Context

The cookbook architecture-demos spec tracks Granite as status: blocked. Earlier QA work (#230, closed 2026-02-13) characterized failure patterns at the inference layer; this issue is for the prerequisite — adding the loader YAML so apr convert and aprender::rosetta::load_family can handle Granite checkpoints.

Family

• Name: granite
• Vendor: IBM
• HF architectures: GraniteForCausalLM
• HF pattern: ibm-granite/*
• Reference checkpoints: ibm-granite/granite-3.1-2b-instruct, ibm-granite/granite-3.1-8b-instruct

Acceptance criteria

• contracts/model-families/granite.yaml exists with size_variants for 2b and 8b
• Loader handles Granite-specific attention scaling (attention_multiplier) and embedding scaling (embedding_multiplier)
• Discriminator field documented (Granite shares Llama-shape but has the multiplier fields)
• Inference smoke pass against granite-3.1-2b resolves the failure pattern from QA: Granite-3.1-2B has 3-test failure pattern (89.3% pass rate) #230

Unblock impact

• Cookbook manifest flips from blocked to certified
• Re-opens path to closing the 89.3%-pass-rate Granite QA campaign documented in QA: Granite-3.1-2B has 3-test failure pattern (89.3% pass rate) #230

Cookbook reference

• manifest.yaml entry — name: granite block