Inferra-Qwen-LoRA

Lightweight conversational AI model built using QLoRA fine-tuning on top of Qwen2.5-3B-Instruct.

---

Features

• QLoRA Fine-Tuning
• 4-bit Quantization
• Fast Inference
• Low VRAM Usage
• Hugging Face Compatible
• LoRA Adapter Based

---

Model Details

Property	Value
Base Model	Qwen2.5-3B-Instruct
Fine-Tuning	QLoRA
Precision	4-bit
Trainable Params	~15M
Total Params	~3.1B
GPU Used	NVIDIA T4
Platform	Kaggle

---

Installation

pip install -U transformers accelerate peft bitsandbytes unsloth

---

Hugging Face Model

mohdmusheer/inferra

---

Load Model

from unsloth import FastLanguageModel
from peft import PeftModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Qwen/Qwen2.5-3B-Instruct",
    max_seq_length=1024,
    load_in_4bit=True,
)

model = PeftModel.from_pretrained(
    model,
    "mohdmusheer/inferra",
)

FastLanguageModel.for_inference(model)

---

Inference Example

messages = [
    {
        "role": "user",
        "content": "Explain machine learning in simple words."
    }
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(
    text,
    return_tensors="pt",
).to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=200,
)

response = tokenizer.decode(
    outputs[0],
    skip_special_tokens=True,
)

print(response)

---

Training Config

max_seq_length = 1024
batch_size = 1
gradient_accumulation_steps = 8
max_steps = 500
learning_rate = 2e-4

---

Docker Build

docker build -t inferra-qwen .

---

Docker Run

docker run --gpus all -p 8000:8000 inferra-qwen

---

Dockerfile

FROM pytorch/pytorch:2.4.0-cuda12.1-cudnn9-runtime

WORKDIR /app

COPY . .

RUN pip install -U \
    transformers \
    accelerate \
    peft \
    bitsandbytes \
    unsloth

CMD ["python", "app.py"]

---

Project Structure

├── app.py
├── Dockerfile
├── inference.py
├── training.ipynb
└── README.md

---

Limitations

• Not full fine-tuning
• Not a frontier reasoning model
• Adapter-based conversational tuning

---

Future Improvements

• Reasoning datasets
• Coding specialization
• RAG integration
• GGUF export
• Ollama support

---

License

Please follow the license terms of:

• Base model
• Datasets
• Hugging Face ecosystem