LFM2.5-1.2B-Instruct-smoltalk-LoRA

LFM2.5-1.2B-Instruct LoRA Adapter

LoRA adapter for LiquidAI/LFM2.5-1.2B-Instruct fine-tuned on SmolTalk via the official SFT tutorial.

Training Details

Target Modules

• Attention: [q_proj, k_proj, v_proj, out_proj]
• MLP (GLU): [w1, w2, w3]
• Conv: [in_proj, out_proj]

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

MODEL_ID = "LiquidAI/LFM2.5-1.2B-Instruct"
LORA_ID = "LiquidAI/LFM2.5-1.2B-Instruct-smoltalk-LoRA"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
base_model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

prompt = "What are some ideas for a good short story about a city not on a planet, but rather a generation ship, or on the moon of a gas giant, or somewhere else unusual?"
messages = [{"role": "user", "content": prompt}]
input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(input_text, return_tensors="pt").to(base_model.device)

# Generate with base model (no LoRA)
with torch.no_grad():
    outputs = base_model.generate(**inputs, max_new_tokens=100, do_sample=False)
print("Base:", tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

# Load LoRA and generate
lora_model = PeftModel.from_pretrained(base_model, LORA_ID)
with torch.no_grad():
    outputs = lora_model.generate(**inputs, max_new_tokens=100, do_sample=False)
print("LoRA:", tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Verification: LoRA Produces Different Outputs

With greedy decoding (do_sample=False), base and LoRA models produce different outputs:

vLLM Usage

Launch Server with LoRA

vllm serve LiquidAI/LFM2.5-1.2B-Instruct \
  --host 0.0.0.0 \
  --port 30000 \
  --dtype float16 \
  --enable-lora \
  --max-lora-rank 8 \
  --lora-modules "smoltalk=LiquidAI/LFM2.5-1.2B-Instruct-smoltalk-LoRA"

Call with LoRA

To use the LoRA adapter, set the model field to the LoRA adapter name (smoltalk):

curl -s http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "smoltalk",
    "messages": [{"role": "user", "content": "What are some ideas for a good short story about a city not on a planet, but rather a generation ship, or on the moon of a gas giant, or somewhere else unusual?"}],
    "max_tokens": 100,
    "temperature": 0.0
  }'

Call without LoRA (base model)

curl -s http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "LiquidAI/LFM2.5-1.2B-Instruct",
    "messages": [{"role": "user", "content": "What are some ideas for a good short story about a city not on a planet, but rather a generation ship, or on the moon of a gas giant, or somewhere else unusual?"}],
    "max_tokens": 100,
    "temperature": 0.0
  }'

SGLang Usage

Note: Requires the internal SGLang implementation with LFM2 LoRA support.

Launch Server with LoRA

python -m sglang.launch_server \
    --model-path LiquidAI/LFM2.5-1.2B-Instruct \
    --port 30000 \
    --enable-lora \
    --max-lora-rank 8 \
    --lora-paths "smoltalk=LiquidAI/LFM2.5-1.2B-Instruct-smoltalk-LoRA" \
    --lora-target-modules q_proj k_proj v_proj out_proj w1 w2 w3 in_proj

Call with LoRA

Option 1: Using lora_path parameter

curl -s http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "LiquidAI/LFM2.5-1.2B-Instruct",
    "lora_path": "smoltalk",
    "messages": [{"role": "user", "content": "What are some ideas for a good short story about a city not on a planet, but rather a generation ship, or on the moon of a gas giant, or somewhere else unusual?"}],
    "max_tokens": 100,
    "temperature": 0.0
  }'

Option 2: Using colon syntax in model name

curl -s http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "LiquidAI/LFM2.5-1.2B-Instruct:smoltalk",
    "messages": [{"role": "user", "content": "What are some ideas for a good short story about a city not on a planet, but rather a generation ship, or on the moon of a gas giant, or somewhere else unusual?"}],
    "max_tokens": 100,
    "temperature": 0.0
  }'

Call without LoRA (base model)

curl -s http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "LiquidAI/LFM2.5-1.2B-Instruct",
    "messages": [{"role": "user", "content": "What are some ideas for a good short story about a city not on a planet, but rather a generation ship, or on the moon of a gas giant, or somewhere else unusual?"}],
    "max_tokens": 100,
    "temperature": 0.0
  }'