PyTorch DPU Backend

A minimal PyTorch backend built on PrivateUse1, modeled after ascend/pytorch.

Minimum PyTorch version: 2.7.0

Build

# Install PyTorch first
pip install torch==2.7.0

# Generate codegen stubs
bash generate_code.sh python3

# Build and install (editable)
pip install -e . --no-build-isolation

A successful build produces torch_dpu/_C.*.so.

Quick Start

import torch
import torch_dpu

# Create tensors on DPU device
x = torch.ones(3, 3, device="dpu")
y = torch.empty(3, 3, device="dpu").fill_(2.0)

# Basic operations
z = x + y                     # addition
result = torch.cat([x, y])    # concatenation

Supported Operators

Category	Operators
Creation	empty, empty_strided
Math	add, isfinite
Comparison	eq, ne
Manipulation	view, as_strided, resize_, cat, fill_, set_
Copy	to, copy_
Access	item()

See docs/skills.md for adding new operators.

Troubleshooting

• Symbol not found errors: Use --no-build-isolation and remove old torch_dpu/_C.*.so artifacts before rebuilding
• Compilation fails: Report errors - do not proceed to verification