# ltx-bench

Profiling and benchmarking tools for LTX-2.

## Running benchmarks

```bash
# All benchmarks (requires H100 GPU + model checkpoints at /models/)
uv run pytest -m bench --bench

# With Nsight Systems profiling
uv run pytest -m bench --bench --nsys

# Custom output directory and iterations
uv run pytest -m bench --bench \
  --bench-output=./my_results \
  --bench-warmup=3 \
  --bench-repeats=10
```

## Benchmark types

### Components

Individual model components benchmarked in isolation:

| Benchmark | What it measures |
|-----------|-----------------|
| `transformer_forward_video_only` | X0Model forward pass (video only) |
| `transformer_forward_av` | X0Model forward pass (video + audio) |
| `video_vae_decode` | Video VAE decoder |
| `video_vae_encode` | Video VAE encoder |
| `audio_decode` | Audio VAE decoder + vocoder |
| `audio_encode` | Audio VAE encoder |
| `text_encoder_forward` | Gemma text encoder (positive + negative prompt) |
| `spatial_upsampler` | Latent spatial upsampler 2x |

### Pipelines

End-to-end pipeline benchmarks with per-stage instrumentation:

| Benchmark | Pipeline |
|-----------|----------|
| `distilled` | 8-step distilled inference |
| `ti2vid_one_stage` | Text-to-video one-stage |
| `ti2vid_two_stages` | Text-to-video two-stage with upsampling |
| `keyframe_interpolation` | Keyframe interpolation two-stage |
| `ic_lora` | IC-LoRA with depth conditioning |
| `a2vid_two_stage` | Audio-to-video two-stage |
| `retake` | Video retake with regeneration |

## Trend analysis

Generate trend charts from GCS historical data:

```bash
python -m ltx_bench.trends --branch main --last 20 --output ./trend_report
```

## Report output

Each run produces a `REPORT.md` with:
- Per-component GPU time, peak VRAM, and peak RAM
- Per-pipeline total time with stage breakdown (text encoding, denoising, VAE decode, etc.)
- Denoising throughput in s/it