rave-cli

Command-line interface for the RAVE GPU video pipeline.

rave-cli wires all workspace crates into an operator-facing executable with commands for pipeline execution, benchmarking, and environment probing.

Commands

• upscale: full decode -> preprocess -> infer -> encode path
• benchmark: throughput-focused run with optional encode skip/fallback
• probe: CUDA/provider probe for runtime diagnostics
• devices: list visible CUDA devices

JSON Contract

• --json emits one final JSON object on stdout
• progress is emitted on stderr via --progress (human or jsonl)
• payloads include schema_version for parser stability

Quickstart

# Human-readable probe
target/debug/rave probe

# JSON probe
target/debug/rave probe --json

# Benchmark without encode
target/debug/rave benchmark --input in.mp4 --model model.onnx --skip-encode

# Benchmark JSON + output file
target/debug/rave benchmark --input in.mp4 --model model.onnx --skip-encode --json --json-out /tmp/bench.json

# Full upscale
target/debug/rave upscale --input in.mp4 --output out.mp4 --model model.onnx

Feature Flags

• debug-alloc: enables allocation instrumentation across rave-core, rave-pipeline, and rave-tensorrt

Runtime Requirements

• NVIDIA CUDA driver + toolkit runtime
• NVDEC/NVENC runtime libraries
• FFmpeg runtime libraries
• ONNX Runtime + TensorRT provider dependencies

Notes

• Non-JSON human output and JSON machine output are intentionally separated.
• Progress streams are opt-in to keep stdout stable for scripting.