RVC-Project/Retrieval-based-Voice-Conversion-WebUI

MacOS | MPS | torch metal work on Train -> step-1 then Step-2 but not on training the model (Step-3)

Open

#991 opened on Aug 12, 2023

View on GitHub
 (0 comments) (0 reactions) (0 assignees)Python (2,849 forks)batch import
help wanted

Repository metrics

Stars
 (18,427 stars)
PR merge metrics
 (No merged PRs in 30d)

Description

Hi is there a way to make the model train on torch metal, mps, even when passing the "mps" manually to the G and D models, it still uses the cpu, I also monitored the GPU performence, it peaks only on the first two steps, but on the model training it peaks the CPU any idea? and it stuck sometimes on "INFO:torch.nn.parallel.distributed:Reducer buckets have been rebuilt in this iteration." for a long time, I was trying to train on a 3 minutes speech, it took more than 5hrs for 20 epoch Thanks is advance

Contributor guide

Research direction
Check how models and data are moved to MPS device in training step. Ensure all tensors and model parameters are on the same device. Investigate if any part of the training loop forces CPU fallback. Look for environment variables like PYTORCH MPS HIGH WATERMARK RATIO and adjust batch size to avoid memory issues that cause fallback to CPU. Verify that the reducer bucket rebuild message indicates distributed training issues; ensure single GPU training is not using distributed wrappers inadvertently.
Tech stack
pythonpytorch
Domain
machine learningbackend
Issue type
Bug
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
Half day
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Mostly clear
Prerequisites
PythonPyTorchMPS backend knowledge
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
60