The repository benchmark measures the complete local path: task submission, neural TTS, ASS subtitle generation, FFmpeg composition, MP4 download, and encoder reporting. Run it with:
MPT_BENCH_URL=http://127.0.0.1:8501 ./scripts/benchmark.shThe current validation environment exposes FFmpeg software support for h264_nvenc, h264_qsv, and h264_vaapi, but it has no usable NVIDIA runtime, no /dev/dri/renderD128, and no detected GPU. Therefore MPT_GPU_BACKEND=auto correctly selected CPU / libx264.
A real end-to-end benchmark using a ten-second custom narration completed successfully in approximately 7.095 seconds and produced a valid MP4. A separate ASS integration run completed in approximately 6.768 seconds, generated captions.ass, served it through the artifact endpoint, and passed ffprobe validation.
For parallel batch measurements, set MPT_BENCH_BATCH_COUNT=2 before running the benchmark. The task engine renders the visual background and final MP4 for each variant concurrently, while keeping narration and subtitle preparation shared. MPT_MAX_BATCH_WORKERS defaults to 2 to avoid uncontrolled FFmpeg process growth; increase it only after measuring CPU, GPU memory, disk throughput, and thermal behavior on the target machine.
In the current validation environment, a real batch of two custom-narration videos completed in approximately 12.130 seconds with CPU / libx264. This is a host-specific measurement, not a universal speedup claim.
GPU backends are selected only when both the FFmpeg encoder and a usable runtime are detected. Explicitly requesting an unavailable backend fails early with a clear API validation error; auto always remains safe and falls back to CPU. Because no GPU was present in this environment, a numeric GPU-vs-CPU speedup claim cannot be made without hardware-specific measurements.