Flux support inside MPIExecutor, plus FluxExecutor

[jlnav]

Addresses #1647

Credit to Claude, as usual

Flux Usage Examples

Option 1: MPIExecutor with flux run wrapper

This wraps flux run as a subprocess, similar to how srun or mpirun work:

from libensemble.executors.mpi_executor import MPIExecutor

# Auto-detect Flux (works when FLUX_URI is set in environment)
exctr = MPIExecutor()

# Or explicit configuration
exctr = MPIExecutor(custom_info={"mpi_runner": "flux"})
exctr.register_app(full_path="/path/to/app.x", app_name="my_app")

# In sim function:
task = exctr.submit(app_name="my_app", num_procs=4, num_nodes=2)
task.wait()

Option 2: Native FluxExecutor

Uses Flux Python API directly — no subprocess spawning of flux run:

from libensemble.executors.flux_executor import FluxExecutor

exctr = FluxExecutor()  # Requires FLUX_URI to be set
exctr.register_app(full_path="/path/to/app.x", app_name="my_app")

# In sim function - submits via flux.job.submit() Python API
task = exctr.submit(app_name="my_app", num_procs=4, num_nodes=2)
task.wait()

Option 3: Container Workflow (FluxExecutor recommended)

This is the key use case mentioned in the issue — running libEnsemble inside a container where srun/mpiexec aren't available but Flux manages the resources from outside:

# Inside container (e.g., podman-hpc, Singularity, or Docker)
# Flux instance runs on the host and FLUX_URI is passed into the container

from libensemble.executors.flux_executor import FluxExecutor
from libensemble import Ensemble
from libensemble.specs import LibeSpecs, SimSpecs, GenSpecs

# FluxExecutor uses the Flux Python API to submit jobs
# No need for mpiexec/srun inside the container!
exctr = FluxExecutor()
exctr.register_app(full_path="/app/simulation.x", app_name="sim")

def sim_f(H, persis_info, sim_specs, libE_info):
    """Simulation function running inside container"""
    x = H["x"][0]

    # Submit work to Flux - Flux handles launching on compute nodes
    # even though we're inside a container
    task = exctr.submit(
        app_name="sim",
        app_args=f"--input {x[0]} {x[1]}",
        num_procs=4,
        num_nodes=1,
        num_gpus=2,  # Flux handles GPU assignment
    )
    task.wait()

    # Parse output...
    result = parse_output(task.read_stdout())

    H_out = np.zeros(1, dtype=sim_specs["out"])
    H_out["f"] = result
    return H_out, persis_info

# Configure libEnsemble for container environment
libE_specs = LibeSpecs(
    nworkers=4,
    comms="local",
    platform="flux",  # Use Flux platform configuration
)

# ... rest of ensemble setup

• Flux's Python bindings communicate directly with the Flux instance via FLUX_URI
• Flux running on the host handles all resource allocation and job launching
• Works with podman-hpc, Singularity, Docker, or any container runtime that can pass through the Flux socket