Context

The cookbook architecture-demos spec tracks Nemotron as status: blocked. Issue #228 ("QA: Nemotron-Nano-4B largely unsupported (12.3% pass rate)", closed 2026-02-13) characterized the failure surface; this issue is the prerequisite loader work.

Family

• Name: nemotron
• Vendor: NVIDIA
• HF architectures: NemotronForCausalLM
• HF pattern: nvidia/Nemotron-* (e.g., nvidia/Nemotron-Nano-4B, nvidia/Nemotron-4-340B-Instruct)
• Reference checkpoints: nvidia/Nemotron-Nano-4B, nvidia/Nemotron-4-15B

Acceptance criteria

• contracts/model-families/nemotron.yaml exists with size_variants for at least 4B
• Loader handles Nemotron-specific squared ReLU activation (mlp_bias, qk_layernorm)
• Discriminator field documented (Nemotron has mlp_bias: true + partial_rotary_factor)
• Inference smoke pass against Nemotron-Nano-4B improves the 12.3% pass rate from QA: Nemotron-Nano-4B largely unsupported (12.3% pass rate) #228

Unblock impact

• Cookbook manifest flips from blocked to certified
• Re-opens path to the Nemotron QA campaign (currently 12.3% — way below the QA campaign threshold)

Cookbook reference

• manifest.yaml entry — name: nemotron block
• Closed QA predecessor: aprender#228