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

Libraries

How to use LiquidAI/LFM2.5-350M-Base with Transformers:

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

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

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("LiquidAI/LFM2.5-350M-Base")
model = AutoModelForCausalLM.from_pretrained("LiquidAI/LFM2.5-350M-Base", 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 LiquidAI/LFM2.5-350M-Base with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/LiquidAI/LFM2.5-350M-Base

SGLang

How to use LiquidAI/LFM2.5-350M-Base 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 "LiquidAI/LFM2.5-350M-Base" \
    --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": "LiquidAI/LFM2.5-350M-Base",
		"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 "LiquidAI/LFM2.5-350M-Base" \
        --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": "LiquidAI/LFM2.5-350M-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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

LFM2.5-350M-Base degenerates on plain LM continuation (no chat template)

by yuyijiong - opened 5 days ago

Discussion

yuyijiong

5 days ago

•

edited 4 days ago

I'm testing LFM2.5-350M-Base as a standard causal LM with plain text continuation (no apply_chat_template, greedy decoding, add_special_tokens=False).

The model shows severe degeneration: repetitive tokens/phrases and failure on simple completions.

After other tests, I find LFM2.5-2.6B-Base doesn't show such degeneration, while 1.2B, 350M, 230M all show the same degeneration.

Minimal repro:

prefix = "The capital of France is"
inputs = tokenizer(prefix, return_tensors="pt", add_special_tokens=False).to(model.device)
out = model.generate(**inputs, max_new_tokens=16, do_sample=False)

Example: next-token top-1 is ' the'; generation becomes unrelated or repetitive (e.g. by by by..., = = =..., the the the...). On a small 17-case LM continuation suite: 0/17 matches.

Questions:

Is plain LM continuation expected to work for this -Base checkpoint?
Is this repetitive behavior known/expected?
Should users always use chat templating, even for -Base?

mlabonne

Liquid AI org 3 days ago

I think you might just be missing the BOS token, can you try the following?

prompt = "The capital of France is"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=16, do_sample=False)
print(tokenizer.decode(output[0], skip_special_tokens=True))

yuyijiong

2 days ago

I think you might just be missing the BOS token, can you try the following?

prompt = "The capital of France is"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=16, do_sample=False)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Thank you. Adding BOS token solved this problem. It seems small size models are quite sensitive to BOS.

yuyijiong changed discussion status to closed 2 days ago

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment