Add SmolLM2-1.7B-Instruct support (Apache 2.0, 1.7B)

[stikves]

Summary

Add HuggingFace SmolLM2-1.7B-Instruct for on-device inference. Apache 2.0 licensed, standard Llama architecture.

Changes

• Add llama model_type to registry (maps to existing MistralForCausalLM)
• Add smollm2-1.7b-instruct preset (INT4 default)
• models/smollm2/README.md — Usage documentation

Architecture

SmolLM2 uses standard Llama architecture — no new model code needed:

• 24 layers, 2048 hidden, MHA (32/32 heads), head_dim=64
• SiLU MLP, RMSNorm, standard RoPE (theta=130000)
• tie_word_embeddings = True
• Vocabulary: 49K tokens

Performance (Apple Silicon, INT4)

Metric	Value
Prefill	~502 t/s
Generation	~176 t/s
Model Size	~1 GB

Export

uv run coreai.llm.export HuggingFaceTB/SmolLM2-1.7B-Instruct

Test plan

• Model loads via existing MistralForCausalLM class
• Export completes (INT4)
• E2E inference produces correct output ("The capital of Japan is Tokyo")
• 176 t/s generation speed