optimum-benchmark
optimum-benchmark copied to clipboard
Process isolation is not working on MPS device
Setting up the launcher to process
on an MPS device causes the benchmark to be stuck permanently on “Sending report to main process.”
Steps to reproduce: Run this configuration on an MPS device (Mac M-series).
defaults:
- benchmark
- scenario: inference
- **launcher: process # launcher: inline works**
- backend: pytorch
- _base_
- _self_
name: pytorch_bert
scenario:
latency: true
memory: true
input_shapes:
batch_size: 1
sequence_length: 128
backend:
device: mps
no_weights: true
model: bert-base-uncased
(example adapted from pytorch_bert.yaml
)
Temporary Solution: set to launcher: inline
Remarks: launcher.start_method: fork
does not also work on mac