Context

The cookbook architecture-demos spec tracks tiny_starcoder_py as status: blocked because GPTBigCodeForCausalLM has no upstream loader. GPTBigCode is StarCoder-1's architecture (multi-query attention with a single shared K/V across all heads), distinct from StarCoder2 (which has grouped-query attention; tracked separately).

Family

• Name: tiny_starcoder_py (canonical loader: gptbigcode)
• Vendor: BigCode
• HF architectures: GPTBigCodeForCausalLM
• HF pattern: bigcode/tiny_starcoder_py, bigcode/starcoder (StarCoder-1)
• Reference checkpoints: bigcode/tiny_starcoder_py (164M, fits in CI smoke), bigcode/starcoderbase-1b

Acceptance criteria

• contracts/model-families/gptbigcode.yaml exists (or tiny_starcoder_py.yaml if scoped narrower)
• Loader handles GPTBigCode's MQA: single shared K/V across all attention heads (multi_query: true)
• Discriminator distinguishes GPTBigCode from GPT-2 (GPTBigCode has multi_query: true, GPT-2 has standard MHA)
• Inference smoke pass against tiny_starcoder_py (164M is small enough for fast CI)

Unblock impact

• Cookbook manifest flips from blocked to certified
• 164M variant is the smallest causal-LM in the supported set — useful for fast smoke tests

Cookbook reference

• manifest.yaml entry — name: tiny_starcoder_py block