Explains the v1 tied-output flag, omitted output tensor behavior, index routing, parameter reduction, and malformed-file rejection.
Implementation evidence: this topic is grounded in the reviewed GGUF.MiRust.com source snapshot. It documents observed code and artifacts without claiming broad deployment, model quality, or production readiness.
Flag contract
FLAG_TIED_OUTPUT = 1. When set, output.weight may be absent and logits use tok_embeddings.weight. Without the flag, output weight remains required.
Load-time routing
The top-level model index stores output_projection = tok_embeddings for tied models. Token generation then uses the same stable index path as an independent output tensor.
Fixture
The tied fixture has 11 tensors and 2,744 parameters instead of the ordinary tiny fixture’s 12 tensors and 4,824 parameters. Unit tests verify the missing tensor, index identity, and successful generation.
Compatibility requirement
A producer must only set the flag when embedding and output shapes and semantics are genuinely compatible. A loader must not infer tying from a missing tensor.
Scope
This starter page defines the questions, boundaries, evidence, and failure modes that should be recorded before a capability is presented as supported.
Engineering considerations
- Identify the source, version, target environment, and owner.
- Separate observed values from estimates and externally reported values.
- Record trade-offs, unsupported cases, and fallback behavior.
- Link performance statements to a compatible benchmark methodology.
Verification questions
- What exact artifact, revision, backend, and environment were reviewed?
- Which assumptions could change the result?
- Which data should be retained so another engineer can reproduce the conclusion?