Kiln Documentation

Start here

Quickstart

Install Kiln, start Qwen3.5-4B, verify a chat response, and choose the next workflow.

This is the command-focused reference path. For platform-specific installers, screenshots, troubleshooting cues, and copy buttons, use the guided Quickstart.

1. Install

Download the current release for your accelerator from GitHub Releases, or build one backend explicitly from source.

# Example: Vulkan build
cargo build --release --locked -p kiln-server \
  --features vulkan --no-default-features --bin kiln

Do not combine accelerator feature sets in one binary unless the build documentation for that target says to do so.

2. Start Qwen3.5-4B

./target/release/kiln serve \
  --model-path /path/to/Qwen3.5-4B \
  --host 127.0.0.1 \
  --port 8420

Wait for readiness, then inspect the resolved backend rather than assuming the requested accelerator was selected:

curl -fsS http://127.0.0.1:8420/health | jq
curl -fsS http://127.0.0.1:8420/v1/config | jq '.accelerator_runtime'

3. Send one request

curl -fsS http://127.0.0.1:8420/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{
    "model": "Qwen3.5-4B",
    "messages": [{"role": "user", "content": "Say hello in five words."}],
    "temperature": 0,
    "max_tokens": 32
  }' | jq

Open http://127.0.0.1:8420/ui/ for the dashboard and playground.

4. Connect an OpenAI client

Point the client at http://127.0.0.1:8420/v1. A placeholder API key is accepted for clients that require one syntactically. Kiln does not authenticate requests itself, so keep the default loopback bind or put remote access behind an authenticated, trusted network boundary.

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8420/v1", api_key="local")
response = client.chat.completions.create(
    model="Qwen3.5-4B",
    messages=[{"role": "user", "content": "Give me one debugging rule."}],
    temperature=0,
)
print(response.choices[0].message.content)

5. Choose the next guide

  • OpenEnv training: run seed-matched stateful RL episodes and train directly from environment rewards.
  • GRPO: score completions, train, evaluate, and hot-swap an adapter.
  • Evals: build a suite and compare base versus adapter.
  • Configuration: set the small number of values most deployments need.
  • Troubleshooting: diagnose startup, loading, backend, request, and adapter failures.
  • HTTP API: browse serving, training, eval, and diagnostics endpoints.

Success criteria

You are done with the quickstart when:

  • /health is ready and names the backend you intended;
  • one chat request completes with non-empty content;
  • the dashboard loads;
  • the server log has no device-loss, allocation, or model-load error.

Throughput is a separate check. Use Benchmarks and reproduce the workload that matches your use case instead of treating one unlabeled tokens-per-second number as universal.