ignis-dl

Neural-net layers + transformers for Ruby, on the Ignis GPU/autograd stack. Installing this pulls the whole chain (ignis + ignis-autograd).

• NN: Linear, Embedding, LayerNorm, RMSNorm, Dropout; SGD/Adam/AdamW; losses.
• Transformers: multi-head + grouped-query attention, RoPE, SwiGLU, KV cache.
• Loaders: real GPT-2 and Llama (HuggingFace safetensors) — matches HF logits.

require "ignis-dl"
model = Ignis::AI::LlamaLoader.from_pretrained("path/to/Llama-3.2-1B")
# ... matches HuggingFace; "The capital of France is" => " Paris"

Loads real GPT-2 and Llama-3.2 (cosine 1.0 vs HuggingFace) and trains transformers from scratch. MIT.