
+        name: TRT engine name
     Returns:
         TRTModule or TRTModuleNext
     """
-    interp = TRTInterpreter(
+    interpreter = TRTInterpreter(
         module,
         InputTensorSpec.from_tensors(inputs),
         explicit_batch_dimension=True,
         logger_level=(trt.Logger.VERBOSE if settings.debug else trt.Logger.WARNING),
     )
 
-    r = interp.run(
-        max_workspace_size=settings.workspace_size,
+    interpreter_result = interpreter.run(
+        workspace_size=settings.workspace_size,
         lower_precision=settings.precision,
         profiling_verbosity=(
             trt.ProfilingVerbosity.VERBOSE
             if settings.debug
             else trt.ProfilingVerbosity.LAYER_NAMES_ONLY
         ),
+        max_aux_streams=settings.max_aux_streams,
+        version_compatible=settings.version_compatible,
+        optimization_level=settings.optimization_level,
     )
 
-    return TRTModule(*r)
+    if settings.use_experimental_rt:
+        with io.BytesIO() as engine_bytes:
+            engine_bytes.write(interpreter_result.engine.serialize())
+            engine_str = engine_bytes.getvalue()
+        return TRTModuleNext(
+            serialized_engine=engine_str,
+            name=name,
+            input_binding_names=interpreter_result.input_names,
+            output_binding_names=interpreter_result.output_names,
+        )
+    else:
+        return TRTModule(
+            engine=interpreter_result.engine,
+            input_names=interpreter_result.input_names,
+            output_names=interpreter_result.output_names,
+        )