| # 🛠️ ZetaGrid-Ollama Patch: Setup Guide | |
| To run **RTH-LM** (Fractal TCN) natively in your local environment via Ollama, you need to add support for the TCN operators to the underlying `llama.cpp` engine. | |
| ### 📦 Prerequisites | |
| 1. **GGUF Model:** Download [rth_lm_25b_v1.gguf](https://huggingface.co/RthItalia/Rth-lm-25b) | |
| 2. **Ollama Source:** Clone the official repository or use our fork. | |
| ### 🛠️ Step 1: Add Custom Kernels | |
| Copy the provided C++ files into the `llama.cpp` source tree: | |
| - Move `rth_tcn_ops.cpp` and `rth_tcn_ops.h` to `llama.cpp/src/` | |
| - Register `GGML_OP_CAUSAL_CONV1D` and `GGML_OP_FRACTAL_GATE` in `ggml.c` | |
| ### 🏗️ Step 2: Compile | |
| Rebuild `llama.cpp` or your Ollama binary: | |
| ```bash | |
| make -j | |
| # or for Ollama | |
| go generate ./... | |
| go build . | |
| ``` | |
| ### 🚀 Step 3: Create & Run | |
| Use the provided `Modelfile_RTH-LM` to register the model: | |
| ```bash | |
| ollama create rth-lm -f Modelfile_RTH-LM | |
| ollama run rth-lm | |
| ``` | |
| --- | |
| **Why this matters:** RTH-LM is a pioneer in non-Transformer architectures for local inference. By applying this patch, you are at the forefront of the TCN revolution. 🫡🌌🦙 | |