mpsify

Run CUDA-hardcoded PyTorch repos on Apple Silicon (MPS) — with zero source edits.

You inherit someone's training script. It's full of .cuda(), device='cuda', torch.cuda.amp, map_location='cuda'. PyTorch-MPS could run the actual math fine on your Mac — but the code dies before it gets the chance. mpsify patches torch at import time so all of that transparently retargets to MPS.

pip install mpsify

Usage

Point it at any script — no edits to their code:

python -m mpsify train.py --epochs 10 --lr 1e-3
# or, after install, the console script:
mpsify train.py --epochs 10 --lr 1e-3

Or drop one line at the top of your entry file:

import mpsify  # patches torch on import

The wrapper (python -m mpsify) is preferred: it sets PYTORCH_ENABLE_MPS_FALLBACK=1 before torch loads, which the import form can't always guarantee.

What it does

Ops with no Metal kernel fall back to CPU automatically. mpsify catches those fallbacks, warns once per op live, and prints a summary at exit — so you can see exactly which ops are your latency hot spots.

Libraries with no Metal backend at all (bitsandbytes, apex, deepspeed, flash_attn, triton) are detected and reported loudly instead of crashing cryptically.

Pre-flight check

See what mpsify will do to a script without running it:

python -m mpsify doctor train.py

Reports every CUDA call it will remap and any library it can't fix (flash-attention, bitsandbytes, ...). Exit code is non-zero if something is unshimmable — handy in CI.

Loading a CUDA-trained checkpoint

Weights are just numbers — a model trained on CUDA runs on MPS unchanged. The only snag is torch.load restoring tensors to their saved cuda device. mpsify.load handles that (and upcasts fp16 → fp32, since MPS half-precision has rough edges):

import mpsify
state = mpsify.load("model_trained_on_a100.pth")   # lands on MPS
model.load_state_dict(state)

Diagnosing slow ops

python -m mpsify --profile train.py

Runs a dispatch-level profiler that counts calls and times ops. This adds per-op overhead — use it for a diagnostic pass, not production.

Knobs

Tested on

Real pretrained models, output verified numerically identical to CPU (max abs diff ~1e-6) on macOS + torch 2.5.1:

• torchvision resnet18, efficientnet_b0, vit_b_16
• timm mobilenetv3_small_100
• HuggingFace transformers — DistilBERT sentiment: .to('cuda'), forward, and pipeline(device=0) all run on MPS
• A 3-step fine-tuning loop (loss decreases, params on mps:0)
• Loading fp16/CUDA checkpoints via mpsify.load

Note: transformers 5.x refuses to torch.load a .bin checkpoint unless torch ≥ 2.6 (a CVE guard). Use models that ship safetensors (use_safetensors=True) — unrelated to mpsify.

Scope

Handles pure-PyTorch repos (torchvision / timm models — ResNet, EfficientNet, ViT, etc.). It does not translate custom .cu/Triton kernels or make CUDA-only libraries (flash-attention, DeepSpeed, apex) actually work — those are detected and reported, not fixed.

License

MIT