mycelya-torch 0.1.3

Mycelya: PyTorch extension for transparent remote GPU execution on cloud infrastructure

Mycelya: Remote GPU Computing for PyTorch

Run your PyTorch code anywhere and power it with cloud GPUs. Mycelya integrates a remote GPU backend into PyTorch, allowing tensor operations to execute on cloud infrastructure with minimal code changes.

import torch
import mycelya_torch

# Create a remote machine with cloud GPU
machine = mycelya_torch.RemoteMachine("modal", "A100")
cuda_device = machine.device("cuda")

# Your existing PyTorch code just works
x = torch.randn(1000, 1000, device=cuda_device)
y = torch.randn(1000, 1000).to(cuda_device)  # Move tensor to remote GPU
result = x @ y  # Computed on remote A100!

# Transfer result back to local machine
result_local = result.cpu()
print(f"Result: {result_local}")

Supported GPUs (Modal)

8 GPU Types: T4, L4, A10G, A100, L40S, H100, H200, B200

Installation

Prerequisites

• Python 3.10+
• PyTorch 2.6+ (must be pre-installed)
• Modal account (free tier available)

Quick Install

# 1. Install PyTorch 2.6+ first
pip install --upgrade torch

# 2. Install mycelya-torch
pip install --no-build-isolation mycelya-torch

# Or install from source
pip install --no-build-isolation git+https://github.com/alyxya/mycelya-torch.git

# 3. Setup Modal if not setup yet
modal setup

Note: Modal is currently the only supported GPU cloud provider. Support for other providers (AWS, etc.) will be added in future releases.

Quick Start

MNIST Training

import torch
import torch.nn as nn
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
import mycelya_torch

# Setup remote GPU
machine = mycelya_torch.RemoteMachine("modal", "T4")
device = machine.device("cuda")

# Load MNIST data
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.1307,), (0.3081,))
])
train_data = datasets.MNIST("./data", train=True, download=True, transform=transform)
train_loader = DataLoader(train_data, batch_size=128, shuffle=True)

# Define model - all operations run on remote GPU
model = nn.Sequential(
    nn.Flatten(),
    nn.Linear(784, 128),
    nn.ReLU(),
    nn.Linear(128, 10)
).to(device)

optimizer = torch.optim.Adam(model.parameters(), lr=0.001)

# Train on remote GPU
for data, target in train_loader:
    data, target = data.to(device), target.to(device)
    optimizer.zero_grad()
    output = model(data)
    loss = nn.functional.cross_entropy(output, target)
    loss.backward()
    optimizer.step()

LLM Inference

import mycelya_torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Define remote functions for model loading and inference
@mycelya_torch.remote
def load_model(model_name: str):
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForCausalLM.from_pretrained(
        model_name, torch_dtype="auto", device_map="auto"
    )
    return model, tokenizer

@mycelya_torch.remote
def generate_text(model, tokenizer, prompt: str):
    messages = [{"role": "user", "content": prompt}]
    text = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

    generated_ids = model.generate(**model_inputs, max_new_tokens=512)
    output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
    return tokenizer.decode(output_ids, skip_special_tokens=True)

# Create remote machine with required packages
machine = mycelya_torch.RemoteMachine(
    "modal", "A100", pip_packages=["transformers", "accelerate"]
)

# Load model and generate text - all on remote GPU
model, tokenizer = load_model("Qwen/Qwen3-4B-Instruct-2507")
content = generate_text(model, tokenizer, "Explain quantum computing briefly.")
print("Response:", content)

Diffusion Model Inference

import torch
import mycelya_torch
from diffusers import DiffusionPipeline

# Define remote functions for pipeline loading and image generation
@mycelya_torch.remote
def load_pipeline(model_name: str):
    pipe = DiffusionPipeline.from_pretrained(model_name, torch_dtype=torch.bfloat16).to("cuda")
    return pipe

@mycelya_torch.remote
def generate_image(pipe, prompt: str, height: int, width: int, seed: int):
    image = pipe(
        prompt,
        height=height,
        width=width,
        num_inference_steps=50,
        true_cfg_scale=4.0,
        generator=torch.Generator(device="cuda").manual_seed(seed)
    ).images[0]
    return image

# Create remote machine with required packages
machine = mycelya_torch.RemoteMachine(
    "modal", "H100", pip_packages=["diffusers", "transformers", "accelerate"]
)

# Load pipeline and generate image - all on remote GPU
pipe = load_pipeline("Qwen/Qwen-Image")
image = generate_image(pipe, "A cat holding a sign that says hello world",
                       height=1024, width=1024, seed=0)
image.save("cat.png")

API Reference

RemoteMachine

# Create remote machine with cloud GPU
machine = mycelya_torch.RemoteMachine(
    "modal", "A100",
    gpu_count=1,                                  # 1-8 GPUs
    pip_packages=["transformers", "diffusers"],   # Pre-install for remote functions
    idle_timeout=300                              # Pause after 5 min inactivity
)
device = machine.device("cuda")

# Install packages dynamically
machine.pip_install("numpy")

# Pause to save costs, resume when needed
machine.pause()   # Offload state and stop compute
machine.resume()  # Restart and reload state

@remote Decorator

# Execute entire function remotely
@mycelya_torch.remote
def custom_function(x: torch.Tensor) -> torch.Tensor:
    return torch.relu(x @ x.T)

result = custom_function(x)  # Runs on remote GPU

# Async execution
@mycelya_torch.remote(run_async=True)
def async_function(x: torch.Tensor) -> torch.Tensor:
    return x @ x.T

future = async_function(x)
result = future.result()

License

AGPL-3.0-or-later - See LICENSE file for details.