Skip to content

Decision record · ADR

ADR-0046: Carry auxiliary weight artifacts as governed-kind server slots

Status: accepted | Date: 2026-09-04

References: RFC-0002, RFC-0003, RFC-0006, ADR-0042

Real serving launches increasingly consume more than one weight artifact. Speculative decoding deployments such as DSpark load draft weights that are independent of the main model, and the same shape appears across mtp, eagle3, and dflash; framework spellings differ (the vLLM speculative-config JSON model key, the SGLang draft-model path option, the TensorRT-LLM draft model directory). Today the only way to express such a launch in InferLab is to hardcode the auxiliary weight path inside framework escape-hatch arguments. That path bypasses per-machine model_weights resolution (RFC-0002:C-LOCAL-PLACEMENT), so the same server cannot be placed across machines with different draft-weight locations; it leaves no typed evidence in dry-run output or serve records; and it is invisible to validation until the framework fails at startup.

ADR-0042 deliberately kept draft-model locators out of the serving contract because the demonstrated need at the time was only a controlled acceptance length. The demonstrated need now exists. What ADR-0042 rejected permanently still stands: the speculative method vocabulary must not enter the contract, and framework-specific spelling remains owned by the integrations (RFC-0006:C-INTEGRATIONS).

Constraints shaping the choice: planning receives logical model identity only while rendering receives machine-resolved locators; cases must not change the selected model or model locator (RFC-0003:C-RECIPE-COMPOSITION); portable outputs must exclude weight locators by construction; and the repository keeps one authority per fact rather than introducing a second configuration channel.

Server definitions gain an auxiliary_models declaration (RFC-0003:C-SERVE-AUXILIARY-MODELS): a map from a governed auxiliary-weight kind to an ordinary workspace model identifier. The kind names the role of the weight artifact, not a speculative method — the initial vocabulary contains exactly draft-model, the separate weights a framework-owned speculative decoding configuration drafts with, and a new kind enters the vocabulary only with a demonstrated workflow that requires it.

The referenced identifier selects an ordinary workspace model entry, and its weights resolve through the standing model-weight bindings unchanged — shared fallback locator plus per-machine locators, with the same Machine/Fallback provenance. No parallel resolution machinery is introduced. The declaration is server-level only; cases do not declare or override auxiliary models, consistent with the standing case invariance over model and model locator.

The adapter protocol hard-cuts to a new version. This is forced, not preferred: ADR-0042 extended version 9 in place because it was still unpublished, and version 9 has since shipped with product 0.13.2, so in-place extension is no longer available. The new version carries auxiliary weight artifacts in two places mirroring the main model split: planning receives the logical identities (kind, model id, served name), and each model-rank rendering allocation receives the machine-resolved locator per kind. Each integration consumes the kinds it supports and splices the resolved locator into its own framework spelling; planning fails with a typed error when a declared kind is not consumed by the selected integration, and likewise when the operator’s framework configuration already spells an artifact a declared kind covers. Dry-run and serve-record evidence preserve the declared mapping and each effective auxiliary locator with its kind and resolution provenance; portable outputs continue to exclude weight locators by construction.

Positive: a launch consuming several independent weight artifacts becomes expressible with per-machine resolution, planning-time validation, and typed evidence, while the speculative method vocabulary stays framework-owned. The design mirrors the synthetic-acceptance overlay split — InferLab supplies a governed value, the integration owns the spelling — so operators learn one pattern. Governed kinds make every consumed artifact role auditable in serve records.

Negative: the adapter protocol hard-cuts again, so the adapter SDK, framework integrations, and workspace pins move in lockstep and older integrations reject the new control plane. Every new artifact kind is an RFC amendment — deliberate friction that could read as ceremony if kinds proliferate. InferLab still cannot verify that the framework actually honors the spliced path; a wrong-but-well-formed auxiliary locator fails at framework startup, the same validation horizon as the synthetic-acceptance overlay. The escape-hatch path remains possible and ungoverned — an operator can still splice a literal draft path into framework arguments with no resolution, validation, or evidence; that is the pre-existing state, not a regression, and mixing both paths for the same artifact is rejected at planning so the declaration stays the single authority.

Neutral: draft-model composes with but does not require synthetic acceptance — a run with a real draft model and a synthetic overlay remains the AgentX qualification shape, while a real acceptance measurement simply omits the overlay. Separate-process components with their own devices remain a topology concern outside this mechanism. Case-level auxiliary declarations are not offered; a demonstrated per-case need would amend the RFC.

Carry auxiliary weight artifacts as governed-kind server slots that reference ordinary workspace models and resolve through the standing model-weight bindings (accepted)

Section titled “Carry auxiliary weight artifacts as governed-kind server slots that reference ordinary workspace models and resolve through the standing model-weight bindings (accepted)”
  • Pros: Reuses the models registry and per-machine resolution; no parallel machinery, Planning-time rejection of unknown or unconsumed kinds; typed evidence names the artifact role, The governed vocabulary is artifact-role-shaped, not method-shaped, so it does not reopen the moving speculative-method vocabulary
  • Cons: Each new artifact kind is a deliberate governance action, Requires an adapter protocol hard cut with SDK, integration, and workspace pins moving in lockstep

Free-form operator-chosen slot names with per-integration naming conventions (rejected)

Section titled “Free-form operator-chosen slot names with per-integration naming conventions (rejected)”
  • Pros: No governance step when a new artifact kind appears
  • Cons: Misspellings and divergent conventions surface only at framework startup, not at planning, The contract cannot validate or record which artifact role was consumed; per-integration conventions drift into a de facto contract without review

Interpolate weight references inside framework escape-hatch arguments (rejected)

Section titled “Interpolate weight references inside framework escape-hatch arguments (rejected)”
  • Pros: No protocol or schema change
  • Cons: String templating becomes a second configuration channel; typed validation, per-machine resolution, and value-level evidence are lost

Model separate-process components such as a standalone vision encoder as topology roles with their own model (rejected)

Section titled “Model separate-process components such as a standalone vision encoder as topology roles with their own model (rejected)”
  • Pros: Covers components that own devices and placement
  • Cons: No demonstrated separate-process workflow; that need is a topology feature, not an auxiliary-weights feature, and can arrive later without changing this mechanism