Module mps 

Experimental CUDA MPS (Multi-Process Service) support (GPU-SHARE §1.5).

MPS is not auto-started. This module provides opt-in setup for users who understand the risks:

• A GPU fault in any MPS client kills ALL clients on that GPU
• Thread percentage is static once set (no rebalancing)
• Jetson MPS is experimental below 30% thread allocation

Usage

apr finetune model.apr --vram 8 --experimental-mps --gpu-share 50

Structs

MpsConfig

MPS configuration for experimental GPU sharing.

MpsValidation

Result of MPS configuration validation.

Functions

is_mps_daemon_running

Check if MPS daemon appears to be running.

print_mps_warning

Print MPS safety warning to stderr.

setup_mps_env

Set MPS environment variables before CUDA context creation.

validate_mps_config

Validate MPS configuration for known issues.