mila-llm

The LLM stack you can read and understand.

Mila is a C++23/CUDA runtime for open large language models — inference and training, built from explicit neural-network components. Device and precision are compile-time decisions, every forward pass is explicit, and there is no hidden execution engine. This package is its Python projection.

pip install mila-llm

import mila

mila.initialize("warning")

# Once: fetch a published model into the local store (~6.8 GB).
mila.ModelStore().pull("gemma-4-12b-it-fp4", mila.default_hub_owner())

tokenizer = mila.BpeTokenizer.from_store("gemma-4-12b-it-fp4")
model = mila.GemmaModel.from_store("gemma-4-12b-it-fp4", 4096)

reason = model.generate(tokenizer.encode(prompt), print)

A model is named, not pathed. from_store reads the local store's record, which is what knows the weights are already FP4 — so nothing pairs a weights path with a tokenizer path, and nothing has to be told what the bytes are. Pull and load are separate verbs: a load never reaches the network, so an uninstalled name is an error rather than a surprise download.

generate hands each token to the callback as it is produced and returns why it stopped — stop, length, context_limit or cancelled — which the tokens themselves cannot tell you.

The GIL is released around generation, so the callback runs on a live interpreter and StopController cancels a decode loop already in flight.

Requirements

An NVIDIA GPU. The CUDA runtime libraries arrive as dependencies (nvidia-cublas, nvidia-curand, and nvidia-cuda-runtime on Linux, where the extension links cudart dynamically) — no CUDA Toolkit installation is required. An installed Toolkit is used as a fallback if those are absent.

gemma-4-12b-it-fp4 wants roughly 12 GB of VRAM at a 4096 context; a Llama 3.2 3B is the smaller first run.

What it exposes

The store is shared with Mila's chat harness and inference server, so a model installed by any of them is loadable by all of them.

What it does not

Stated because the limits are documentation, not an omission from it.

• No weights are bundled. They are fetched on request into a local store, over a transport this package supplies from the standard library — the wheel carries no HTTP client of its own.
• A load never downloads. pull and from_store are separate calls on purpose.
• No GPT-2. It exists in the C++ library and is not bound.
• A published model's quantization is fixed — its bytes are already FP4 or FP8. Choosing a quantization applies only to unquantized weights loaded by path.
• No training, no batching, and a model instance is not thread-safe: serialize calls through a single worker thread.
• Text in, text out. No embeddings, logits, or hidden-state access.

Links

• Documentation: https://mila.toddt.me
• Source: https://github.com/toddthomson/Mila
• Models: https://huggingface.co/mila-llm — what pull fetches from