TRT-LLM loading mechanism tool

dev_dep_versions.yml

@@ -1,2 +1,3 @@
 __cuda_version__: "12.8"
 __tensorrt_version__: "10.11.0"
+__tensorrt_llm_version__: "0.17.0.post1"

py/torch_tensorrt/dynamo/_compiler.py

@@ -103,6 +103,7 @@ def cross_compile_for_windows(
     tiling_optimization_level: str = _defaults.TILING_OPTIMIZATION_LEVEL,
     l2_limit_for_tiling: int = _defaults.L2_LIMIT_FOR_TILING,
     offload_module_to_cpu: bool = _defaults.OFFLOAD_MODULE_TO_CPU,
+    use_distributed_mode_trace: bool = _defaults.USE_DISTRIBUTED_MODE_TRACE,
     **kwargs: Any,
 ) -> torch.fx.GraphModule:
     """Compile an ExportedProgram module using TensorRT in Linux for Inference in Windows
@@ -177,6 +178,7 @@ def cross_compile_for_windows(
         enable_weight_streaming (bool): Enable weight streaming.
         tiling_optimization_level (str): The optimization level of tiling strategies. A higher level allows TensorRT to spend more time searching for better tiling strategy. We currently support ["none", "fast", "moderate", "full"].
         l2_limit_for_tiling (int): The target L2 cache usage limit (in bytes) for tiling optimization (default is -1 which means no limit).
+        use_distributed_mode_trace (bool):  Using aot_autograd to trace the graph. This is enabled when DTensors or distributed tensors are present in distributed model
         **kwargs: Any,
     Returns:
         torch.fx.GraphModule: Compiled FX Module, when run it will execute via TensorRT
@@ -339,6 +341,7 @@ def cross_compile_for_windows(
         "enable_weight_streaming": enable_weight_streaming,
         "tiling_optimization_level": tiling_optimization_level,
         "l2_limit_for_tiling": l2_limit_for_tiling,
+        "use_distributed_mode_trace": use_distributed_mode_trace,
     }
 
     # disable the following settings is not supported for cross compilation for windows feature
@@ -439,6 +442,7 @@ def compile(
     tiling_optimization_level: str = _defaults.TILING_OPTIMIZATION_LEVEL,
     l2_limit_for_tiling: int = _defaults.L2_LIMIT_FOR_TILING,
     offload_module_to_cpu: bool = _defaults.OFFLOAD_MODULE_TO_CPU,
+    use_distributed_mode_trace: bool = _defaults.USE_DISTRIBUTED_MODE_TRACE,
     **kwargs: Any,
 ) -> torch.fx.GraphModule:
     """Compile an ExportedProgram module for NVIDIA GPUs using TensorRT
@@ -516,6 +520,7 @@ def compile(
         tiling_optimization_level (str): The optimization level of tiling strategies. A higher level allows TensorRT to spend more time searching for better tiling strategy. We currently support ["none", "fast", "moderate", "full"].
         l2_limit_for_tiling (int): The target L2 cache usage limit (in bytes) for tiling optimization (default is -1 which means no limit).
         offload_module_to_cpu (bool): Offload the module to CPU. This is useful when we need to minimize GPU memory usage.
+        use_distributed_mode_trace (bool):  Using aot_autograd to trace the graph. This is enabled when DTensors or distributed tensors are present in distributed model
         **kwargs: Any,
     Returns:
         torch.fx.GraphModule: Compiled FX Module, when run it will execute via TensorRT
@@ -688,6 +693,7 @@ def compile(
         "tiling_optimization_level": tiling_optimization_level,
         "l2_limit_for_tiling": l2_limit_for_tiling,
         "offload_module_to_cpu": offload_module_to_cpu,
+        "use_distributed_mode_trace": use_distributed_mode_trace,
     }
 
     settings = CompilationSettings(**compilation_options)
@@ -1029,6 +1035,7 @@ def convert_exported_program_to_serialized_trt_engine(
     tiling_optimization_level: str = _defaults.TILING_OPTIMIZATION_LEVEL,
     l2_limit_for_tiling: int = _defaults.L2_LIMIT_FOR_TILING,
     offload_module_to_cpu: bool = _defaults.OFFLOAD_MODULE_TO_CPU,
+    use_distributed_mode_trace: bool = _defaults.USE_DISTRIBUTED_MODE_TRACE,
     **kwargs: Any,
 ) -> bytes:
     """Convert an ExportedProgram to a serialized TensorRT engine
@@ -1093,6 +1100,7 @@ def convert_exported_program_to_serialized_trt_engine(
         enable_weight_streaming (bool): Enable weight streaming.
         tiling_optimization_level (str): The optimization level of tiling strategies. A higher level allows TensorRT to spend more time searching for better tiling strategy. We currently support ["none", "fast", "moderate", "full"].
         l2_limit_for_tiling (int): The target L2 cache usage limit (in bytes) for tiling optimization (default is -1 which means no limit).
+        use_distributed_mode_trace: bool = _defaults.USE_DISTRIBUTED_MODE_TRACE,
     Returns:
         bytes: Serialized TensorRT engine, can either be saved to a file or deserialized via TensorRT APIs
     """
@@ -1215,6 +1223,7 @@ def convert_exported_program_to_serialized_trt_engine(
         "tiling_optimization_level": tiling_optimization_level,
         "l2_limit_for_tiling": l2_limit_for_tiling,
         "offload_module_to_cpu": offload_module_to_cpu,
+        "use_distributed_mode_trace": use_distributed_mode_trace,
     }
 
     settings = CompilationSettings(**compilation_options)

py/torch_tensorrt/dynamo/conversion/converter_utils.py

@@ -1,8 +1,6 @@
 import collections
-import ctypes
 import functools
 import logging
-import os
 from typing import (
     Any,
     Callable,
@@ -1116,69 +1114,6 @@ def args_bounds_check(
     return args[i] if len(args) > i and args[i] is not None else replacement
 
 
-def load_tensorrt_llm() -> bool:
-    """
-    Attempts to load the TensorRT-LLM plugin and initialize it.
-
-    Returns:
-        bool: True if the plugin was successfully loaded and initialized, False otherwise.
-    """
-    try:
-        import tensorrt_llm as trt_llm  # noqa: F401
-
-        _LOGGER.info("TensorRT-LLM successfully imported")
-        return True
-    except (ImportError, AssertionError) as e_import_error:
-        # Check for environment variable for the plugin library path
-        plugin_lib_path = os.environ.get("TRTLLM_PLUGINS_PATH")
-        if not plugin_lib_path:
-            _LOGGER.warning(
-                "TensorRT-LLM is not installed. Please install TensorRT-LLM or set TRTLLM_PLUGINS_PATH to the directory containing libnvinfer_plugin_tensorrt_llm.so to use converters for torch.distributed ops",
-            )
-            return False
-
-        _LOGGER.info(f"TensorRT-LLM Plugin lib path found: {plugin_lib_path}")
-        try:
-            # Load the shared library
-            handle = ctypes.CDLL(plugin_lib_path)
-            _LOGGER.info(f"Successfully loaded plugin library: {plugin_lib_path}")
-        except OSError as e_os_error:
-            _LOGGER.error(
-                f"Failed to load libnvinfer_plugin_tensorrt_llm.so from {plugin_lib_path}"
-                f"Ensure the path is correct and the library is compatible",
-                exc_info=e_os_error,
-            )
-            return False
-
-        try:
-            # Configure plugin initialization arguments
-            handle.initTrtLlmPlugins.argtypes = [ctypes.c_void_p, ctypes.c_char_p]
-            handle.initTrtLlmPlugins.restype = ctypes.c_bool
-        except AttributeError as e_plugin_unavailable:
-            _LOGGER.warning(
-                "Unable to initialize the TensorRT-LLM plugin library",
-                exc_info=e_plugin_unavailable,
-            )
-            return False
-
-        try:
-            # Initialize the plugin
-            TRT_LLM_PLUGIN_NAMESPACE = "tensorrt_llm"
-            if handle.initTrtLlmPlugins(None, TRT_LLM_PLUGIN_NAMESPACE.encode("utf-8")):
-                _LOGGER.info("TensorRT-LLM plugin successfully initialized")
-                return True
-            else:
-                _LOGGER.warning("TensorRT-LLM plugin library failed in initialization")
-                return False
-        except Exception as e_initialization_error:
-            _LOGGER.warning(
-                "Exception occurred during TensorRT-LLM plugin library initialization",
-                exc_info=e_initialization_error,
-            )
-            return False
-    return False
-
-
 def promote_trt_tensors_to_same_dtype(
     ctx: ConversionContext, lhs: TRTTensor, rhs: TRTTensor, name_prefix: str
 ) -> tuple[TRTTensor, TRTTensor]:

py/torch_tensorrt/dynamo/conversion/custom_ops_converters.py

@@ -11,11 +11,11 @@
 from torch_tensorrt.dynamo.conversion._ConverterRegistry import (
     dynamo_tensorrt_converter,
 )
-from torch_tensorrt.dynamo.conversion.converter_utils import load_tensorrt_llm
 from torch_tensorrt.dynamo.lowering.passes.fuse_distributed_ops import (
     tensorrt_fused_nccl_all_gather_op,
     tensorrt_fused_nccl_reduce_scatter_op,
 )
+from torch_tensorrt.dynamo.utils import load_tensorrt_llm
 
 _LOGGER: logging.Logger = logging.getLogger(__name__)