TSLM WebUI

TSLM WebUI is a Kotlin Multiplatform web interface and HTTP service for the TSLM torrent-title recognition model. It serves a Compose for Web frontend and a Ktor API from one JVM process.

Project Layout

.
├── tslm-java/      # Git submodule: TSLM inference library
└── tslm-server/    # Ktor server and Compose for Web client

tslm-java is a Gradle subproject and is required to build the server.

Requirements

• JDK 21
• Docker, for container builds
• NVIDIA driver and NVIDIA Container Toolkit, for CUDA inference
• A TSLM ONNX model

The GPU model is tslm-b-fp16.onnx; use the CPU model with TSLM_WEBUI_USE_CUDA=false when CUDA is unavailable.

Clone And Build

Clone with submodules:

git clone --recurse-submodules <repository-url>

For an existing checkout:

git submodule update --init --recursive

Build a local JVM distribution, including the production WebAssembly client:

./gradlew :tslm-server:installJvmDist

The runnable distribution is written to:

tslm-server/build/install/tslm-server-jvm/

Docker

The project provides production and development Docker build tasks.

# Builds the production image. HEAD must have a vMAJOR.MINOR.PATCH tag, or set
# docker.releaseVersion in gradle.properties.
./gradlew :tslm-server:buildDockerImage

# Builds the JVM distribution locally, then copies it into a runtime image.
./gradlew :tslm-server:buildDevDockerImage

Production images are published to GitHub Container Registry as ghcr.io/torrenkt/tslm-webui. Both runtime images use CUDA 12.9 with cuDNN and install Zulu 21.

Deploy A Public Instance

A public instance does not require PostgreSQL, a super token, or a database network. It requires Docker Compose, an NVIDIA driver, and the NVIDIA Container Toolkit. The model file must be available on the host.

Create a compose.yaml next to a writable data/ directory:

services:
  tslm-webui:
    image: ghcr.io/torrenkt/tslm-webui:v1
    ports:
      - "2156:2156"
    restart: always
    volumes:
      - ./data/model:/app/model:ro
      - ./data/nv-cache:/root/.nv/ComputeCache
    environment:
      TSLM_WEBUI_PORT: "2156"
      TSLM_WEBUI_PUBLIC_INSTANCE: "true"
      TSLM_WEBUI_USE_CUDA: "true"
      TSLM_WEBUI_FLAVOR: "TSLM_B" # or TSLM_C
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              capabilities: [gpu]
              count: 1

The nv-cache mount persists the CUDA driver JIT cache generated during model preheating and is recommended for faster subsequent starts.

Start the service:

docker compose up -d

Open http://<host>:2156. To update the image, run docker compose pull && docker compose up -d.

For CPU inference, remove the deploy.resources.reservations.devices section, set TSLM_WEBUI_USE_CUDA to "false", and mount a CPU-compatible model.

License

Copyright 2026 TorrenKt.

Licensed under the Apache License, Version 2.0.