[orangedeng]

Refer to issue:

• Support serving decision models/APIs gpustack#6436

Parse the ".decision.type" metadata the llama.cpp conversion tools write, mirroring common_get_decision_type: the presence of the key alone marks a decision model serving "POST /v1/systemone", whatever its value says.

Expose it as GGUFArchitecture.DecisionType and a Decision flag on the llama.cpp estimates, plus a DECISION column in the CLI table. The declared marker outranks the reranking heuristics: general backbones such as modern-bert carry classifier tensors and rank pooling that would otherwise read as reranking.

Cover both with offline fixture tests and a first-party Laya GGUF from ModelScope.

[gpustack-code-review]

✅ OpenCodeReview: Review complete: 0 finding(s) across 5 selected item(s).

[Copilot]

🟡 Changes recommended

Detection misses valid markers, malformed metadata can panic, and the live-model test bypasses the required opt-in behavior.

3 open issues

• Validate decision marker metadata type before accessing it · New
• Use TEST_MODEL_PATH and skip unavailable model tests · New
• Detect decision models by marker presence regardless of attention type · New

What changed in this PR

Adds decision-model identification to the GGUF library and CLI, supporting GPUStack’s decision-model serving work in #6436.

Changes:

• Parses <architecture>.decision.type and exposes decision flags in llama.cpp estimates and summaries.
• Gives decision markers precedence over reranking heuristics and adds a CLI Decision column.
• Adds synthetic fixtures and a Laya model check, and documents decision-model screening.

File	Description
README.md	Adds Decision to the supported model-screening types.
file_estimate__llamacpp.go	Adds decision classification and propagates it to summaries.
file_decision_test.go	Tests metadata, serialization, classification, and Laya detection.
file_architecture.go	Reads and exposes decision-type metadata.
cmd/​gguf-parser/​main.go	Displays decision support in estimate tables.

_{Review effort: Balanced}

[thxCode]

LGTM