SkyRL tx is an open-source library that implements a backend for the Tinker API, allowing you to set up your own Tinker-like service running on your own hardware. It provides a unified interface for both training and inference, enabling seamless online learning, cost-effective multi-tenancy through LoRA, and simplified ML infrastructure.

✨ Key Features

• Unified Training & Inference — Single engine for forward passes, backward passes, and sampling
• Multi-User LoRA Support — Efficient GPU sharing across users with individual adapters
• SFT & RL Support — Supervised fine-tuning and reinforcement learning with PPO and custom loss functions
• Multi-Node Training — FSDP and tensor parallelism for distributed training
• Multiple Model Architectures — Support for Qwen3 (dense & MoE), Llama 3, and DeepSeek V3
• External Inference Engine — Optional vLLM integration for optimized inference
• Production Ready — PostgreSQL support, cloud storage checkpoints, and database migrations

🏗️ Architecture

SkyRL tx consists of four main components:

┌─────────────────────────────────────────────────────────────────┐
│                        REST API Server                          │
│                    (FastAPI - handles requests)                 │
└─────────────────────────────────────────────────────────────────┘
                                 │
                                 ▼
┌─────────────────────────────────────────────────────────────────┐
│                           Database                              │
│         (SQLite/PostgreSQL - metadata, job queue)               │
└─────────────────────────────────────────────────────────────────┘
                                 │
                                 ▼
┌─────────────────────────────────────────────────────────────────┐
│                            Engine                               │
│        (Scheduling & batching across users/adapters)            │
└─────────────────────────────────────────────────────────────────┘
                                 │
                                 ▼
┌─────────────────────────────────────────────────────────────────┐
│                            Worker                               │
│       (Model execution, forward/backward, optimizer)            │
└─────────────────────────────────────────────────────────────────┘

🚀 Quick Start

Installation

git clone https://github.com/NovaSky-AI/SkyRL
cd SkyRL/skyrl-tx

# For GPU
uv run --extra gpu --extra tinker -m tx.tinker.api --base-model <model>

# For TPU
uv run --extra tpu --extra tinker -m tx.tinker.api --base-model <model>

Basic Training Example (Pig Latin)

Start the server:

uv run --extra gpu --extra tinker -m tx.tinker.api --base-model "Qwen/Qwen3-0.6B"

Run a simple training loop:

import tinker
import numpy as np
from tinker import types

# Connect to the local server
service_client = tinker.ServiceClient(base_url="http://localhost:8000", api_key="tml-dummy")
training_client = service_client.create_lora_training_client(base_model="Qwen/Qwen3-0.6B")
tokenizer = training_client.get_tokenizer()

# Training examples
examples = [
    {"input": "banana split", "output": "anana-bay plit-say"},
    {"input": "quantum physics", "output": "uantum-qay ysics-phay"},
    {"input": "coding wizard", "output": "oding-cay izard-way"},
]

def process_example(example, tokenizer):
    prompt = f"English: {example['input']}\nPig Latin:"
    prompt_tokens = tokenizer.encode(prompt, add_special_tokens=True)
    completion_tokens = tokenizer.encode(f" {example['output']}\n\n", add_special_tokens=False)

    tokens = prompt_tokens + completion_tokens
    weights = [0] * len(prompt_tokens) + [1] * len(completion_tokens)

    return types.Datum(
        model_input=types.ModelInput.from_ints(tokens=tokens[:-1]),
        loss_fn_inputs=dict(weights=weights[1:], target_tokens=tokens[1:])
    )

processed = [process_example(ex, tokenizer) for ex in examples]

# Training loop
for _ in range(6):
    fwdbwd = training_client.forward_backward(processed, "cross_entropy").result()
    training_client.optim_step(types.AdamParams(learning_rate=1e-4)).result()

    logprobs = np.concatenate([o['logprobs'].tolist() for o in fwdbwd.loss_fn_outputs])
    weights = np.concatenate([e.loss_fn_inputs['weights'].tolist() for e in processed])
    print(f"Loss: {-np.dot(logprobs, weights) / weights.sum():.4f}")

Sampling

# After training, create a sampling client
sampling_client = training_client.save_weights_and_get_sampling_client(name='my-model')

# Sample from the model
prompt = types.ModelInput.from_ints(tokenizer.encode("English: coffee break\nPig Latin:"))
params = types.SamplingParams(max_tokens=20, temperature=0.0)
result = sampling_client.sample(prompt=prompt, sampling_params=params, num_samples=8).result()

for i, seq in enumerate(result.sequences):
    print(f"{i}: {tokenizer.decode(seq.tokens)}")

📖 Usage Examples

Dense Model Training (Qwen3-8B on 8×H100)

# Start the server
uv run --extra gpu --extra tinker -m tx.tinker.api \
    --base-model Qwen/Qwen3-8B \
    --backend-config '{"max_lora_adapters": 2, "max_lora_rank": 1, "tensor_parallel_size": 8, "train_micro_batch_size": 1}'

# Run training (using tinker-cookbook)
export TINKER_API_KEY="tml-dummy"
uv run --with wandb --with tinker sl_loop.py \
    base_url=http://localhost:8000 \
    model_name=Qwen/Qwen3-8B lora_rank=1

MoE Model Training (Qwen/Qwen3-30B-A3B)

# Start the server
uv run --extra gpu --extra tinker -m tx.tinker.api \
    --base-model Qwen/Qwen3-30B-A3B \
    --backend-config '{"max_lora_adapters": 2, "max_lora_rank": 1, "tensor_parallel_size": 8, "train_micro_batch_size": 1, "shard_attention_heads": false}'

# Run training (using tinker-cookbook)
export TINKER_API_KEY="tml-dummy"
uv run --with wandb --with tinker sl_loop.py \
    base_url=http://localhost:8000 \
    model_name=Qwen/Qwen3-30B-A3B lora_rank=1 max_length=512

Reinforcement Learning (Qwen/Qwen3-8B)

# Start server
uv run --extra gpu --extra tinker -m tx.tinker.api \
    --base-model Qwen/Qwen3-8B \
    --backend-config '{"max_lora_adapters": 3, "max_lora_rank": 1, "tensor_parallel_size": 8, "train_micro_batch_size": 8, "sample_max_num_sequences": 256}' > out.log

# Run RL loop
uv run --with wandb --with tinker rl_loop.py \
    base_url=http://localhost:8000 \
    model_name="Qwen/Qwen3-8B" \
    lora_rank=1 max_length=1024

Multi-Node Training

# Node 0 (coordinator + API server)
CUDA_VISIBLE_DEVICES=0,1,2,3 uv run --extra gpu --extra tinker -m tx.tinker.api \
    --base-model Qwen/Qwen3-8B \
    --backend-config '{
        "max_lora_adapters": 3,
        "max_lora_rank": 1,
        "tensor_parallel_size": 4,
        "fully_sharded_data_parallel_size": 2,
        "train_micro_batch_size": 8,
        "sample_max_num_sequences": 256,
        "coordinator_address": "node0:7777",
        "num_processes": 2
    }' > out.log

# Node 1 (worker)
CUDA_VISIBLE_DEVICES=4,5,6,7 uv run --extra gpu --extra tinker -m tx.tinker.backends.jax \
    --coordinator-address "node0:7777" \
    --num-processes 2 \
    --process-id 1

With External vLLM Inference

# Start vLLM
VLLM_ALLOW_RUNTIME_LORA_UPDATING=True \
VLLM_PLUGINS=lora_filesystem_resolver \
VLLM_LORA_RESOLVER_CACHE_DIR=/tmp/lora_models/ \
CUDA_VISIBLE_DEVICES=4,5,6,7 uv run --with vllm vllm serve Qwen/Qwen3-4B \
    --tensor-parallel-size 4 --port 7999 --enable-lora

# Start SkyRL tx with external inference
CUDA_VISIBLE_DEVICES=0,1,2,3 uv run --extra gpu --extra tinker -m tx.tinker.api \
    --base-model Qwen/Qwen3-4B \
    --external-inference-url "http://0.0.0.0:7999" \
    --backend-config '{"max_lora_adapters": 3, "max_lora_rank": 1, "tensor_parallel_size": 4, "train_micro_batch_size": 8}' > out.log

🎯 Supported Features

🗺️ Roadmap

• Performance — Expert parallelism, context parallelism, optimized kernels
• Models — More architectures, PyTorch model definitions via torchax
• API Coverage — Full Tinker API compatibility
• Operations — Dashboard/frontend, improved logging and metrics
• Integration — SkyRL-train Tinkerification

🤝 Contributing

We welcome contributions! The project is early and hackable — now is a great time to get involved.

Ways to contribute:

• Try examples from the Tinker documentation or cookbook
• Fix issues or implement features from our issue tracker
• Improve documentation
• Add support for more models
• Performance optimizations

📚 Resources

• Ray Summit Talk — SkyRL tx: A unified training and inference engine
• Slides — Presentation slides
• Tinker Documentation — Official Tinker API docs
• Tinker Cookbook — Example recipes

📝 Blog Posts

• Introducing SkyRL tx
• SkyRL tx v0.0.2
• SkyRL tx v0.0.3
• SkyRL tx v0.1.0
• SkyRL tx v0.2.0
• SkyRL tx v0.2.1

📬 Contact

• Slack: #skyrl-tx
• GitHub: NovaSky-AI/SkyRL/skyrl-tx
• Twitter/X: @NovaSkyAI

📄 License

See LICENSE for details.