Instructions to use Nichonauta/LFM2.5-350M-ToMoE-INT4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Nichonauta/LFM2.5-350M-ToMoE-INT4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Nichonauta/LFM2.5-350M-ToMoE-INT4", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Nichonauta/LFM2.5-350M-ToMoE-INT4", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Nichonauta/LFM2.5-350M-ToMoE-INT4", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Nichonauta/LFM2.5-350M-ToMoE-INT4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Nichonauta/LFM2.5-350M-ToMoE-INT4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Nichonauta/LFM2.5-350M-ToMoE-INT4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Nichonauta/LFM2.5-350M-ToMoE-INT4

SGLang

How to use Nichonauta/LFM2.5-350M-ToMoE-INT4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Nichonauta/LFM2.5-350M-ToMoE-INT4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Nichonauta/LFM2.5-350M-ToMoE-INT4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Nichonauta/LFM2.5-350M-ToMoE-INT4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Nichonauta/LFM2.5-350M-ToMoE-INT4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Nichonauta/LFM2.5-350M-ToMoE-INT4 with Docker Model Runner:
```
docker model run hf.co/Nichonauta/LFM2.5-350M-ToMoE-INT4
```

Nichonauta commited on 2 days ago

Commit

d3b809f

verified ·

1 Parent(s): fd16fe9

Upload README.md with huggingface_hub

Browse files

Files changed (1) hide show

README.md +71 -0

README.md ADDED Viewed

	@@ -0,0 +1,71 @@

+---
+license: other
+license_name: lfm1.0
+license_link: https://huggingface.co/LiquidAI/LFM2.5-350M/blob/main/LICENSE
+language:
+- en
+- ar
+- zh
+- fr
+- de
+- ja
+- ko
+- es
+- pt
+base_model: Nichonauta/LFM2.5-350M-ToMoE
+tags:
+- lfm2
+- mixture-of-experts
+- tomoe
+- quantization
+pipeline_tag: text-generation
+library_name: transformers
+---
+# LFM2.5-350M-ToMoE-INT4
+Weight-only **INT4** quantization of the real ToMoE MoE — **[Nichonauta/LFM2.5-350M-ToMoE](https://huggingface.co/Nichonauta/LFM2.5-350M-ToMoE)** (derived from [LiquidAI/LFM2.5-350M](https://huggingface.co/LiquidAI/LFM2.5-350M)).
+Unlike the GGUF companions (which are **dense-equivalents** because llama.cpp cannot represent the channel-MoE), this repository stores the **actual pruned MoE weights** quantized in safetensors format:
+- `LinearINT4` replaces the MLP/attention/conv Linear layers (per-channel symmetric for INT8; per-group of 64 with zero-point for INT4), packed 2-nibbles-per-byte for INT4.
+- The **routers, embeddings and norms stay fp32** (tiny and sensitive); weights dequantize lazily in the forward pass.
+- Same custom model definition (`modeling_lfm2_moe_final.py`, `trust_remote_code`), plus the `quantization` flag in `config.json`.
+## Metrics vs the fp32 MoE
+| Metric | fp32 MoE | INT8 | INT4 |
+|---|---|---|---|
+| PPL wikitext-2 (raw, 1810 tok) | 994 | 987 | 1221 |
+| PPL chat-formatted (prompt+answer) | 314 | 785 | 941 |
+| `model.safetensors` size | 1226 MB | 400 MB | — |
+The 350M MoE sits at a fragile equilibrium (chat PPL 314 is already near the repetition-collapse regime): INT8 pushes it to ~785 and INT4 to ~941 on chat, while raw-text PPL stays roughly unchanged. This is a property of the pruned model, not of the quantizer.
+## Usage (transformers, trust_remote_code)
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+model = AutoModelForCausalLM.from_pretrained(
+    "Nichonauta/LFM2.5-350M-ToMoE-INT4",
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+    device_map="cuda:0",
+    attn_implementation="eager",
+)
+tokenizer = AutoTokenizer.from_pretrained("Nichonauta/LFM2.5-350M-ToMoE-INT4")
+```
+The `quantization` field in `config.json` makes the custom modeling file build the quantized Linear layers automatically.
+## Files
+- `model.safetensors` — INT4 weights + fp32 scales/zero-points + fp32 routers/embeddings/buffers
+- `modeling_lfm2_moe_final.py` — custom model definition with `LinearINT4` support (trust_remote_code)
+- `config.json` — model configuration (`quantization: "int4"`, `auto_map`)
+- `tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`, `generation_config.json`, `LICENSE`
+## License
+Derivative of LiquidAI/LFM2.5-350M — released under the **LFM Open License v1.0** (see `LICENSE`).