July 4, 2026 · 15 min read

Loihi 2 PyTorch Without Lava

Intel archived Lava. Teams still need Loihi 2 oriented PyTorch workflows. This guide covers Loihi 2 PyTorch without Lava using NeuroCUDA: convert, simulate IF dynamics, export NIR, and know what still requires silicon.

For years, loihi 2 pytorch without lava was not a common search phrase because Lava was the assumed bridge between PyTorch-style development and Intel's neuromorphic stack. In 2026 the lava-nc repositories are archived. No public successor SDK ships yet. Teams with half-finished Lava projects and new PyTorch checkpoints need a path that does not depend on unmaintained Intel packaging.

NeuroCUDA is one answer - not a Lava drop-in, but a pip-installable pipeline from PyTorch to a Loihi 2 equation simulator plus GPU/CPU validation. This guide explains what that covers, what it does not, and how it compares to staying on archived Lava.

What Lava was and why archival matters

Intel Lava provided a Python framework for describing spiking processes, mapping them to Loihi hardware, and integrating with Lava-DL (SLAYER) for gradient-based training. When Intel archived Lava:

• Repositories remain readable but receive no updates or issue responses
• Dependency chains rot as Python and PyTorch evolve
• New Loihi-oriented projects lack an officially maintained onboarding path
• Documentation still references Lava in older tutorials and university labs

Full context: Intel Lava archived: alternatives guide. The question is no longer "how do I install Lava" but "what replaces the PyTorch-to-Loihi validation loop."

Loihi 2 PyTorch without Lava: the NeuroCUDA path

pip install neurocuda[all]

import neurocuda

# 1. Start from trained PyTorch (not Lava Process models)
snn = neurocuda.convert(pytorch_model, calibration_loader, timesteps=32)

# 2. Validate on GPU first (fast iteration)
neurocuda.compile(snn, target="gpu")
gpu_acc = neurocuda.evaluate(snn, test_loader)

# 3. Switch to Loihi 2 IF simulator
neurocuda.compile(snn, target="loihi2_sim")
loihi_sim_acc = neurocuda.evaluate(snn, test_loader)

# 4. Optional NIR export for ecosystem tooling
neurocuda.to_nir(snn, "model_loihi_ready.nir")

This is Loihi 2 PyTorch without Lava in practice: standard torch.nn in, spiking network out, Loihi IF dynamics in simulation. No lava.proc imports, no Lava message passing runtime, no SLAYER training loop required.

Lava vs NeuroCUDA comparison

What loihi2_sim validates

NeuroCUDA's Loihi 2 backend implements integrate-and-fire dynamics per Intel's published Loihi neuron specification. Validation protocol from the technical report:

• 100,000+ spike comparisons against reference equations
• Zero reported deviations in published tests
• Simulator only - not execution on Nahuku boards or Oheo Gulch systems

Use this backend to catch threshold and reset bugs before you invest in hardware porting. Do not cite simulator accuracy as Loihi silicon benchmarks. See Loihi 2 vs GPU energy for reporting discipline.

Recommended workflow in 2026

1. Train ANN in PyTorch using your existing ML stack (same as pre-Lava archival)
2. Convert with NeuroCUDA - QCFS + BPTT preserves accuracy (94.61% ResNet-18/CIFAR-10 published)
3. Validate GPU vs CPU - bit-exact spike parity builds confidence in the graph
4. Run loihi2_sim - equation-level Loihi IF check without Lava
5. Export NIR - hand off to Open Neuromorphic tooling if needed
6. Pursue silicon via INRC when Intel's next SDK or hardware path is clear

Step 6 is intentionally last. Most teams searching loihi 2 pytorch without lava need steps 1-5 today.

Who should still use archived Lava

Honest exceptions:

• You have a large production codebase already running on Lava and migration cost exceeds benefit short-term
• You need exact reproduction of a published Lava experiment for paper replication
• You have active INRC hardware access with tooling that still expects Lava graph formats

For greenfield PyTorch perception models, NeuroCUDA is the lower-friction path. For legacy Lava maintenance, read lava archived alternative before pinning old Python versions indefinitely.

Alternatives beyond NeuroCUDA

See also PyTorch to Loihi 2 guide and best neuromorphic compiler for the wider landscape.

Code example: ResNet on Loihi sim

import torch
import torchvision
import neurocuda
from torch.utils.data import DataLoader

model = torchvision.models.resnet18(num_classes=10)
model.load_state_dict(torch.load("resnet18_cifar10.pth"))
model.eval()

test_loader = DataLoader(...)  # CIFAR-10 test
cal_loader = DataLoader(...)   # calibration split

snn = neurocuda.convert(model, cal_loader, timesteps=32)
neurocuda.compile(snn, target="loihi2_sim")

acc = neurocuda.evaluate(snn, test_loader)
print(f"Loihi 2 sim accuracy: {acc:.2%}")  # expect ~94.61% per PDF

Full tutorial: ResNet-18 SNN conversion.

Common misconceptions

• "Without Lava means without Intel": You still align to Intel's neuron equations in sim; you skip the archived SDK.
• "Simulator equals chip": It does not. Energy, latency, and noise differ on silicon.
• "NeuroCUDA programs Loihi": It prepares and validates SNNs; chip programming remains Intel's domain.
• "I need Lava for NIR": NeuroCUDA exports NIR directly after conversion.

Robotics and ROS2 angle

Teams building neuromorphic perception on robots often paired Lava with custom bridges. NeuroCUDA ROS2 wraps the compiler in standard nodes so Loihi 2 PyTorch without Lava extends to Jetson and x86 deploy graphs without maintaining a fork of archived Intel packages. Event camera pipelines: DVS ROS2 guide.

Frequently asked questions

Start: pip install neurocuda · Lava archived alternative · Product page