Architecture
Generated with Grok Build: Grok 4.5 · xAI Imagine (/imagine)
A CPU-first numerical + MoE routing kernel:
Tensor / ops ──► Transformer blocks / LM
──► MoE router (GGUF bridge + gate scores + top-k)
| Area | Detail |
|---|---|
| Tensor math | Row-major f32, matmul, norms, mask, embedding |
| Transformer | Attention, residual block, decoder LM |
| MoE routing | Gate scores, softmax, top-k, L2, resample, sim modes |
| Family adapters | OLMoE, Qwen3-MoE, Gemma4, DeepSeek2, Llama-MoE |
| Dequant to f32 | Q8_0, Q5_K, F16 (supported set) |
| Consumer GGUF bridge | mmap access + embedding extraction (transitional) |
| Area | Owner / note |
|---|---|
| Canonical GGUF v3 parse + expert raw extract |
engram-parser (#8) |
| Safetensors manifests / MoE discovery | planned safetensors-parser (#9) |
| CUDA / GPU execution | Never — consume Tensor elsewhere |
| Autograd | Non-goal |
| Tokenization | Pair with external tokenizer |
| SNN dynamics | neuromod |
| ANN→SNN orchestration | hybrid-fusion |
GPU backends (cust), frameworks (candle, tch, ort), domain orchestration crates, and any dependency on rmems/corinth-canal. Extraction is a one-way copy.
Surgically extracted from a larger hybrid codebase (corinth-canal / historical SpikeLMo lineage), then cleaned to pure Rust.
Last updated: July 29, 2026
Updated by: Grok Build: Grok 4.5
Package tip reference: 37c5a21 (main)