Devy1
/

CodeLlama-7b-hf-AQLM-2bit-code-1x15

@@ -1,199 +1,229 @@
 ---
 library_name: transformers
-tags: []
 ---
-# Model Card for Model ID
-<!-- Provide a quick summary of what the model is/does. -->
-## Model Details
-### Model Description
-<!-- Provide a longer summary of what this model is. -->
-This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
-- **Developed by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
-- **Model type:** [More Information Needed]
-- **Language(s) (NLP):** [More Information Needed]
-- **License:** [More Information Needed]
-- **Finetuned from model [optional]:** [More Information Needed]
-### Model Sources [optional]
-<!-- Provide the basic links for the model. -->
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
-## Uses
-<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
-### Direct Use
-<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
-[More Information Needed]
-### Downstream Use [optional]
-<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
-[More Information Needed]
-### Out-of-Scope Use
-<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
-[More Information Needed]
-## Bias, Risks, and Limitations
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
-[More Information Needed]
-### Recommendations
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
-Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
-## How to Get Started with the Model
-Use the code below to get started with the model.
-[More Information Needed]
-## Training Details
-### Training Data
-<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
-[More Information Needed]
-### Training Procedure
-<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
-#### Preprocessing [optional]
-[More Information Needed]
-#### Training Hyperparameters
-- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
-#### Speeds, Sizes, Times [optional]
-<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
-[More Information Needed]
-## Evaluation
-<!-- This section describes the evaluation protocols and provides the results. -->
-### Testing Data, Factors & Metrics
-#### Testing Data
-<!-- This should link to a Dataset Card if possible. -->
-[More Information Needed]
-#### Factors
-<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
-[More Information Needed]
-#### Metrics
-<!-- These are the evaluation metrics being used, ideally with a description of why. -->
-[More Information Needed]
-### Results
-[More Information Needed]
-#### Summary
-## Model Examination [optional]
-<!-- Relevant interpretability work for the model goes here -->
-[More Information Needed]
-## Environmental Impact
-<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
-Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
-- **Hardware Type:** [More Information Needed]
-- **Hours used:** [More Information Needed]
-- **Cloud Provider:** [More Information Needed]
-- **Compute Region:** [More Information Needed]
-- **Carbon Emitted:** [More Information Needed]
-## Technical Specifications [optional]
-### Model Architecture and Objective
-[More Information Needed]
-### Compute Infrastructure
-[More Information Needed]
-#### Hardware
-[More Information Needed]
-#### Software
-[More Information Needed]
-## Citation [optional]
-<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
-**BibTeX:**
-[More Information Needed]
-**APA:**
-[More Information Needed]
-## Glossary [optional]
-<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
-[More Information Needed]
-## More Information [optional]
-[More Information Needed]
-## Model Card Authors [optional]
-[More Information Needed]
-## Model Card Contact
-[More Information Needed]

 ---
 library_name: transformers
+tags:
+- aqlm
+base_model:
+- codellama/CodeLlama-7b-hf
+base_model_relation: quantized
 ---
+# Quantizing Large Language Models for Code Generation: A Differentiated Replication
+## Table of Contents
+1. [Introduction](#1-introduction)
+2. [Model details](#2-model-details)
+3. [Experiments](#3-experiments)
+4. [Replication](#4-replication)
+## 1. Introduction
+HuggingFace repository containing the quantized models from the paper _"Quantizing Large Language Models for Code Generation: A Differentiated Replication."_.
+In this study, we evaluate the performance of compressed Deep Learning models on the code generation task. Specifically, we quantize code models such as CodeLlama and DeepSeek Coder at different levels of precision, namely 8, 4, 3, and 2 bits per model parameter, using a SOTA quantization technique for extreme model compression, that is [AQLM](https://github.com/Vahe1994/AQLM) (Additive Quantization of Language Models).
+## 2. Model details
+The complete list of models used in this study is available in our [model collection](https://huggingface.co/collections/Devy1/quantization-for-code-generation-67c9b83b34ed9a5a84fb714d), which is organized by order of appearance in the paper discussion.
+More specifically, we named the models as follows:
+**\<base-model\>**-AQLM-**\<precision\>**-**\<calibration\>**-**\<finetuned?\>**-**\<hyperparameters\>**
+- **\<base-model\>**: define the starting model that was used for quantization.
+- **\<precision\>**: the average number of bits per model weight.
+- **\<calibration\>**: the type of calibration performed. It can be 'rnd' (random), 'code' (code-specific), and 'mixed' (using both code and technical language).
+- **\<finetuned?\>**: if the model was fine-tuned after quantization, this tag will appear as "-finetuned". Otherwise it will not be present.
+- **\<hyperparameters\>**: number of codebooks and codebook size used for quantization. Expressed in the format **\<codebooks\>**x**\<bits\>**.
+For example, the model **Devy1/CodeLlama-7b-hf-AQLM-2bit-rnd-1x15** has the following features:
+1. This model is a compressed version of [CodeLlama-7b-hf](https://huggingface.co/codellama/CodeLlama-7b-hf).
+2. On average, each parameter is represented by **2 bits**.
+3. We used a (**random**) sample of the RedPajama dataset for the calibration process.
+4. The model was **not fine-tuned** after quantization (because the -finetuned tag does not appear after the calibration dataset type).
+5. We used **1** codebook of **15** bits to quantize the model. The default group size used for each model is 8.
+More information about the quantization process and hyperparameters can be found in our paper and in the config.json file from this repository.
+## 3. Experiments
+Below, we present the code generation performance of each quantized model across different experiments. Performance is computed on Python and Java languages using [MultiPL-E](https://github.com/nuprl/MultiPL-E) and [McEval](https://mceval.github.io/) benchmarks. More details on the research approach can be found in our paper.
+Results are listed by research question and benchmark. By clicking on the "precision" value, you will be redirected to the corresponding model.
+### RQ1. How does low-bit quantization affect the model’s code generation ability?
+#### MultiPL-E benchmark
+| Model                | Params | Precision | Size      | Python pass@1 | Java pass@1 |
+|----------------------|--------:|-------------------------|-----------:|---------------:|-------------:|
+| CodeLlama - Base     | 7B     | [Float16](https://huggingface.co/codellama/CodeLlama-7b-hf)   | 13.48 GB  | 29.8          | 32.2        |
+|                      |        | [8-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-8bit-rnd-4x15)     | 7.47 GB   | 29.7          | 31.6        |
+|                      |        | [4-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-4bit-rnd-2x15)     | 4.00 GB   | 29.1          | 30.7        |
+|                      |        | [3-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-rnd-2x12)     | 3.80 GB   | 24.3          | 26.5        |
+|                      |        | [2-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-rnd-1x15)     | 2.26 GB   | 16.4          | 14.1        |
+| DeepSeek-Coder - Base| 7B     | [Float16](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base)   | 13.48 GB  | 45.8          | 41.4        |
+|                      |        | [8-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-8bit-rnd-4x15)     | 7.48 GB   | 46.2          | 41.9        |
+|                      |        | [4-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-4bit-rnd-2x15)     | 4.00 GB   | 45.2          | 41.4        |
+|                      |        | [3-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-rnd-2x12)     | 3.80 GB   | 41.1          | 37.7        |
+|                      |        | [2-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-rnd-1x15)     | 2.27 GB   | 27.6          | 23.2        |
+#### McEval benchmark
+| Model                | Params | Precision | Size      | Python pass@1 | Java pass@1 |
+|----------------------|--------:|-------------------------|-----------:|---------------:|-------------:|
+| CodeLlama - Base     | 7B     | [Float16](https://huggingface.co/codellama/CodeLlama-7b-hf)   | 13.48 GB  | 12.9          | 29.3        |
+|                      |        | [8-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-8bit-rnd-4x15)     | 7.47 GB   | 12.9          | 29.2        |
+|                      |        | [4-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-4bit-rnd-2x15)     | 4.00 GB   | 15.2          | 25.3        |
+|                      |        | [3-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-rnd-2x12)     | 3.80 GB   | 10.0          | 21.3        |
+|                      |        | [2-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-rnd-1x15)     | 2.26 GB   | 5.6           | 11.4        |
+| DeepSeek-Coder - Base| 7B     | [Float16](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base)   | 13.48 GB  | 41.8          | 42.6        |
+|                      |        | [8-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-8bit-rnd-4x15)     | 7.48 GB   | 42.5          | 42.8        |
+|                      |        | [4-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-4bit-rnd-2x15)     | 4.00 GB   | 40.7          | 45.9        |
+|                      |        | [3-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-rnd-2x12)     | 3.80 GB   | 36.2          | 34.5        |
+|                      |        | [2-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-rnd-1x15)     | 2.27 GB   | 13.7          | 23.6        |
+### RQ1. Impact of end-to-end fine-tuning after quantization
+#### MultiPL-E benchmark
+| Model                | Params | Precision               | Size      | Python pass@1 | Java pass@1 |
+|----------------------|--------:|-------------------------|-----------:|---------------:|-------------:|
+| CodeLlama - Base     | 7B     | [Float16](https://huggingface.co/codellama/CodeLlama-7b-hf)                 | 13.48 GB  | 29.8          | 32.2        |
+|                      |        | [3-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-rnd-2x12)                   | 3.80 GB   | 24.3          | 26.5        |
+|                      |        | [2-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-rnd-1x15)                   | 2.26 GB   | 16.4          | 14.1        |
+|                      |        | [3-bit + Fine-tuning](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-rnd-finetuned-2x12)     | 3.80 GB   | <span style="color:red;">&#9660;</span> **24.0** | <span style="color:green;">&#9650;</span> **27.8** |
+|                      |        | [2-bit + Fine-tuning](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-rnd-finetuned-1x15)     | 2.26 GB   | <span style="color:green;">&#9650;</span> **19.9** | <span style="color:green;">&#9650;</span> **19.0** |
+| DeepSeek-Coder - Base| 7B     | [Float16](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base)                 | 13.48 GB  | 45.8          | 41.4        |
+|                      |        | [3-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-rnd-2x12)                   | 3.80 GB   | 41.1          | 37.7        |
+|                      |        | [2-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-rnd-1x15)                   | 2.27 GB   | 27.6          | 23.2        |
+|                      |        | [3-bit + Fine-tuning](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-rnd-finetuned-2x12)     | 3.80 GB   | <span style="color:green;">&#9650;</span> **41.8** | <span style="color:red;">&#9660;</span> **37.7** |
+|                      |        | [2-bit + Fine-tuning](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-rnd-finetuned-1x15)     | 2.27 GB   | <span style="color:green;">&#9650;</span> **33.0** | <span style="color:green;">&#9650;</span> **26.8** |
+#### McEval benchmark
+| Model                | Params | Precision               | Size      | Python pass@1 | Java pass@1 |
+|----------------------|--------:|-------------------------|-----------:|---------------:|-------------:|
+| CodeLlama - Base     | 7B     | [Float16](https://huggingface.co/codellama/CodeLlama-7b-hf)                 | 13.48 GB  | 12.9          | 29.3        |
+|                      |        | [3-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-rnd-2x12)                   | 3.80 GB   | 10.0          | 21.3        |
+|                      |        | [2-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-rnd-1x15)                   | 2.26 GB   | 5.6           | 11.4        |
+|                      |        | [3-bit + Fine-tuning](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-rnd-finetuned-2x12)     | 3.80 GB   | <span style="color:green;">&#9650;</span> **10.8** | <span style="color:green;">&#9650;</span> **22.0** |
+|                      |        | [2-bit + Fine-tuning](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-rnd-finetuned-1x15)     | 2.26 GB   | <span style="color:green;">&#9650;</span> **7.6** | <span style="color:green;">&#9650;</span> **14.3** |
+| DeepSeek-Coder - Base| 7B     | [Float16](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base)                 | 13.48 GB  | 41.8          | 42.6        |
+|                      |        | [3-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-rnd-2x12)                   | 3.80 GB   | 36.2          | 34.5        |
+|                      |        | [2-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-rnd-1x15)                   | 2.27 GB   | 13.7          | 23.6        |
+|                      |        | [3-bit + Fine-tuning](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-rnd-finetuned-2x12)     | 3.80 GB   | <span style="color:red;">&#9660;</span> **35.6** | <span style="color:red;">&#9660;</span> **32.4** |
+|                      |        | [2-bit + Fine-tuning](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-rnd-finetuned-1x15)     | 2.27 GB   | <span style="color:green;">&#9650;</span> **20.2** | <span style="color:green;">&#9650;</span> **27.0** |
+### RQ2. Which impact does the calibration dataset have on model performance?
+#### MultiPL-E benchmark
+| Model                | Params | Precision               | Size      | Python pass@1 | Java pass@1 |
+|----------------------|--------|-------------------------|----------:|--------------:|------------:|
+| CodeLlama - Base     | 7B     | [Float16 - Baseline](https://huggingface.co/codellama/CodeLlama-7b-hf)      | 13.48 GB  |          29.8 |        32.2 |
+|                      |        | [8-bit with Random samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-8bit-rnd-4x15) |  7.47 GB  |          29.7 |        31.6 |
+|                      |        | [8-bit with Mixed samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-8bit-mixed-4x15)  |  7.47 GB  | <span style="color:red;">&#9660;</span> 29.7 | <span style="color:green;">&#9650;</span> 32.3 |
+|                      |        | [8-bit with Code samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-8bit-code-4x15)   |  7.47 GB  | <span style="color:red;">&#9660;</span> 29.2 | <span style="color:green;">&#9650;</span> 32.0 |
+|                      |        | [4-bit with Random samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-4bit-rnd-2x15) |  4.00 GB  |          29.1 |        30.7 |
+|                      |        | [4-bit with Mixed samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-4bit-mixed-2x15)  |  4.00 GB  | <span style="color:red;">&#9660;</span> 29.0 | <span style="color:green;">&#9650;</span> 31.4 |
+|                      |        | [4-bit with Code samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-4bit-code-2x15)   |  4.00 GB  | <span style="color:green;">&#9650;</span> 30.2 | <span style="color:red;">&#9660;</span> 29.8 |
+|                      |        | [3-bit with Random samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-rnd-2x12) |  3.80 GB  |          24.3 |        26.5 |
+|                      |        | [3-bit with Mixed samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-mixed-2x12)  |  3.80 GB  | <span style="color:green;">&#9650;</span> 28.2 | <span style="color:green;">&#9650;</span> 28.4 |
+|                      |        | [3-bit with Code samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-code-2x12)   |  3.80 GB  | <span style="color:green;">&#9650;</span> 27.0 | <span style="color:green;">&#9650;</span> 28.0 |
+|                      |        | [2-bit with Random samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-rnd-1x15) |  2.26 GB  |          16.4 |        14.1 |
+|                      |        | [2-bit with Mixed samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-mixed-1x15)  |  2.26 GB  | <span style="color:green;">&#9650;</span> 23.9 | <span style="color:green;">&#9650;</span> 21.5 |
+|                      |        | [2-bit with Code samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-code-1x15)   |  2.26 GB  | <span style="color:green;">&#9650;</span> 24.1 | <span style="color:green;">&#9650;</span> 19.4 |
+| DeepSeek-Coder - Base| 7B     | [Float16 - Baseline](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base)      | 13.48 GB  |          45.8 |        41.4 |
+|                      |        | [8-bit with Random samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-8bit-rnd-4x15) |  7.48 GB  |          46.2 |        41.9 |
+|                      |        | [8-bit with Mixed samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-8bit-mixed-4x15)  |  7.48 GB  | <span style="color:red;">&#9660;</span> 45.4 | <span style="color:green;">&#9650;</span> 43.2 |
+|                      |        | [8-bit with Code samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-8bit-code-4x15)   |  7.48 GB  | <span style="color:red;">&#9660;</span> 45.9 | <span style="color:red;">&#9660;</span> 41.7 |
+|                      |        | [4-bit with Random samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-4bit-rnd-2x15) |  4.00 GB  |          45.2 |        41.4 |
+|                      |        | [4-bit with Mixed samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-4bit-mixed-2x15)  |  4.00 GB  | <span style="color:red;">&#9660;</span> 44.5 | <span style="color:green;">&#9650;</span> 41.8 |
+|                      |        | [4-bit with Code samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-4bit-code-2x15)   |  4.00 GB  | <span style="color:red;">&#9660;</span> 44.2 | <span style="color:red;">&#9660;</span> 40.6 |
+|                      |        | [3-bit with Random samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-rnd-2x12) |  3.80 GB  |          41.1 |        37.7 |
+|                      |        | [3-bit with Mixed samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-mixed-2x12)  |  3.80 GB  | <span style="color:green;">&#9650;</span> 43.7 | <span style="color:green;">&#9650;</span> 39.1 |
+|                      |        | [3-bit with Code samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-code-2x12)   |  3.80 GB  | <span style="color:green;">&#9650;</span> 42.5 | <span style="color:green;">&#9650;</span> 38.7 |
+|                      |        | [2-bit with Random samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-rnd-1x15) |  2.27 GB  |          27.6 |        23.2 |
+|                      |        | [2-bit with Mixed samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-mixed-1x15)  |  2.27 GB  | <span style="color:green;">&#9650;</span> 35.7 | <span style="color:green;">&#9650;</span> 27.4 |
+|                      |        | [2-bit with Code samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-code-1x15)   |  2.27 GB  | <span style="color:green;">&#9650;</span> 34.8 | <span style="color:green;">&#9650;</span> 27.5 |
+#### McEval benchmark
+| Model                | Params | Precision               | Size      | Python pass@1 | Java pass@1 |
+|----------------------|--------|-------------------------|----------:|--------------:|------------:|
+| CodeLlama - Base     | 7B     | [Float16 - Baseline](https://huggingface.co/codellama/CodeLlama-7b-hf)      | 13.48 GB  |          12.9 |        29.3 |
+|                      |        | [8-bit with Random samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-8bit-rnd-4x15) |  7.47 GB  |          12.9 |        29.2 |
+|                      |        | [8-bit with Mixed samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-8bit-mixed-4x15)  |  7.47 GB  | <span style="color:green;">&#9650;</span> 13.7 | <span style="color:red;">&#9660;</span> 28.6 |
+|                      |        | [8-bit with Code samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-8bit-code-4x15)   |  7.47 GB  | <span style="color:red;">&#9660;</span> 12.3 | <span style="color:green;">&#9650;</span> 29.5 |
+|                      |        | [4-bit with Random samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-4bit-rnd-2x15) |  4.00 GB  |          15.2 |        25.3 |
+|                      |        | [4-bit with Mixed samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-4bit-mixed-2x15)  |  4.00 GB  | <span style="color:red;">&#9660;</span> 13.0 | <span style="color:green;">&#9650;</span> 30.3 |
+|                      |        | [4-bit with Code samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-4bit-code-2x15)   |  4.00 GB  | <span style="color:red;">&#9660;</span> 11.1 | <span style="color:green;">&#9650;</span> 25.8 |
+|                      |        | [3-bit with Random samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-rnd-2x12) |  3.80 GB  |          10.0 |        21.3 |
+|                      |        | [3-bit with Mixed samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-mixed-2x12)  |  3.80 GB  | <span style="color:green;">&#9650;</span> 12.3 | <span style="color:green;">&#9650;</span> 25.5 |
+|                      |        | [3-bit with Code samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-3bit-code-2x12)   |  3.80 GB  | <span style="color:green;">&#9650;</span> 10.8 | <span style="color:red;">&#9660;</span> 19.9 |
+|                      |        | [2-bit with Random samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-rnd-1x15) |  2.26 GB  |           5.6 |        11.4 |
+|                      |        | [2-bit with Mixed samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-mixed-1x15)  |  2.26 GB  | <span style="color:green;">&#9650;</span> 11.1 | <span style="color:green;">&#9650;</span> 12.8 |
+|                      |        | [2-bit with Code samples](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-code-1x15)   |  2.26 GB  | <span style="color:green;">&#9650;</span>  6.1 | <span style="color:green;">&#9650;</span> 12.8 |
+| DeepSeek-Coder - Base| 7B     | [Float16 - Baseline](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base)      | 13.48 GB  |          41.8 |        42.6 |
+|                      |        | [8-bit with Random samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-8bit-rnd-4x15) |  7.48 GB  |          42.5 |        42.8 |
+|                      |        | [8-bit with Mixed samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-8bit-mixed-4x15)  |  7.48 GB  | <span style="color:green;">&#9650;</span> 42.7 | <span style="color:red;">&#9660;</span> 42.5 |
+|                      |        | [8-bit with Code samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-8bit-code-4x15)   |  7.48 GB  | <span style="color:red;">&#9660;</span> 41.3 | <span style="color:red;">&#9660;</span> 42.7 |
+|                      |        | [4-bit with Random samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-4bit-rnd-2x15) |  4.00 GB  |          40.7 |        45.9 |
+|                      |        | [4-bit with Mixed samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-4bit-mixed-2x15)  |  4.00 GB  | <span style="color:red;">&#9660;</span> 39.0 | <span style="color:red;">&#9660;</span> 42.8 |
+|                      |        | [4-bit with Code samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-4bit-code-2x15)   |  4.00 GB  | <span style="color:red;">&#9660;</span> 39.8 | <span style="color:green;">&#9650;</span> 46.3 |
+|                      |        | [3-bit with Random samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-rnd-2x12) |  3.80 GB  |          36.2 |        34.5 |
+|                      |        | [3-bit with Mixed samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-mixed-2x12)  |  3.80 GB  | <span style="color:red;">&#9660;</span> 35.5 | <span style="color:green;">&#9650;</span> 42.8 |
+|                      |        | [3-bit with Code samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-3bit-code-2x12)   |  3.80 GB  | <span style="color:green;">&#9650;</span> 36.5 | <span style="color:green;">&#9650;</span> 45.6 |
+|                      |        | [2-bit with Random samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-rnd-1x15) |  2.27 GB  |          13.7 |        23.6 |
+|                      |        | [2-bit with Mixed samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-mixed-1x15)  |  2.27 GB  | <span style="color:green;">&#9650;</span> 26.2 | <span style="color:green;">&#9650;</span> 29.1 |
+|                      |        | [2-bit with Code samples](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-code-1x15)   |  2.27 GB  | <span style="color:green;">&#9650;</span> 24.6 | <span style="color:green;">&#9650;</span> 28.0 |
+### RQ3. How does extreme quantization affect model accuracy across different model sizes?
+#### MultiPL-E benchmark
+| Model                | Params | Precision               | Size (GB) | Python pass@1 | Dec (%) | Java pass@1 | Dec (%) |
+|----------------------|--------|-------------------------|----------:|--------------:|--------:|------------:|--------:|
+| CodeLlama - Base     | 7B     | [Float16](https://huggingface.co/codellama/CodeLlama-7b-hf)                 |     13.48 |          29.8 |     --- |        32.2 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-mixed-1x15)                  |      2.26 |          23.9 |   -19.8 |        21.5 |   -33.2 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-mixed-finetuned-1x15)      |      2.26 |          25.5 |   -14.4 |        26.5 |   -17.7 |
+|                      | 13B    | [Float16](https://huggingface.co/codellama/CodeLlama-13b-hf)                 |     24.25 |          34.3 |     --- |        38.3 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/CodeLlama-13b-hf-AQLM-2bit-mixed-1x15)                   |      3.98 |          30.9 |    -9.9 |        27.7 |   -27.7 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/CodeLlama-13b-hf-AQLM-2bit-mixed-finetuned-1x15)      |      3.98 |          30.1 |   -12.2 |        32.8 |   -14.4 |
+|                      | 34B    | [Float16](https://huggingface.co/codellama/CodeLlama-34b-hf)                 |     62.74 |          41.9 |     --- |        44.1 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/CodeLlama-34b-hf-AQLM-2bit-mixed-1x15)                   |      9.54 |          37.1 |   -11.5 |        32.7 |   -25.9 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/CodeLlama-34b-hf-AQLM-2bit-mixed-finetuned-1x15)      |      9.54 |          36.0 |   -14.1 |        36.1 |   -18.1 |
+| DeepSeek-Coder - Base| 1B     | [Float16](https://huggingface.co/deepseek-ai/deepseek-coder-1.3b-base)                 |      2.57 |          28.4 |     --- |        28.8 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/DeepSeek-Coder-1.3b-base-AQLM-2bit-mixed-1x14)                   |      0.61 |          13.9 |   -51.1 |         6.6 |   -77.1 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/DeepSeek-Coder-1.3b-base-AQLM-2bit-mixed-finetuned-1x14)      |      0.61 |          21.7 |   -23.6 |        14.7 |   -49.0 |
+|                      | 7B     | [Float16](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base)                 |     13.48 |          45.8 |     --- |        41.4 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-mixed-1x15)                   |      2.27 |          35.7 |   -22.1 |        27.4 |   -33.8 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-mixed-finetuned-1x15)      |      2.27 |          36.4 |   -20.5 |        32.8 |   -20.8 |
+|                      | 33B    | [Float16](https://huggingface.co/deepseek-ai/deepseek-coder-33b-base)                 |     62.16 |          52.1 |     --- |        47.3 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/DeepSeek-Coder-33b-base-AQLM-2bit-mixed-1x15)                   |      9.38 |          43.4 |   -16.7 |        34.5 |   -27.1 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/DeepSeek-Coder-33b-base-AQLM-2bit-mixed-finetuned-1x15)      |      9.38 |          43.0 |   -17.5 |        38.7 |   -18.2 |
+#### McEval benchmark
+| Model                | Params | Precision               | Size (GB) | Python pass@1 | Dec (%) | Java pass@1 | Dec (%) |
+|----------------------|--------|-------------------------|----------:|--------------:|--------:|------------:|--------:|
+| CodeLlama - Base     | 7B     | [Float16](https://huggingface.co/codellama/CodeLlama-7b-hf)                 |     13.48 |          12.9 |     --- |        29.3 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-mixed-1x15)                   |      2.26 |          11.1 |   -14.0 |        12.8 |   -56.3 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/CodeLlama-7b-hf-AQLM-2bit-mixed-finetuned-1x15)      |      2.26 |          13.0 |    -0.8 |        18.3 |   -37.5 |
+|                      | 13B    | [Float16](https://huggingface.co/codellama/CodeLlama-13b-hf)                 |     24.25 |          18.9 |     --- |        40.9 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/CodeLlama-13b-hf-AQLM-2bit-mixed-1x15)                   |      3.98 |           9.4 |   -50.3 |        22.3 |   -45.5 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/CodeLlama-13b-hf-AQLM-2bit-mixed-finetuned-1x15)      |      3.98 |          10.4 |   -45.0 |        27.8 |   -32.0 |
+|                      | 34B    | [Float16](https://huggingface.co/codellama/CodeLlama-34b-hf)                 |     62.74 |          29.0 |     --- |        39.2 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/CodeLlama-34b-hf-AQLM-2bit-mixed-1x15)                   |      9.54 |          17.6 |   -39.3 |        25.2 |   -35.7 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/CodeLlama-34b-hf-AQLM-2bit-mixed-finetuned-1x15)      |      9.54 |          19.0 |   -34.5 |        31.6 |   -19.4 |
+| DeepSeek-Coder - Base| 1B     | [Float16](https://huggingface.co/deepseek-ai/deepseek-coder-1.3b-base)                 |      2.57 |          23.8 |     --- |        42.0 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/DeepSeek-Coder-1.3b-base-AQLM-2bit-mixed-1x14)                   |      0.61 |           4.4 |   -81.5 |         8.5 |   -79.8 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/DeepSeek-Coder-1.3b-base-AQLM-2bit-mixed-finetuned-1x14)      |      0.61 |           6.9 |   -71.0 |        15.5 |   -63.1 |
+|                      | 7B     | [Float16](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base)                 |     13.48 |          41.8 |     --- |        42.6 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-mixed-1x15)                   |      2.27 |          26.2 |   -37.3 |        29.1 |   -31.7 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/DeepSeek-Coder-6.7b-base-AQLM-2bit-mixed-finetuned-1x15)      |      2.27 |          30.1 |   -28.0 |        31.0 |   -27.2 |
+|                      | 33B    | [Float16](https://huggingface.co/deepseek-ai/deepseek-coder-33b-base)                 |     62.16 |          55.5 |     --- |        57.0 |     --- |
+|                      |        | [2-bit](https://huggingface.co/Devy1/DeepSeek-Coder-33b-base-AQLM-2bit-mixed-1x15)                   |      9.38 |          36.9 |   -33.5 |        39.2 |   -31.2 |
+|                      |        | [2-bit + Finetuning](https://huggingface.co/Devy1/DeepSeek-Coder-33b-base-AQLM-2bit-mixed-finetuned-1x15)      |      9.38 |          39.8 |   -28.3 |        44.0 |   -22.8 |
+## 4. Replication
+The scripts used to quantize and evaluate the models are available in our GitHub repository ([link](https://github.com/Devy99/lowbit-quantization)).
+Model predictions, statistical results, and datasets are instead available in our Zenodo repository ([link](https://doi.org/10.5281/zenodo.13752774)).