Lambda Azure Engine (LAE)

Lambda Azure Engine is an experimental sandbox for exploring ternary quantization, toy Mixture‑of‑Experts routing, and bit‑packed kernels. It is not a full inference engine and it does not run a 14B/200B model. Many numbers and terms in earlier docs were aspirational; this README describes what actually exists in the repo today.

What this repo contains

Python prototypes (lambda-azure-engine/python-bridge)

• lambda_engine.py: A toy “engine” that simulates routing and layer execution with random tensors.
• expert_paging.py: A mock NVMe → pinned‑memory → VRAM prefetch flow using numpy.memmap and CUDA streams.
• kv_cache_homology.py: A simple Union‑Find clustering demo over key vectors (L2 distance), used as a stand‑in for “KV compression.”
• quantize_to_ternary.py: Ternary Weight Networks (TWN) quantization and 2‑bit packing utilities.
• triton_ternary_gemm.py: Packed ternary GEMM with a CPU reference path and an optional Triton kernel.
• train_ternary_moe.py: A minimal MoE training stub on dummy data.
• integration_test.py: A small smoke test that exercises the toy engine.

Rust kernel (lambda-azure-engine/rust-kernel)

• Bit‑packed ternary math (lattice_compute.rs).
• Toy p‑adic weight helpers (p_adic.rs) and a simple “tilt” mapping (perfectoid.rs).
• Mock boundary/paging structures (boundary.rs, bulk_space.rs) that simulate large‑file access.
• Union‑Find for connected components (homology.rs).

What this repo does not contain

• A real 14B/200B model or a production MoE inference pipeline.

• End‑to‑end training, evaluation, or deployment tooling.

If you need a working large‑scale model, this repository is not that; it is a collection of small, illustrative experiments.

Quick start (toy demos)

These scripts are designed for experimentation and may require a CUDA‑capable system and torch. The Triton kernel is optional.

Suggested entry points:

• python lambda-azure-engine/python-bridge/kv_cache_homology.py
• python lambda-azure-engine/python-bridge/triton_ternary_gemm.py
• python lambda-azure-engine/python-bridge/quantize_to_ternary.py

Notes and caveats

• rust-kernel/src/main.rs attempts to create a very large file if you run it as‑is. Review it first if disk usage matters.
• The Colab notebook generator (build_colab.py) assumes large downloads and long‑running steps; treat it as a demo script, not a reproducible pipeline.

Directory layout

• lambda-azure-engine/python-bridge/: Python prototypes and demos.
• lambda-azure-engine/rust-kernel/: Rust crate with toy kernels and helpers.
• 200B_Scaling_Plan.md: Aspirational notes; not an implemented plan.