mpsops 0.1.5

CUDA-only PyTorch operations for Apple Silicon - metapackage

MPS Ops

CUDA-only PyTorch operations for Apple Silicon.

This metapackage installs all mpsops packages - Metal implementations of operations that are typically CUDA-only.

Installation

Install everything:

pip install mpsops

Or install individual packages:

pip install mps-flash-attn    # Flash attention
pip install mps-bitsandbytes  # Quantized ops
pip install mps-deform-conv   # Deformable convolution
pip install mps-correlation   # Optical flow correlation
pip install mps-carafe        # Content-aware upsampling
pip install mps-conv3d        # 3D convolution for video

Packages

Package	Description	Use Case
mps-flash-attn	Flash Attention	Transformers, LLMs
mps-bitsandbytes	8-bit/4-bit quantization	LLM inference, QLoRA
mps-deform-conv	Deformable convolution	Object detection (DETR, DCN)
mps-correlation	Correlation layer	Optical flow (RAFT, PWC-Net)
mps-carafe	CARAFE upsampling	Segmentation (Mask R-CNN)
mps-conv3d	3D Convolution	Video models (I3D, SlowFast, MMAudio)

Quick Start

import mpsops

# Check what's installed
mpsops.print_status()

# Use the ops directly
from mpsops import flash_attn_func, deform_conv2d, correlation, carafe, conv3d, patch_conv3d

Compatibility

• PyTorch: 2.0+
• macOS: 12.0+ (Monterey)
• Hardware: Apple Silicon (M1/M2/M3/M4)

Why?

Many state-of-the-art models use CUDA-only operations:

• LLMs need flash attention and quantization
• Object detection needs deformable convolution
• Optical flow needs correlation layers
• Segmentation needs specialized upsampling

On Mac, you get errors like:

NotImplementedError: flash_attn not implemented for MPS

MPS Ops provides native Metal implementations so these models run on Apple Silicon.

License

MIT