Any standard GGUF file
A zero-copy, no-throw GGUF v2/v3 parser over a memory mapping. Weights are viewed in their storage format (F16, BF16, Q4_0, Q5_0, Q8_0, Q4_K, Q6_K) and never copied.
Synapse
GGUF models from the GGML ecosystem, run as static, allocation-free compute graphs on MTL5 kernels, with the logits llama.cpp computes.
What Synapse is
Section titled “What Synapse is”Synapse is the ML runtime of the Branes.AI platform: where signals are transmitted and computed. Cortex decides what to compute (a vision-language-action policy, a depth network, a learned motion model), Synapse turns that model into a static compute graph and runs it, and Reflex consumes the results inside control loops that must never wait on an allocation.
Synapse reimplements the ML-environment half of the GGML ecosystem in modern
C++20 and delegates the math to MTL5. It
reads .gguf files unchanged, without carrying GGML’s C code, void* dispatch,
or switch (ggml_type) kernels (the GGML ecosystem).
No allocation during inference
A liveness pass plans every activation into one arena before the first run. Decode and prefill then allocate nothing and throw nothing.
Checked against llama.cpp
Dequantizers bit-exact against gguf-py, kernels against ggml’s, tokenizers
against llama_tokenize, and logits against llama.cpp on Llama, Qwen2 and
Mistral models.
Typed, not switched
Element and quantization types are part of the graph’s node types and pick MTL5 kernels by concepts and traits, so a mismatch is a compile error.
The platform at a glance
Section titled “The platform at a glance”| Repo | Analogue | Responsibility |
|---|---|---|
cortex | Cerebral cortex | Perception (VIO, SLAM, SfM), vision-language-action models, world building, path planning |
synapse | The synapses | Tensor storage, graph construction, memory planning, and execution of ML models |
reflex | Brainstem, spinal cord, autonomic system | Hard real-time control and safety: stabilization, state propagation, MPC, fault recovery |
anatomy | The body plan | The mission framework: data contracts, slots, configs, replay log, harness, simulator |
Where to go next
Section titled “Where to go next”- Introduction: the platform, the design, and the components.
- Build & Test: build, test, run a model, and use synapse from another project.
- The GGML ecosystem: GGUF, ggml, libllama and the conversion tools, and which synapse component corresponds to each.
- What a Llama model computes: the computation step by step, why it is built that way, and what it costs.
- Block-scaled dot performance: an investigation of the quantized kernels’ speed.
- Documentation map: every document, by folder.
Status: synapse loads and runs Llama-family models (
llama,qwen2,mistral3) and the SmolVLM image encoder from their GGUF files, with tokenizers that match llama.cpp’s exactly. Llama-3.2-1B Q4_K_M decodes at 28 to 30 tokens/s on a 20-thread x86 machine with the AVX2 build. GPU and KPU backends are planned.