This document is relevant for: Inf2, Trn1, Trn2, Trn3

nrtpy configuration#

This page documents the module-level functions that configure and manage the nrtpy runtime singleton.

Overview#

nrtpy uses a single-runtime execution model: one Python process holds one runtime singleton managing one libnrt instance. The singleton is created lazily the first time you use an NrtpyTensor or NrtpyModel.

  • Call configure() before the first runtime operation to set which NeuronCores are visible.

  • Call reset() to close the runtime and clear configuration state. After reset(), you may call configure() again before the next operation.

Functions#

nrtpy.configure(visible_cores=None)#

Set runtime configuration before initialization. This only sets the NEURON_RT_VISIBLE_CORES environment variable — it does not initialize the runtime. The runtime is initialized lazily on the first nrtpy operation after configure().

Can only be called when the runtime is not active — either before the first nrtpy operation, or after reset() (which closes the runtime and clears state). Calling configure() is optional; if not called, the runtime uses the system default (all NeuronCores visible).

Parameters:

visible_cores (collections.abc.Iterable[int] or None) – Iterable of NeuronCore IDs, for example [0, 1, 2] or range(4). If None, visible cores are left unchanged (uses the current NEURON_RT_VISIBLE_CORES value or system default).

Raises:

  • RuntimeError – If the runtime is already active. Call reset() first.

  • TypeError – If visible_cores is an int rather than an iterable, or contains a non-integer value.

  • ValueError – If any core ID is negative.

nrtpy.reset()#

Close the current runtime and clear configuration state. If the runtime has not been initialized yet, this is a no-op. Call configure() afterward to set new visible cores before the next nrtpy operation.

Warning

All existing NrtpyTensor and NrtpyModel objects become invalid after this call. Any operations on them will raise errors.

Example#

import nrtpy
from nrtpy import NrtpyModel, NrtpyTensor

# Configure visible NeuronCores before first use
nrtpy.configure(visible_cores=[0, 1])

# The singleton is created lazily on first operation
model = NrtpyModel.load_from_neff("model.neff")
tensor = NrtpyTensor.from_numpy(data, name="input")

# Switch to different cores: reset, then reconfigure
del model, tensor
nrtpy.reset()
nrtpy.configure(visible_cores=[2, 3])

# New runtime uses cores 2 and 3
model2 = NrtpyModel.load_from_neff("model.neff", core_id=0)