This repository contains the 4-bit AWQ (Activation-aware Weight Quantization) version of the Floppanacci/DeepSeek-R1-Distill-Qwen-7B-Floppanacci model.

Model Description

This model is optimized for faster inference and lower memory footprint compared to the original bf16/fp16 fine-tuned model. It's designed for mathematical reasoning tasks, especially Chain-of-Thought style problem-solving relevant to the AIMO competition.

The original model was fine-tuned on the Floppanacci/QWQ-LongCOT-AIMO dataset.

How to Use

With `transformers` (and `autoawq`)

You need to install the autoawq library:

pip install autoawq transformers torch

Then use the model with transformers:

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Floppanacci/DeepSeek-R1-Distill-Qwen-7B-Floppanacci-AWQ"
tokenizer = AutoTokenizer.from_pretrained(model_id)
# Load the AWQ quantized model
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto" # Automatically uses available GPU(s)
)

# Example Prompt (adjust based on how the model expects input)
prompt = "Question: Let $ABCD$ be a unit square. Let $P$ be a point inside the square such that $PA = \sqrt{5}/3$, $PB = \sqrt{2}/3$, and $PC = \sqrt{5}/3$. Find the distance $PD$. Answer:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

# Generate
outputs = model.generate(**inputs, max_new_tokens=300, temperature=0.1, do_sample=False) # Example settings
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(response)

With `vLLM` (Optimized Inference)

For higher throughput and optimized inference, you can use vLLM.

First, install vLLM:

pip install vllm

Then run the following Python code:

python

from vllm import LLM, SamplingParams

# Define prompts
prompts = [
    "Question: Let $ABCD$ be a unit square. Let $P$ be a point inside the square such that $PA = \sqrt{5}/3$, $PB = \sqrt{2}/3$, and $PC = \sqrt{5}/3$. Find the distance $PD$. Answer:",
    "Question: What is the sum of the first 100 positive integers? Answer:",
]

# Define sampling parameters
sampling_params = SamplingParams(temperature=0.1, top_p=0.95, max_tokens=300)

# Initialize the LLM engine with the AWQ model
llm = LLM(model="Floppanacci/DeepSeek-R1-Distill-Qwen-7B-Floppanacci-AWQ",
          quantization="awq",
          dtype="auto", # vLLM will typically use half-precision for activations (use bfloat16 on compatible hardware e.g. L4, A100, H100, etc.)
          trust_remote_code=True
         )

# Generate responses
outputs = llm.generate(prompts, sampling_params)

# Print the outputs
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

Floppanacci/DeepSeek-R1-Distill-Qwen-7B-Floppanacci-AWQ

DeepSeek-R1-Distill-Qwen-7B-Floppanacci (4-bit AWQ Quantized)

Model Description

How to Use

With transformers (and autoawq)

With vLLM (Optimized Inference)

Floppanacci/
DeepSeek-R1-Distill-Qwen-7B-Floppanacci-AWQ

With `transformers` (and `autoawq`)

With `vLLM` (Optimized Inference)