DocumentationArchitecture

SaveWeights

A zoomed-in view of one operation in Piro’s stateful inference loop.

Architecture detail

SaveWeights

Writes consolidated durable snapshots without rewriting unchanged durable components; external state storage is outside this model method.

← Full model

SaveWeights(weights:
    changed = DiffAgainstLoadedManifest(weights)
    if changed is empty:
        return NoOp("nothing changed")
    revision = ReadCurrentRevision() + 1
    for each group in changed:
        if group.owner == "plasticity.fastOverlay":
            component = Encode(group, format = BF16, scales = BF16)
        else if group.owner == "durable.base":
            component = Quantize(group, format = INT4, scales = BF16)
        else:
            component = Encode(group, format = BF16)
        R2Put("models/{modelId}/weights/revisions/{revision}/" + component.name, component.bytes)
    manifest = BuildManifest(revision, changed, parentRevision)
    R2Put("models/{modelId}/weights/revisions/{revision}/manifest.json", manifest)
    R2Put("models/{modelId}/weights/current/manifest.json", manifest)
    return revision

What it does

Diffs the active runtime object against the loaded manifest, encodes changed fast overlays as BF16, encodes dynamic state as BF16, re-quantizes durable base changes as INT4 only during consolidation, and publishes a versioned R2 manifest. Multipart upload is available for large or resumable component transfers.

Why it exists

Plasticity should not rewrite the ~230M-parameter base object on every interaction; R2's storage capacity is not the constraint here, while write volume and operation cost are.

← Back to architecture