torch-pyodide 0.0.80

PyTorch-compatible module with autograd, interpolate/upsample, pooling 1D, loss modules, gradient clipping, nn.init, ModuleList, and float16/bfloat16 — via Pyodide + WebGPU.

torch-pyodide

PyTorch-compatible API that runs entirely in the browser. Built on Pyodide and WebGPU.

Try it now

Open the playground — write and run PyTorch code in your browser.

Install

In Pyodide (browser)

import micropip
await micropip.install("torch-pyodide")
import torch

x = torch.randn((3, 4))
w = torch.randn((4, 5))
y = x.matmul(w)
print(y.shape)  # (3, 5)

Important: in browser, the Python wheel alone is not enough.
Load and install the JavaScript runtime (runtime.mjs) first, then import torch.

Official manifest channel:

• https://celsowm.github.io/torch-pyodide/latest.json

The manifest provides:

• runtimeUrl
• wheelUrl
• runtimeSha256
• wheelSha256

Locally (with Python + Node.js)

pip install torch-pyodide
# Requires Node.js 20+ and a WebGPU-capable browser/device

What works

• Tensor creation: tensor(), zeros, ones, rand, randn, arange, full, empty
• Arithmetic: add, sub, mul, div, pow, matmul, mm, mv, bmm
• Linear algebra: dot, outer, norm (Frobenius, L1, L2, inf)
• Unary ops: relu, sigmoid, tanh, gelu, silu, sqrt, exp, log, sin, cos, and 40+ more
• Comparison: eq, ne, gt, lt, ge, le
• Reductions: sum, mean, max, min, prod, any, all, cumsum, cumprod
• Shape ops: reshape, flatten, squeeze, unsqueeze, transpose, permute, cat, stack, expand
• Indexing: select, slice, index_select, masked_select, masked_fill, where
• Neural network (torch.nn): Linear, Bilinear, Conv2d, BatchNorm1d/2d, LayerNorm, Dropout, pooling, loss functions, activations
• CUDA stub: torch.cuda.is_available(), torch.cuda.device_count(), etc.

License

MIT