genv

Fixed

• genv remote llm supports multiple Linux users

Fixed

• genv llm supports multiple Linux users

Added

• Introduced LLM support in Genv
• Added command genv llm
• Added command genv remote llm
• Added command genv version

Fixed

• Passing environment variables over SSH commands

Added

• Ray integration and genv.ray subpackage

Added

• Added flag --max-devices-for-user to genv enforce and genv remote enforce commands

Changed

• Renamed Python CLI from genvctl to genv
• Installation is now 100% Python-based
• Moved Python source directory to the project root

Removed

• Removed shell code base libexec/ entirely

Added

• Added genv.sdk.activate
• Added genv.sdk.attach
• Added genv.sdk.attached
• Added genv.sdk.configure
• Added genv.sdk.configuration

Changed

• Merged executables into genvctl

Added

• Introducing the Genv CLI genvctl
• Added CLI subcommand genvctl lock for locking over-subscribed devices as access control
• Supporting over-subscription in container toolkit with genv-docker flag --over-subscribe
• Added genv.sdk.env SDK package for the active environment
• Supporting locking multiple devices in genv.core.devices.lock

Changed

• Renamed genv.sdk.lock_devices to genv.sdk.lock

Fixed

• Fixed genv lock

Added

• Using global lock around critical sections instead of locking per module
• Enriched core entities with more fields
• Python package does not run executables in subprocesses anymore
• Supporting cleaning up entities in place with .cleanup()
• Bug fix in serialization of Report objects

Changed

• Major restructure in project directory
• Renamed Snapshot entities
• Renamed genv.env to genv.sdk

Removed

• Removed methods from genv.envs and genv.devices

Added

• Over-subscribe devices with new flag -o --over-subscribe to genv attach
• New Python SDK at genv.env
• Control access to over-subscribed devices with genv lock and genv.env.lock_devices()
• Added fields and actions to genv.devices.Device and genv.devices.Snapshot

Changed

• Renamed previous command genv-devices query to genv-devices find
• Using genv.devices.Snapshot instead of plain JSON in genv-devices and devices.json (backwards compatible)

Removed

• Removed environment variable GENV_ALLOW_DEVICE_OVER_ALLOCATION

Added

• Introduced Genv container toolkit: genv-docker and the Genv container runtime
• Flags --count and --index are optional in genv-devices attach; if none passed, genv-devices attach uses the configured device count if set
• Added genv.devices.attach()
• Added genv.envs.gpus()
• Added genv.envs.activate()
• Added genv.envs.configure()

Changed

• nvidia-smi shim prints warning message when missing information about processes
• nvidia-smi shim supports the case when environment variable CUDA_VISIBLE_DEVICES is not set
• nvidia-smi shim fails if no other nvidia-smi executable found

Added

• Added monitoring features with Prometheus and Grafana using genv monitor and genv remote monitor

Added

• Added enforcement rule env-devices
• Added enforcement rule env-memory
• Added command genv remote query
• Added flag -t --timeout to genv remote to set SSH connection timeout
• Added flag -e --exit-on-error to genv remote to exit on SSH connection issues
• Added flag -q --quiet to genv remote to ignore SSH connection issues
• Added flag --no-prompt to genv remote activate to not change shell prompt
• Ignoring commented lines in hostfile used in genv remote
• Added query uid to genv-envs query
• Set up Google Analytics for documentation site

Changed

• genv remote does not exit on SSH connection issues by default
• Refactor entities and snapshots in genv Python package
• Refactor remote capabilities in genv.remote Python subpackage
• Combined enforcement rules code under genv.enforce.rules
• Improved development setup and in particular nvidia-smi development shim and CPU-only setup for remote features

Added

• Introduced local and remote enforcement features with two enforcement rules: non environment processes and max devices per user
• Added genv-usage executable for taking snapshots and executing enforcement reports
• Supporting querying environment usernames
• Added flag --quiet to genv-devices detach
• Created devel directory and nvidia-smi mock shim
• Added environment variable GENV_TERMINATE_PROCESSES to allow not terminating processes
• Added environment variable GENV_MOCK_NVIDIA_SMI_PIDS to set process identifiers in the nvidia-smi mock shim

Changed

• Created Python package genv
• Major refactor to Python code by adding logic layers (e.g. genv.envs) and entities (e.g. genv.envs.Env)
• Supporting encoding and decoding entities as JSON
• Supporting sending standard input to SSH commands
• Supporting running SSH commands with sudo

Added

• Listing active environments on remote hosts with genv remote envs
• Showing device information on remote hosts with genv remote devices
• Activating an environment on a remote host with genv remote activate

Changed

• Formatting Python code with black
• Linting Python code with flake8

Added

• Configuring environment GPU memory capacity
• GPU memory aware device provisioning
• Supporting device over allocation with multiple environments
• Documentation site

Changed

• docker shim injects environment variable GENV_ENVIRONMENT_ID to containers
• Renamed environment variable GENV_DEVICES to GENV_MOCK_DEVICE_COUNT

Fixed

• nvidia-smi shim supports environment variables with =
• docker shim supports the case when argument --gpus is not passed
• nvidia-smi shim does not pass argument --id in bypass mode

Features

• nvidia-smi shim shows only information relevant to the environment (i.e. device memory, processes)
• Added docker shim to expose containers to devices attached to the environment