Execution profiles

An experiment config says what is measured. Where an interpreter lives, which rendering library to load and where the model cache is differ from machine to machine, and must not change an experiment's identity. They live in the machine's execution profile instead (RFC-0005 part 1).

A profile (hylo.profile/v0.1, hylo.core.profile) is one YAML file per machine:

schema_version: hylo.profile/v0.1
runtimes:
  libero-sim:                                    # the LIBERO simulator
    python: /home/me/hylo/bench/envs/libero/.venv/bin/python
    env:
      MUJOCO_GL: osmesa
      PYOPENGL_PLATFORM: osmesa
      LD_LIBRARY_PATH: ~/.local/mesa/root/usr/lib/x86_64-linux-gnu
      LIBERO_CONFIG_PATH: /home/me/hylo/bench/envs/libero/config
      OMP_NUM_THREADS: '1'
      LP_NUM_THREADS: '2'
  lerobot:                                       # LeRobot policies (SmolVLA, pi0, ...)
    python: /home/me/hylo/bench/envs/libero/.venv/bin/python
    env:
      HF_HOME: /home/me/hylo/bench/envs/hf-home
      HF_HUB_OFFLINE: '1'
      TOKENIZERS_PARALLELISM: 'false'
store: /home/me/hylo/bench/runs           # the run and artifact store

It is found at $HYLO_PROFILE, else ~/.config/hylo/profile.yaml. A machine without a profile has no runtimes and uses ./runs as its store.

Runtimes

A runtime is a named interpreter plus environment variables: the venv a component runs in when it cannot share the main one (a simulator with pinned dependencies, a policy stack with its own torch, a robot SDK). Configs name the runtime and never a path:

env:
  type: libero
  remote: {runtime: libero-sim}                  # the environment runs in a subprocess with this runtime
agent:
  type: remote
  params:
    ref: lerobot.steered
    runtime: lerobot                             # the agent too
    params: {repo_id: lerobot/smolvla_libero}

The same config therefore runs on the workstation, the GPU machine and the robot computer, from any directory, and has the same digest on all of them: env.remote and a remote agent's runtime, python and env are left out of the digest.

Commands

uv run hylo profile show                                       # the profile and where it is
uv run hylo profile set-runtime robot --python ~/robot/.venv/bin/python --env HF_HOME=~/hf
uv run hylo profile remove-runtime robot
uv run hylo store init ~/hylo/runs                     # record this machine's store (see the store guide)

set-runtime adds or replaces a runtime. --python is stored as an absolute path without following symlinks, so a venv's interpreter keeps its venv. --env KEY=VALUE can be repeated; values are stored as given.

bench/envs/libero/setup.sh builds the LIBERO + LeRobot venv and registers both runtimes (libero-sim and lerobot) with the paths of the bench checkout it ran in. Running it again re-registers them.

Resolution rules

When a component starts in a runtime (hylo.core.profile.resolve):

1. the runtime named in the config is looked up in this machine's profile; an unknown name is an error that says which runtimes exist and how to register the missing one;
2. explicit python and env in the config, if any, override the runtime's (for a one-off; they are not measured either);
3. ~ and $VARS in the interpreter path and in every value are expanded when the component starts, not when the profile is written;
4. the subprocess gets the parent's environment plus these variables.

The store directory follows its own order: --root on hylo store, then $HYLO_STORE, then store: in the profile, then ./runs (store).

Machines in a lab

Machine	Typical runtimes	Store
workstation	libero-sim, lerobot	the main store, where campaigns and benchmarks write
GPU machine	lerobot (policy server, learners)	its own, synced with hylo store push/pull
robot computer	robot (the robot SDK venv)	where the workcell writes, pulled by the workstation

The profile is personal to each machine and is not committed anywhere. What is shared (workcell configs, safety limits) belongs in the deployments repository.