Upload lingbot-world-repro/report.md with huggingface_hub

lingbot-world-repro/report.md

@@ -0,0 +1,50 @@
+# LingBot-World Inference Reproduction Report
+
+## Setup
+
+- Repository: https://github.com/robbyant/lingbot-world
+- Conda environment: `lingbot_world_repro`
+- FlashAttention: `2.7.4.post1` prebuilt wheel for CUDA 12 / PyTorch 2.5 / Python 3.10.
+- GPU execution: 1 Slurm node, 4 GPUs, `torchrun --nproc_per_node=4`
+- Generation size argument: `480*832` with source aspect ratio preserved.
+- Frame count: 81 frames at 16 FPS
+
+## Method Difference
+
+- `inference` runs `generate.py` with the LingBot-World Base camera model. It denoises the full video sequence with the normal Wan I2V sampling path.
+- `fast_inference` runs `generate_fast.py` with the LingBot-World Fast weights under `lingbot_world_fast`. It uses causal chunk-by-chunk inference and KV caching to reduce latency for interactive generation.
+
+## Aggregate Metrics
+
+| Method | Completed Videos | Mean Time | Peak Single GPU VRAM | Peak Total VRAM |
+| --- | ---: | ---: | ---: | ---: |
+| inference | 8 | 20m 20.3s | 43.1 GiB | 171.4 GiB |
+| fast_inference | 8 | 7m 06.3s | 37.3 GiB | 123.2 GiB |
+
+## Comparison
+
+- Mean wall time: `fast_inference` is 2.86x faster than `inference` on this 8-case run.
+- Peak single-GPU VRAM: `fast_inference` reduces the observed peak by 5.8 GiB.
+- Peak total VRAM across 4 GPUs: `fast_inference` reduces the observed peak by 48.2 GiB.
+- Video validation: all 16 MP4 outputs open successfully and report 81 frames at 16 FPS. Fourteen outputs are 832x464; the two Great Wall outputs preserve their source aspect ratio at 768x512.
+
+## Per-Video Results
+
+| Method | Case | Status | Time | Peak Single GPU VRAM | Output |
+| --- | --- | --- | ---: | ---: | --- |
+| fast_inference | case00_seed142 | ok | 7m 35.1s | 36.5 GiB | `experiments/lingbot_repro/outputs/fast_inference/fast_inference_case00_seed142.mp4` |
+| fast_inference | case00_seed42 | ok | 8m 29.5s | 36.3 GiB | `experiments/lingbot_repro/outputs/fast_inference/fast_inference_case00_seed42.mp4` |
+| fast_inference | case01_seed143 | ok | 9m 13.2s | 36.5 GiB | `experiments/lingbot_repro/outputs/fast_inference/fast_inference_case01_seed143.mp4` |
+| fast_inference | case01_seed43 | ok | 8m 14.9s | 36.5 GiB | `experiments/lingbot_repro/outputs/fast_inference/fast_inference_case01_seed43.mp4` |
+| fast_inference | case02_seed44 | ok | 9m 05.7s | 36.8 GiB | `experiments/lingbot_repro/outputs/fast_inference/fast_inference_case02_seed44.mp4` |
+| fast_inference | case03_seed45 | ok | 2m 25.9s | 36.7 GiB | `experiments/lingbot_repro/outputs/fast_inference/fast_inference_case03_seed45.mp4` |
+| fast_inference | case04_seed46 | ok | 2m 37.1s | 37.3 GiB | `experiments/lingbot_repro/outputs/fast_inference/fast_inference_case04_seed46.mp4` |
+| fast_inference | case05_seed47 | ok | 9m 08.6s | 36.2 GiB | `experiments/lingbot_repro/outputs/fast_inference/fast_inference_case05_seed47.mp4` |
+| inference | case00_seed142 | ok | 16m 56.2s | 43.1 GiB | `experiments/lingbot_repro/outputs/inference/inference_case00_seed142.mp4` |
+| inference | case00_seed42 | ok | 22m 27.6s | 42.4 GiB | `experiments/lingbot_repro/outputs/inference/inference_case00_seed42.mp4` |
+| inference | case01_seed143 | ok | 10m 48.7s | 42.6 GiB | `experiments/lingbot_repro/outputs/inference/inference_case01_seed143.mp4` |
+| inference | case01_seed43 | ok | 21m 02.5s | 42.9 GiB | `experiments/lingbot_repro/outputs/inference/inference_case01_seed43.mp4` |
+| inference | case02_seed44 | ok | 23m 44.3s | 43.1 GiB | `experiments/lingbot_repro/outputs/inference/inference_case02_seed44.mp4` |
+| inference | case03_seed45 | ok | 23m 23.9s | 43.0 GiB | `experiments/lingbot_repro/outputs/inference/inference_case03_seed45.mp4` |
+| inference | case04_seed46 | ok | 22m 10.6s | 43.1 GiB | `experiments/lingbot_repro/outputs/inference/inference_case04_seed46.mp4` |
+| inference | case05_seed47 | ok | 22m 08.5s | 43.1 GiB | `experiments/lingbot_repro/outputs/inference/inference_case05_seed47.mp4` |