rave-tensorrt

TensorRT-backed inference implementation for RAVE.

rave-tensorrt provides an UpscaleBackend implementation powered by ONNX Runtime with TensorRT/CUDA execution providers and device-memory I/O binding.

Scope

• Backend implementation: TensorRtBackend
• Precision policy controls: PrecisionPolicy
• Batch settings surface: BatchConfig
• Inference/output-ring metrics and telemetry
• Provider bridge handling for Linux/WSL runtime compatibility

Public API Highlights

• TensorRtBackend::new(model_path, ctx, device_id, ring_size, downstream_capacity)
• TensorRtBackend::with_precision(...)
• BatchConfig { max_batch, latency_deadline_us }
• TensorRtBackend implements rave_core::backend::UpscaleBackend

Typical Usage

use std::path::PathBuf;
use std::sync::Arc;

use rave_core::backend::UpscaleBackend;
use rave_core::context::GpuContext;
use rave_core::error::Result;
use rave_tensorrt::tensorrt::TensorRtBackend;

async fn init_backend(ctx: Arc<GpuContext>) -> Result<TensorRtBackend> {
    let backend = TensorRtBackend::new(PathBuf::from("model.onnx"), ctx, 0, 6, 4);
    backend.initialize().await?;
    Ok(backend)
}

Runtime Requirements

• ONNX Runtime shared libs and provider libs discoverable at runtime
• CUDA/TensorRT/cuDNN stack compatible with ORT build
• NVIDIA driver libs discoverable (libcuda, etc.)

Notes

• Data path is GPU-resident via ORT I/O binding.
• BatchConfig is part of the API surface; current processing path is single-frame.
• Provider selection can be controlled externally via RAVE_ORT_TENSORRT.