Assets stay local
Datasets, model weights, locally pinned base models, checkpoints, and private configurations do not cross the coordination boundary.
Native training boundary
Your machine. Your data. Your models.
Neural Forge is built for users who want direct control over their data, models, checkpoints, and execution environment. The public page explains the boundary; the training environment remains under the user's control.
Coordination without custody
The Hub coordinates approved jobs without taking custody of the training environment. It receives bounded metadata, opaque identifiers, resource classes, and lineage—not model or dataset bytes.
What stays local
Execution belongs to the machine that runs it.
The native C++ training workflow includes CUDA and AMD/HIP integration. Runtime and hardware availability are version-specific; private inputs and detailed runtime state remain on the user's machine.
The worker owns runtime
Private kernels, device calls, filesystem paths, device details, optimizer state, and local execution settings remain with the training machine.
Detailed telemetry stays private
Raw process output and detailed telemetry remain local. Public surfaces expose only bounded, sanitized status when the approved entitlement and runtime contract allow it.
What the Hub coordinates
Policy travels as bounded metadata.
The Hidden Canopy Hub coordinates approved jobs through organization-scoped identity, opaque IDs, resource classes, and lineage. Local deployment bindings resolve those IDs into the actual models, datasets, binaries, checkpoints, and hardware available on the training machine.
Approved profiles Callers request a training profile, dataset, base model, checkpoint, resource class, evaluation request, and justification.
Local resolution The machine resolves approved identifiers through its own deployment bindings and private catalog.
Separated authority Evaluation and promotion remain distinct from worker execution; a completed run is not silently promoted.
Fail-closed execution
Ambiguity is a stop condition.
The workflow is designed to reject requests that would require guessing, authority expansion, or hidden access to private state.
Caller-controlled internals
Callers do not select local paths, private kernels, GPU ordinals, optimizers, or private training settings.
Incomplete requests
Invalid, ambiguous, stale, incomplete, or unauthorized requests fail rather than being silently corrected or expanded.
Three training paths
From-scratch training, locally pinned base models, and checkpoint resume all use the same bounded execution boundary.
AMD integration
A version-specific runtime boundary.
AMD work ran on MI300X in V2 and was ported into Foundry; retained packed-operation and native AdamW optimizer checks passed. This is not a full training-family parity claim. ROCm 7.2 is an experimental target. Current Foundry bridge validation uses offline tests; it does not establish a ROCm 7.2 native build, GPU run or performance result.
The result
A private, native, user-controlled training workflow with a narrow and auditable coordination boundary. Your machine. Your data. Your models. The Hub coordinates without taking custody.