Crate atomr_accel_tensorrt 

atomr-accel-tensorrt

TensorRT engine builder + runtime as supervised atomr actors. Wraps NVIDIA’s libnvinfer (and optionally libnvonnxparser) at runtime — the library itself is not vendored because it is proprietary; users opt in via the tensorrt-link feature and either install TensorRT system-wide or set LIBNVINFER_PATH.

Features

• tensorrt-link — actually link libnvinfer at build time. Off-by-default so the crate compiles on hosts without TensorRT (used by CI + unit tests).
• tensorrt-onnx — pull in nvonnxparser for ONNX import.
• tensorrt-plugin — IPluginV3 Rust trampolines.
• tensorrt-int8 — INT8 calibration helpers (entropy / minmax).
• tensorrt-fp8 — FP8 PTQ helpers (Hopper-class GPUs).

Public surface

• actor::TrtActor / actor::TrtMsg — sibling actor to atomr_accel_cuda::DeviceActor. Shares Arc<CudaStream> with the device actor so inference rides the same execution timeline.
• builder::IBuilderConfig — pure-Rust mirror of the TensorRT builder config, with knobs for precision, DLA, structured sparsity, tactic sources, timing cache, and engine refit.
• engine::TrtEngine — owned, immutable engine handle that’s Send + Sync via newtype.
• runtime::TrtRuntime / runtime::ExecutionContext — load serialised plans + drive enqueueV3 on a shared CUDA stream.
• [onnx::OnnxParser] — gated on tensorrt-onnx.
• [calibration] — gated on tensorrt-int8 / tensorrt-fp8.
• [plugin] — gated on tensorrt-plugin.

Re-exports

pub use actor::BuildFromOnnxReply;

pub use actor::BuildReply;

pub use actor::CreateContextReply;

pub use actor::DeserializeReply;

pub use actor::EnqueueReply;

pub use actor::ExecuteReply;

pub use actor::NetworkSource;

pub use actor::RefitReply;

pub use actor::RefitWeights;

pub use actor::TrtActor;

pub use actor::TrtMsg;

pub use builder::BuilderFlags;

pub use builder::DeviceType;

pub use builder::IBuilderConfig;

pub use builder::Precision;

pub use builder::RefitPolicy;

pub use builder::TacticSources;

pub use engine::EnginePlan;

pub use engine::TrtEngine;

pub use engine::TrtRefitter;

pub use error::TrtError;

pub use runtime::EnqueueRequest;

pub use runtime::ExecutionBindings;

pub use runtime::ExecutionContext;

pub use runtime::TensorShape;

pub use runtime::TrtRuntime;

Modules

actor

TrtActor — sibling of atomr_accel_cuda::DeviceActor.

builder

Safe wrappers for nvinfer1::IBuilder and nvinfer1::IBuilderConfig.

engine

Safe wrapper around nvinfer1::ICudaEngine.

error

Error types for the TensorRT actor surface.

runtime

Safe wrapper around nvinfer1::IExecutionContext and the IRuntime deserialiser. This is the inference-time hot path.

sys

Hand-written FFI surface for libnvinfer (and libnvonnxparser when the tensorrt-onnx feature is on).

Functions

init_logger

No-op when the tensorrt-link feature is off so callers don’t need to gate their init_logger() calls.