Package: Rllm 0.1.0

Rllm: Native Quantized Matrix Products and LLM Inference over 'Rfmalloc' Tensors

The composition layer of the 'Rfmalloc' ecosystem: it registers 'Rggml' (a vendored 'GGML' build with runtime-SIMD-dispatched quantized kernels) as a codec-aware matrix-multiply backend for 'Rfmalloc', so that products of file-backed, quantized tensors run natively in quantized space (weights stay 'Q4_K'/'Q6_K'/... encoded, activations are quantized on the fly) instead of being decoded to double first. Combined with 'Rgguf', which exposes 'GGUF' model weights as borrowed typed spans, this lets a larger-than-memory model's linear layers be multiplied through 'GGML's SIMD-accelerated dot kernels zero-copy from the original model mapping. Typed semantic programs written through an R module-and-pipe architecture language describe llama, Qwen3.5, LFM2MoE, EmbeddingGemma, ESM-2, OpenSpliceAI, Evo 2 and Tiny Recursive Models. Bound programs join those serializable graphs to mapped parameter storage. Implemented lowerings provide persistent state, pooled embeddings, byte-level generation, and CPU or optional CUDA execution; a reference interpreter executes the same dataflow through explicit R operator tables and structured recurrence.

Authors:Sounkou Mahamane Toure [aut, cre]