Instructions to use RegalInvader/Marco-Nano-Instruct-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use RegalInvader/Marco-Nano-Instruct-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0
# Run inference directly in the terminal:
llama cli -hf RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0
# Run inference directly in the terminal:
llama cli -hf RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0

Use Docker

docker model run hf.co/RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0

Llama new
LM Studio
Jan

vLLM

How to use RegalInvader/Marco-Nano-Instruct-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "RegalInvader/Marco-Nano-Instruct-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "RegalInvader/Marco-Nano-Instruct-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0

Ollama
How to use RegalInvader/Marco-Nano-Instruct-GGUF with Ollama:
```
ollama run hf.co/RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0
```
Unsloth Desktop

How to use RegalInvader/Marco-Nano-Instruct-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use RegalInvader/Marco-Nano-Instruct-GGUF with Docker Model Runner:
```
docker model run hf.co/RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0
```

Lemonade

How to use RegalInvader/Marco-Nano-Instruct-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0

Run and chat with the model

lemonade run user.Marco-Nano-Instruct-GGUF-TQ1_0

List all available models

lemonade list

Hermes Agent

How to use RegalInvader/Marco-Nano-Instruct-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use RegalInvader/Marco-Nano-Instruct-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "RegalInvader/Marco-Nano-Instruct-GGUF:TQ1_0" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Marco-Nano-Instruct-GGUF / code /src /lowbit /quant /importance.py

RegalInvader

Release: Marco-Nano-Instruct 1-, 2- and 3-bit GGUF for phones

e5235fb verified 4 days ago

Raw History Blame Contribute Delete

1.69 kB

"""D-57 importance weights for the N-bit scale search: llama.cpp's imatrix (mean squared input activation per input
channel, per expert) for each layer's experts, as the trainer and exporter read them.
`scripts/imatrix_to_importance.py` turns llama.cpp's imatrix.gguf into a safetensors file with, per layer i,
`layers.{i}.gate_up` [E, hidden] (gate and up read the same input) and `layers.{i}.down` [E, moe_intermediate] in fp32.
An expert the calibration text never routed to (count 0) gets its layer's mean over the routed experts.
"""
from __future__ import annotations

from pathlib import Path

import numpy as np
import torch


def from_imatrix(gguf_path: str | Path, gguf) -> dict[str, torch.Tensor]:
    """llama.cpp imatrix.gguf → {"layers.{i}.gate_up": [E, in], "layers.{i}.down": [E, in]} (fp32)."""
    r = gguf.GGUFReader(str(gguf_path))
    by = {t.name: np.asarray(t.data, dtype=np.float32) for t in r.tensors}
    out = {}
    i = 0
    while f"blk.{i}.ffn_gate_exps.weight.in_sum2" in by:
        for ours, theirs in (("gate_up", "ffn_gate_exps"), ("down", "ffn_down_exps")):
            s2 = by[f"blk.{i}.{theirs}.weight.in_sum2"]            # [E, in]
            n = by[f"blk.{i}.{theirs}.weight.counts"].reshape(-1, 1)  # [E, 1]
            seen = n[:, 0] > 0
            imp = np.where(seen[:, None], s2 / np.maximum(n, 1), 0.0)
            imp[~seen] = imp[seen].mean(0) if seen.any() else 1.0
            out[f"layers.{i}.{ours}"] = torch.from_numpy(np.ascontiguousarray(imp, dtype=np.float32))
        i += 1
    return out


def load(path: str | Path) -> dict[str, torch.Tensor]:
    from safetensors.torch import load_file
    return load_file(str(path))

	"""D-57 importance weights for the N-bit scale search: llama.cpp's imatrix (mean squared input activation per input
	channel, per expert) for each layer's experts, as the trainer and exporter read them.

	`scripts/imatrix_to_importance.py` turns llama.cpp's imatrix.gguf into a safetensors file with, per layer i,
	`layers.{i}.gate_up` [E, hidden] (gate and up read the same input) and `layers.{i}.down` [E, moe_intermediate] in fp32.
	An expert the calibration text never routed to (count 0) gets its layer's mean over the routed experts.
	"""
	from __future__ import annotations

	from pathlib import Path

	import numpy as np
	import torch


	def from_imatrix(gguf_path: str \| Path, gguf) -> dict[str, torch.Tensor]:
	"""llama.cpp imatrix.gguf → {"layers.{i}.gate_up": [E, in], "layers.{i}.down": [E, in]} (fp32)."""
	r = gguf.GGUFReader(str(gguf_path))
	by = {t.name: np.asarray(t.data, dtype=np.float32) for t in r.tensors}
	out = {}
	i = 0
	while f"blk.{i}.ffn_gate_exps.weight.in_sum2" in by:
	for ours, theirs in (("gate_up", "ffn_gate_exps"), ("down", "ffn_down_exps")):
	s2 = by[f"blk.{i}.{theirs}.weight.in_sum2"] # [E, in]
	n = by[f"blk.{i}.{theirs}.weight.counts"].reshape(-1, 1) # [E, 1]
	seen = n[:, 0] > 0
	imp = np.where(seen[:, None], s2 / np.maximum(n, 1), 0.0)
	imp[~seen] = imp[seen].mean(0) if seen.any() else 1.0
	out[f"layers.{i}.{ours}"] = torch.from_numpy(np.ascontiguousarray(imp, dtype=np.float32))
	i += 1
	return out


	def load(path: str \| Path) -> dict[str, torch.Tensor]:
	from safetensors.torch import load_file
	return load_file(str(path))