zylora

Serverless GPU functions — deploy ML models with a decorator, scale to zero, pay per GPU-second.

Quickstart

pip install zylora

Define a GPU function

import zylora

@zylora.fn(gpu="H100")
def embed(text: str) -> list[float]:
    from transformers import AutoModel
    model = AutoModel.from_pretrained("BAAI/bge-large-en-v1.5")
    return model.encode(text).tolist()

Deploy

zy deploy

Invoke

# Local (development)
result = embed("hello world")

# Remote (deployed)
result = embed.remote("hello world")

# Batch
results = embed.map(["hello", "world", "foo"])

# Streaming (for LLMs)
for token in generate.stream("Tell me a joke"):
    print(token, end="", flush=True)

# Async
job = embed.remote_async("hello")
result = job.result()

Client API (without decorator)

from zylora import Zylora

zy = Zylora()
result = zy.invoke("embed", {"text": "hello"})

Authentication

Set your API key via environment variable:

export ZYLORA_API_KEY="zy_live_..."

Or use the CLI:

zy login

Configuration

@zylora.fn() options

Parameter	Type	Default	Description
gpu	str	required	GPU type: t4, l4, rtx4090, a100_40gb, a100_80gb, l40s, h100, h200, b200, mi300x
name	str	function name	Custom function name
packages	list[str]	[]	pip packages to install
model	str	None	Model to download during build
timeout	int	300	Execution timeout (seconds)
min_instances	int	0	Minimum warm instances
max_instances	int	10	Maximum concurrent instances
concurrency	int	1	Requests per instance
image	str	None	Custom base Docker image
runtime	str	"python312"	Python runtime version
routing	str	"cost_optimized"	Routing strategy
visibility	str	"private"	Function visibility

Documentation

Full documentation: docs.zylora.dev

License

MIT