Skip to content

README.md

Latest commit

 

History

History
96 lines (73 loc) · 2.87 KB

README.md

File metadata and controls

96 lines (73 loc) · 2.87 KB

FlashVSR Example

Video Super-Resolution using FlashVSR model with streaming inference support.

Model Source

Platform Link
HuggingFace lzx1413/FlashVSR-v1.1-BF16
ModelScope lzx1413/FlashVSR-v1.1-BF16

Feature Support

Feature Support
CFG Parallel (CFGP) N/A
Ulysses Sequence Parallel (USP) ✔️
LoRA N/A
FP8 Quantization ❔
FSDP ✔️
Encoder Parallel N/A
Async Pipeline ❔
Feature Cache (AdaTaylor) N/A
Distilled Model N/A
Server API N/A

Start

  1. install block sparse attention
git clone https://github.com/mit-han-lab/Block-Sparse-Attention.git
cd Block-Sparse-Attention
pip install packaging
pip install ninja
python setup.py install
  1. install tf-kernel refer to tf-kernel sub-folder

Model Files

The model root directory should contain:

  • flashvsr11_dit_streaming_dmd_5dc619.safetensors - DiT model
  • TCDecoder.ckpt - VAE decoder

Files

flashvsr_stream.py

Streaming video super-resolution example that processes video in chunks.

Purpose: Demonstrates streaming inference for video super-resolution, suitable for processing long videos with limited GPU memory.

Usage:

# Basic usage (auto-detect resolution)
python examples/flashvsr/flashvsr_stream.py -i /path/to/video.mp4 -s 4

# Specify input resolution
python examples/flashvsr/flashvsr_stream.py -i /path/to/video.mp4 -s 4 --height 480 --width 854

# Multi-GPU inference
python examples/flashvsr/flashvsr_stream.py -i /path/to/video.mp4 -s 4 --gpu_num 2

# Custom model path and output
python examples/flashvsr/flashvsr_stream.py \
    -i /path/to/video.mp4 \
    -s 4 \
    --model_root /path/to/FlashVSR-v1.1 \
    -o output.mp4

Parameters:

Parameter Default Description
-i, --input_video (required) Path to input low-quality video
-s, --scale 4 Upscaling factor
-h, --height auto Input video height (auto-detect if not specified)
-w, --width auto Input video width (auto-detect if not specified)
--gpu_num 1 Number of GPUs to use
--model_root /dev/shm/zuoxin/flashvsr Root directory containing model files
-o, --output auto Output video path
--seed 0 Random seed

Performance

Config Device Target Resolution Speed Max VRAM (GB)
flashvsr_stream 5090*1 1920x1080 1.5it/s (per chunk) 22
flashvsr_stream 5090*2 1920x1080 2.1it/s (per chunk) 20

Notes

  • The example uses streaming mode for memory-efficient inference
  • local_range parameter: 9 for sharper details, 11 for more stable results
  • Multi-GPU uses Ulysses sequence parallelism