8mb.local – Self-Hosted GPU Video Compressor

8mb.local is a self-hosted, fire-and-forget video compressor. Drop a file, choose a target size (e.g., 8 MB, 25 MB, 50 MB, 100 MB), and let GPU-accelerated encoding produce compact outputs with AV1/HEVC/H.264. Supports NVIDIA NVENC, Intel Quick Sync, Windows AMD AMF, and Linux VAAPI (including AMD) with automatic CPU fallback. The Docker deployment uses a SvelteKit UI, FastAPI backend, Celery worker, Redis broker, and real-time progress via Server-Sent Events (SSE). The native Windows installer runs the same UI/API/worker code with a local in-process queue.

Video Demo

• Features
• Screenshots
• Projects
• Architecture
• Installation
• Usage
• Configuration
• Performance & Concurrency
• Reverse Proxy Configuration
• Troubleshooting
• License

• NVIDIA NVENC, Intel QSV, Windows AMD AMF, and Linux VAAPI hardware encoding with automatic CPU fallback when a GPU or driver is unavailable
• Robust encoder validation at startup — tests actual encoder initialization, not just availability
• AV1, HEVC (H.265), and H.264 encoding via NVENC, QSV, AMF, VAAPI, or CPU software encoders
• Drag-and-drop UI with helpful presets and advanced options (codec, container, tune, audio bitrate)
• Configurable codec visibility — enable/disable specific codecs in the Settings page
• Resolution control — set max width/height while maintaining aspect ratio
• Video trimming — specify start/end times (seconds or HH:MM:SS format)
• Real-time progress tracking using output size, time processed, bitrate, and wall-clock estimates
• Real-time FFmpeg logs streamed during compression
• Live queue management — view all active jobs with real-time progress, cancel individual jobs, or clear entire queue
• Automatic file size optimization — re-encodes with adjusted bitrate if output exceeds target by >2%
• Batch processing — compress multiple files in a single operation
• Job history enabled by default
• Auto-download enabled by default
• Output container choice: MP4 or MKV, with compatibility safeguards
• Version tracking — UI displays current version, backend provides /api/version

Main Interface	GPU Support List	Settings Panel
Live Queue	Compressing (Real-time Logs)	Encoder Validation Tests
Job History	Advanced Options

Public instances run by people who offer their 8mb.local install for anyone to use (community compressors, demos, mirrors). If you run a public deployment and want it listed here, open a pull request that adds a row to this section.

This repository is one shared source codebase for the frontend, backend, worker, Windows packages, and Docker image. The root VERSION file is the single active application-version source. Generated UI/backend version modules and packaging metadata are synchronized by scripts\set-version.ps1 and verified by scripts\check-version.ps1.

One command runs the automated checks and builds the portable EXE, installer EXE, Store-submission MSIX, and local Docker image:

.\release-local.ps1 -Version 140.0.0.0

GitHub is not required to generate these files. GitHub Actions may still provide an independent compatibility check later. The local workflow never pushes its Docker image, publishes a release, deploys the application, or submits the MSIX; Microsoft Partner Center remains a separate manual submission step.

The UI, backend API, EXE metadata, installer metadata, MSIX manifest, Docker metadata/tag, artifact names, and release directory all derive from the requested four-part version. By default, outputs and stage logs are written to dist\release\<version>\:

• 8mblocal.exe
• 8mblocal-Setup.exe
• 8mblocal_<version>_x64.msix
• 8mblocal-docker.tar
• SHA256SUMS.txt
• BUILD-MANIFEST.json
• TEST-RESULTS.md

Required tools depend on the selected stages: Python 3.11-3.13, Node.js, npm, Docker with Compose, Inno Setup, Windows SDK MakeAppx, and 7-Zip when the Windows build must download its tested FFmpeg bundle. The script checks tools before building and fails with a clear missing-tool message; it does not install unrelated software or perform system upgrades.

Prepare the Python test environment once after cloning:

python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt httpx==0.27.2

Useful commands:

# Validate tools and show the plan without changing versions or building
.\release-local.ps1 -Version 140.0.0.0 -DryRun

# Full local release
.\release-local.ps1 -Version 140.0.0.0

# Windows artifacts only
.\release-local.ps1 -Version 140.0.0.0 -SkipDocker

# Docker artifact only
.\release-local.ps1 -Version 140.0.0.0 -SkipWindows

# Windows EXE and installer without MSIX
.\release-local.ps1 -Version 140.0.0.0 -SkipMsix

-SkipTests is troubleshooting-only and marks the result incomplete. -OutputDir selects another output folder, -KeepTemp preserves temporary files, and -Overwrite may reuse only a release directory previously created and marked by this script. Arbitrary existing directories, source directories, and ancestor paths are protected from overwrite.

Verify generated checksums from the release directory:

Get-Content .\dist\release\140.0.0.0\SHA256SUMS.txt
Get-FileHash .\dist\release\140.0.0.0\8mblocal.exe -Algorithm SHA256

When a build fails, inspect TEST-RESULTS.md, BUILD-MANIFEST.json, and the named stage log, correct the source or environment issue, and rerun into a new output directory. A run that skips required stages is never reported as release-ready.

flowchart LR
  A[Browser / SvelteKit UI] -- Upload / SSE --> B(FastAPI Backend)
  B -- Enqueue --> C[Redis]
  D[Celery Worker + FFmpeg GPU/CPU] -- Progress / Logs --> C
  B -- Pub/Sub relay --> A
  D -- Files --> E[outputs/]
  A -- Download --> B

Components

Data & files

• uploads/ — incoming files (cleaned up after FILE_RETENTION_HOURS)
• outputs/ — compressed results (cleaned up on the same schedule)

All components run in a single container via supervisord.

The Windows installer packages the same frontend, backend, worker, and FFmpeg path in one executable. It replaces only Redis/Celery transport with an in-process bounded queue, stores data under the user's local application data directory, binds to localhost, and opens a native WebView2 window without visible terminal windows. The standalone 8mblocal.exe does not require Docker, Redis, Python, Node.js, or a separate FFmpeg installation. See windows/README.md.

docker run -d \
  --name 8mblocal \
  --gpus all \
  -e NVIDIA_DRIVER_CAPABILITIES=compute,video,utility \
  -p 8001:8001 \
  -v ./uploads:/app/uploads \
  -v ./outputs:/app/outputs \
  jms1717/8mblocal:latest

The -e NVIDIA_DRIVER_CAPABILITIES=compute,video,utility flag is required — it tells the NVIDIA Container Toolkit to mount NVENC libraries into the container.

docker run -d \
  --name 8mblocal \
  -p 8001:8001 \
  -v ./uploads:/app/uploads \
  -v ./outputs:/app/outputs \
  jms1717/8mblocal:latest

Access the web UI at http://localhost:8001.

For Linux hosts with Intel or AMD graphics, use the DRI-enabled compose profile. It discovers /dev/dri/renderD*, validates QSV/VAAPI at startup, and falls back to CPU when the device cannot encode:

docker compose -f docker-compose.vaapi.yml up -d --build

services:
  8mblocal:
    image: jms1717/8mblocal:latest
    container_name: 8mblocal
    ports:
      - "8001:8001"
    volumes:
      - ./uploads:/app/uploads
      - ./outputs:/app/outputs
      - ./.env:/app/.env  # optional
    gpus: all
    environment:
      - NVIDIA_DRIVER_CAPABILITIES=compute,video,utility
    restart: unless-stopped

services:
  8mblocal:
    image: jms1717/8mblocal:latest
    container_name: 8mblocal
    ports:
      - "8001:8001"
    volumes:
      - ./uploads:/app/uploads
      - ./outputs:/app/outputs
      - ./.env:/app/.env  # optional
    restart: unless-stopped

Then run:

docker compose up -d

The maximum application memory budget is user-configurable in the host .env file. MEDIA_MEMORY_LIMIT_GB is the application admission ceiling, while MEDIA_SHM_SIZE is the Docker /dev/shm capacity ceiling. Neither value reserves that amount of host RAM up front.

Recommended adaptive mode: use memory when the live budget fits, and fall back to disk automatically when it does not:

cp .env.example .env
# Edit .env:
MEDIA_STORAGE=auto
MEDIA_MEMORY_LIMIT_GB=10
MEDIA_SHM_SIZE=10g

docker compose up -d --build
docker compose exec 8mblocal df -h /dev/shm

To require memory-backed uploads on Linux/Docker, choose a budget that fits the host and set the shared-memory ceiling at least as high. Jobs that cannot safely fit are rejected instead of silently writing to disk:

# Edit .env:
MEDIA_STORAGE=memory
MEDIA_MEMORY_LIMIT_GB=4
MEDIA_SHM_SIZE=4g

docker compose up -d --build

To force disk-backed temporary uploads, use MEDIA_STORAGE=disk. You can change MEDIA_MEMORY_LIMIT_GB and MEDIA_SHM_SIZE to another host-appropriate value, then recreate the container with docker compose up -d. In auto mode, MEDIA_MEMORY_LIMIT_GB is still enforced as the maximum application budget; available shared memory and host headroom can reduce the amount used.

Default (NVIDIA GPU): requires NVIDIA Container Toolkit and a working docker run --rm --gpus all nvidia/cuda:12.2.0-base-ubuntu22.04 nvidia-smi on the host.

git clone https://github.com/JMS1717/8mb.local.git
cd 8mb.local
docker compose up -d --build

CPU only (no GPU passthrough — e.g. macOS or machine without GPU access):

docker compose -f docker-compose.cpu.yml up -d --build

For a Docker-free Windows install, run the manual Build native Windows executable GitHub Actions workflow and download 8mblocal-Setup.exe from its 8mblocal-windows artifact. The per-user installer creates a Start Menu shortcut and optionally a Desktop shortcut. The release also includes a standalone 8mblocal.exe that can run without installation. Both open the same native WebView2 interface on localhost and probe NVENC, Quick Sync, and AMD AMF before falling back to CPU encoding. See windows/README.md for the installer, Windows security warning, hardware probes, and build details.

The repository includes a disposable scenario harness. It generates a small media corpus, starts the local runtime, uploads real files, runs selected codecs, verifies downloaded outputs with FFprobe, and exercises parallel batch upload plus ZIP download. Explicit hardware requests are useful on CPU-only machines too: a healthy runtime should complete them through its documented CPU fallback path.

# Source/local runtime (no Docker or GPU required)
python scripts/e2e_test.py --mode local

# A representative quicker run
python scripts/e2e_test.py --mode local --profile quick \
  --codecs libx264,h264_qsv,h264_vaapi,h264_nvenc

# A disposable Docker run; build the image first if it is not already present
docker build -t 8mb.local:e2e .
python scripts/e2e_test.py --mode docker --docker-image 8mb.local:e2e

# Hardware-specific Docker runs on the matching host
python scripts/e2e_test.py --mode docker --docker-image 8mb.local:e2e --docker-gpu nvidia
python scripts/e2e_test.py --mode docker --docker-image 8mb.local:e2e --docker-gpu vaapi

The Docker harness uses a unique container name and temporary bind-mounted directories, then stops and removes only the container it created. Use --keep when retaining logs and outputs for diagnosis. The full codec list is the default; narrow it with --codecs when iterating on one hardware path. The manual Docker end-to-end smoke workflow builds a CPU container and runs the same representative scenarios in GitHub Actions.

On Windows, the release script performs the same health, frontend, upload, transcode, status, download, and FFprobe checks. Add -Install to silently install and uninstall the Inno Setup package in an isolated temporary folder:

.\windows\test-release.ps1 -Build -Install

# Check container status
docker ps | grep 8mblocal

# Check NVIDIA GPU access
docker exec 8mblocal nvidia-smi

# List available encoders
docker exec 8mblocal bash -c "ffmpeg -hide_banner -encoders | grep -E 'nvenc|264|265|av1'"

# View startup logs
docker logs 8mblocal

docker compose pull
docker compose up -d

Or with docker run:

docker pull jms1717/8mblocal:latest
docker stop 8mblocal && docker rm 8mblocal
# Re-run your docker run command

1. Drop a video — drag and drop or click Choose File. Analysis runs automatically.
2. Pick a target size — click a preset button or enter a custom MB value.
3. Optional: open Advanced Options
   • Video Codec: AV1 (best quality, RTX 40/50), HEVC (H.265), or H.264 (widest compatibility)
   • Audio Codec: Opus (default) or AAC — MP4 containers auto-switch to AAC
   • Speed/Quality: NVENC presets P1 (fastest) through P7 (best quality), default P6
   • Container: MP4 (most compatible) or MKV (best with Opus audio)
   • Tune: HQ (default), Low Latency, Ultra-Low Latency, or Lossless
   • Resolution: Set max width/height to downscale while preserving aspect ratio
   • Trimming: Set start/end times to compress only a portion
4. Click Compress and watch progress/logs in real time. Cancel anytime. Download starts automatically.

Tips

• For very small targets, prefer AV1 or HEVC and keep audio around 96–128 kbps.
• For speed, try Low Latency tune with a faster preset (P1–P4).
• MP4 + Opus is not supported; the worker auto-switches to AAC for MP4 containers.
• MP4 outputs include +faststart for better web/streaming playback.
• HEVC MP4 outputs use the Apple-compatible hvc1 sample-entry tag so iPhone, iPad, Safari, and other strict players can recognize the video stream.

Create a .env file and mount it at /app/.env:

# Authentication (also configurable via Settings UI)
AUTH_ENABLED=false
AUTH_USER=admin
AUTH_PASS=changeme

# File retention
FILE_RETENTION_HOURS=1

# Worker concurrency (max parallel jobs)
WORKER_CONCURRENCY=4

# SVT-AV1 thread-level parallelism. "auto" is safest across mixed hardware;
# optionally set 0..6 after benchmarking a stable host.
SVTAV1_LP=auto

# Codec visibility is persisted in settings.json and managed from Settings.
# Hardware entries are still hidden unless their runtime probe passes.

# Redis / backend (usually no need to change)
REDIS_URL=redis://127.0.0.1:6379/0
BACKEND_HOST=0.0.0.0
BACKEND_PORT=8001

The Compose profiles bind ./uploads and ./outputs for media, ./state for settings/history, and ./redis-data for the local Redis AOF. Multipart upload and FFmpeg temporary files use ./uploads/.tmp, so large batch uploads do not consume the container's writable layer. Put the media directories on the disk with the most free space and back up state/ if you want to preserve the UI configuration.

Uploads use MEDIA_STORAGE=auto by default. The Docker Compose profiles set a 10 GiB /dev/shm ceiling by default through MEDIA_SHM_SIZE=10g; this is a capacity limit, not a preallocation. Tmpfs consumes host memory only as files are written and releases it when they are deleted. On Linux/Docker, auto mode uses /dev/shm/8mb.local/uploads only when there is at least 512 MB free; use MEDIA_STORAGE=memory to require the memory-backed path, or disk to force normal disk-backed uploads. On native Windows, auto and memory modes keep a normal per-user filesystem pathname and apply FILE_ATTRIBUTE_TEMPORARY, a RAM-preferred Windows cache hint. Windows may still spill temporary data to disk under memory pressure, so this is not a guaranteed RAM disk. Explicit memory mode also applies a conservative per-upload admission budget based on MEDIA_MEMORY_LIMIT_GB, available memory, worker concurrency, FFmpeg working space, and OS headroom. Temporary API upload inputs are removed after the encode, retries, fallback, and final validation finish on every platform; failed partial uploads are removed immediately. Folder Watch source files remain owned by Folder Watch and follow its keep/delete/move policy. Settings, history, Redis data, and final outputs remain on their persistent disk locations.

For Docker, increase or reduce the shared-memory ceiling in .env with MEDIA_SHM_SIZE before recreating the container. Check the active limit with docker compose exec 8mblocal df -h /dev/shm. auto can still fall back to the persistent uploads disk when the available shared memory or the configured MEDIA_MEMORY_LIMIT_GB budget is not sufficient. This keeps RAM use bounded instead of treating the entire host memory pool as a temporary filesystem.

The Folder Watch (Advanced) panel is at the bottom of /settings, is collapsed by default, and can poll an existing Windows, UNC, or Linux-mounted folder. It waits for a file's size and mtime to remain stable, then sends it through the same Celery compression queue used by uploads. It supports recursive scanning, new-only or existing-file processing, an explicit profile, same-folder or specific-folder output, and keep/delete/ move-after-success behavior. Stable seconds is the safety delay before a file is considered finished: the watcher requires the file size and modified time to remain unchanged for that many seconds before starting compression. It is a quiet-period safety delay, not the video's duration and not the total processing time. The default is 5 seconds; use a longer value for slow network copies. The polling interval controls how often the folder is checked and is separate from the stable-file delay. Deletion or moving happens only after the output is non-empty and passes FFprobe validation. Folder Watch state is persisted in the application settings file and is not an arbitrary public-path API.

Manage settings at /settings with no container restart required:

• Authentication — enable/disable, manage credentials
• Default Presets — target size, codec, quality, container defaults
• Codec Visibility — enable/disable NVIDIA, Intel/VAAPI, and CPU codecs
• Preset Profiles — create named presets for quick access
• Worker Concurrency — adjust parallel job limit
• Size Buttons — customize the target size quick-pick buttons
• GPU Support Reference — hardware encoding compatibility at /gpu-support

The Folder Watch panel is also available under Settings for optional cross-platform polling of stable files through the normal queue.

8mb.local supports multiple parallel compression jobs. Configure via Settings UI or WORKER_CONCURRENCY env var.

Considerations

• Most consumer NVIDIA GPUs support 2–3 native NVENC sessions; driver patches or Pro GPUs allow more.
• Each job uses ~200–500 MB RAM and ~100–200 MB VRAM.
• SSD recommended for 6+ concurrent jobs (disk I/O becomes a bottleneck).
• Monitor GPU temps — sustained load above 80 °C may cause throttling.
• Start with 4 concurrent jobs and increase while monitoring utilization.

Container restart required after changing worker concurrency.

SSE (Server-Sent Events) requires special proxy configuration to prevent buffering.

location /api/stream/ {
    proxy_pass http://backend:8001;
    proxy_buffering off;
    proxy_cache off;
    proxy_set_header Connection '';
    chunked_transfer_encoding on;
}

In Nginx Proxy Manager: Edit Proxy Host → Advanced tab → Custom Nginx Configuration.

labels:
  - "traefik.http.middlewares.no-buffer.buffering.maxRequestBodyBytes=0"
  - "traefik.http.middlewares.no-buffer.buffering.maxResponseBodyBytes=0"
  - "traefik.http.routers.8mblocal.middlewares=no-buffer"

<Location /api/stream/>
    ProxyPass http://backend:8001/api/stream/
    ProxyPassReverse http://backend:8001/api/stream/
    SetEnv proxy-sendchunked 1
    SetEnv proxy-interim-response RFC
</Location>

Why this matters: Without proxy_buffering off, your proxy buffers the entire SSE stream and delivers all progress events at once when the job completes — progress appears stuck at 0% until done.

If the host has no NVIDIA GPU, Docker's NVIDIA runtime hook will abort:

nvidia-container-cli: initialization error: WSL environment detected but no adapters were found

Fix: Remove --gpus all and any NVIDIA_* environment variables. The app will start in CPU mode automatically.

1. Missing NVIDIA_DRIVER_CAPABILITIES

Symptom: Cannot load libnvidia-encode.so.1

Fix: Add the environment variable to your docker run or compose:

-e NVIDIA_DRIVER_CAPABILITIES=compute,video,utility

This tells the Container Toolkit to mount NVENC libraries into the container.

2. Driver too old (NVENC API mismatch)

Symptom: Driver does not support the required nvenc API version. Required: 13.0 Found: 12.1

This means your NVIDIA driver is 535.x or older. You need driver 550+.

# Debian 12
wget https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt update && sudo apt install nvidia-driver
sudo reboot

# Ubuntu
sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt update && sudo apt install nvidia-driver-550
sudo reboot

Verify after reboot: nvidia-smi should show driver 550+.

3. Missing NVIDIA Container Toolkit

# Install (Debian/Ubuntu)
curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
  | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
distribution=$(. /etc/os-release; echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/libnvidia-container/$distribution/libnvidia-container.list \
  | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
  | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit
sudo systemctl restart docker

If you can't upgrade the driver, the system will automatically fall back to CPU encoding. Your videos will still compress — just slower.

Cause: Reverse proxy buffering SSE responses.

Fix: Add proxy_buffering off; for the /api/stream/ location. See Reverse Proxy Configuration.

This is handled automatically. If the output exceeds the target by more than 2%, the system re-encodes with a reduced bitrate (up to 2 retries). You'll see a notification and hear an audio alert.

docker ps | grep 8mblocal              # Is it running?
docker logs 8mblocal                   # Startup and runtime logs
docker exec 8mblocal nvidia-smi        # GPU visible?
docker exec 8mblocal ffmpeg -hide_banner -encoders 2>&1 | grep nvenc  # NVENC available?
docker restart 8mblocal                # Restart
docker logs -f 8mblocal               # Live log tail

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)

You are free to use, share, and adapt this project for non-commercial purposes with appropriate attribution. Commercial use requires a separate license — please contact me directly.

Pull requests welcome! Please ensure Docker builds succeed and test with your GPU hardware.

For issues, questions, or feature requests, please open an issue on GitHub.