A100 FlashVSR

One restore command. One run command. 4K video from a Vast.ai A100/A800.

This repository packages a known-good FlashVSR runtime for NVIDIA A100 / A800 (sm80) plus the shell scripts that actually process long videos: overlapped chunks, original frame timing, and a 3840×2160 output.

中文：在 Vast.ai 上租一台 A100/A800，一条命令还原已经编好的 FlashVSR 环境，再一条命令把视频超到 4K。默认保持片源帧率，竖屏会先转横再处理，长视频按重叠分块避免接缝跳变。

---

Why this exists

FlashVSR on a cloud GPU usually fails for boring reasons: disk too small, cuDNN too old, Block-Sparse-Attention rebuilt for the wrong architecture, ffmpeg eating the chunk list from stdin, seams between 8-second segments.

This repo is the path that already worked:

1. Restore a prebuilt conda env + FlashVSR + Block-Sparse-Attention from GitHub Releases (no compile on a fresh box).
2. Run run_flashvsr_video.sh on one file or a folder.

Hardware

Item	Requirement
GPU	A100 80GB or A800 80GB (sm80). Other architectures need a fresh Block-Sparse-Attention build.
CUDA	Driver reporting ≥ 12.8. Runtime in the backup is torch 2.11.0+cu128, cuDNN 91900.
Disk	100 GB minimum, 150 GB better. A 32 GB root disk is not enough.
Image	Vast.ai CUDA 12.8 base / Jupyter image is fine.

Quick start

On a new Vast.ai instance:

cd /workspace
apt-get update
apt-get install -y git
git clone https://github.com/PA5MIN/A100.git
bash /workspace/A100/vastai_setup_flashvsr.sh

Put videos in /workspace/flashvsr_jobs/input, then:

bash /workspace/A100/run_flashvsr_video.sh

Or pass a path:

bash /workspace/A100/run_flashvsr_video.sh /workspace/my_video.mp4

Outputs land in /workspace/flashvsr_jobs/output as *_3840x2160_flashvsr.mp4.

Step-by-step notes: DEPLOY.md. Real failure modes from the first restore: TROUBLESHOOTING.md.

What the runner does

Default: keep source frame timing (no fps= filter). Portrait is rotated to landscape first.

portrait 1088×1920
  → rotate 1920×1088
  → scale 960×544
  → FlashVSR ×4  → 3840×2176
  → crop 8 px top/bottom → 3840×2160
  → mux original audio

portrait 1080×1920
  → rotate 1920×1080
  → scale 960×540
  → pad 2 px → 960×544
  → FlashVSR ×4 → 3840×2176
  → crop → 3840×2160
  → mux original audio

Landscape skips rotation. Matching 960×544 scales directly; 16:9 scales to 960×540 then pads; anything else is center-cropped to fit.

Chunks overlap so FlashVSR’s 8n−3 tail loss does not show up as a jump. Every ffmpeg call uses -nostdin so it cannot steal the chunk list from the shell loop.

Useful overrides

# Force 30 fps (off by default)
FPS=30 bash /workspace/A100/run_flashvsr_video.sh /workspace/video.mp4

# Custom preprocess
PREPROCESS_FILTER='transpose=1,scale=960:544:flags=area,setsar=1' \
  bash /workspace/A100/run_flashvsr_video.sh /workspace/video.mp4

# Opposite portrait rotation
PORTRAIT_ROTATE_FILTER=transpose=2 bash /workspace/A100/run_flashvsr_video.sh /workspace/video.mp4

# Honor container rotation metadata
FFMPEG_INPUT_OPTS='' bash /workspace/A100/run_flashvsr_video.sh /workspace/video.mp4

# More overlap if a chunk is short on frames
OVERLAP_FRAMES=24 bash /workspace/A100/run_flashvsr_video.sh

What’s in the release

Tag a100-cu128-cudnn919-sm80-20260523 restores:

/workspace/envs/flashvsr-fast
/workspace/FlashVSR
/workspace/Block-Sparse-Attention
/workspace/flashvsr_backup_meta

About 18 GB compressed split archives, SHA-256 checked before extract. vastai_setup_flashvsr.sh refuses to start unless /workspace has at least 80 GB free.

Scripts

File	Role
vastai_setup_flashvsr.sh	Download release, verify checksums, install Miniconda, restore env
run_flashvsr_video.sh	Daily path: preprocess → overlapped FlashVSR → concat → audio
flashvsr_overlap_fix.sh	Standalone overlap-trim helper from the first seam-debug session

Credits

• FlashVSR — OpenImagingLab
• FlashVSR-v1.1 weights

This repo only ships restore/run glue and a prebuilt sm80 runtime. Model licenses follow upstream.

License

Scripts and docs in this repository are MIT. Third-party code and weights inside the release keep their original licenses.