https://github.com/zenwor/zwgrad

🦥 Lazy ML Framework
https://github.com/zenwor/zwgrad

dl framework ml python

Last synced: about 1 month ago
JSON representation

🦥 Lazy ML Framework

• Host: GitHub
• URL: https://github.com/zenwor/zwgrad
• Owner: zenwor
• License: mit
• Created: 2025-05-01T23:16:57.000Z (6 months ago)
• Default Branch: main
• Last Pushed: 2025-05-10T09:33:58.000Z (5 months ago)
• Last Synced: 2025-07-23T01:41:50.677Z (3 months ago)
• Topics: dl, framework, ml, python
• Language: Python
• Homepage:
• Size: 33.2 KB
• Stars: 5
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# zwgrad

I decided to implement my own ML framework!

The goals are simple:

- Make it train some simple models, such as ResNet

- CUDA support

- CUDA codegen, with kernel fusions detected from IR

- MLIR, in the long run

# Changelog

## (10.05.2025) `zwgrad` supports standard operations and Linear (`zwg.nn.Lin`) layer

`zwgrad` supports basic set of operations necessary to have a standard Linear layer, with (default) or without a bias:

```py

import zwgrad as zwg

import zwgrad.nn as nn

from loguru import logger

class Model(nn.Module):

    def __init__(self):

        self.lin1 = nn.Lin(4, 8)

        self.lin2 = nn.Lin(8, 16, bias=False)

    def fwd(self, x):  # Forward pass is defined within `fwd` method

        x = self.lin1(x)

        x = self.lin2(x)

        return x

x = zwg.ten([1.0, 2.0, 3.0, 4.0], req_grad=True) # Create a tensor via `zwg.ten` & track gradient

model = Model()  # Instantiate the model

y = model(x)  # Forward pass

logger.info(f"Graph: {y}")  # Get the computation graph

logger.info(f"y = {y.numpy()}")  # Get the exact forward pass value

z = y.sum()

z.bwd()  # Backpropagate

logger.info(f"x.grad = {x.grad}")

logger.info(f"y.grad = {y.grad}")

```