Context

The cookbook architecture-demos spec tracks StarCoder2 as status: blocked. Issue #311 ("qualify: SafeTensors non-LLaMA architecture support (GPT-2, GPT-NeoX, OPT, StarCoder, BERT)", closed 2026-02-27) added SafeTensors handling for several non-LLaMA architectures but didn't ship a StarCoder2-specific loader YAML.

Family

• Name: starcoder2
• Vendor: BigCode
• HF architectures: Starcoder2ForCausalLM
• HF pattern: bigcode/starcoder2-*
• Reference checkpoints: bigcode/starcoder2-3b, bigcode/starcoder2-7b, bigcode/starcoder2-15b

Acceptance criteria

• contracts/model-families/starcoder2.yaml exists with size_variants for 3b, 7b, 15b
• Loader handles StarCoder2's grouped-query attention (n_kv_heads typically 1/4 of n_heads)
• Loader handles RoPE with rope_theta=1000000 (similar to Llama-3 long-context)
• Discriminator field documented (Starcoder2ForCausalLM architecture string + GQA pattern)
• Inference smoke pass against starcoder2-3b (smallest variant)

Unblock impact

• Cookbook manifest flips from blocked to certified
• Adds a code-completion family alongside CodeLlama (still alias-eligible) and Codestral / CodeGemma (alias-eligible via feat(rosetta): FamilyRegistry::detect_from_config_str + register_alias #1562)

Cookbook reference

• manifest.yaml entry — name: starcoder2 block
• Predecessor: aprender#311 (added SafeTensors but not the loader)