CISCai/Qwen2.5-Coder-32B-Instruct-SOTA-GGUF

Qwen2.5-Coder-32B-Instruct - SOTA GGUF

• Model creator: Qwen
• Original model: Qwen2.5-Coder-32B-Instruct

Description

This repo contains State Of The Art quantized GGUF format model files for Qwen2.5-Coder-32B-Instruct.

Quantization was done with an importance matrix that was trained for ~1M tokens (256 batches of 4096 tokens) of answers from the CodeFeedback-Filtered-Instruction dataset.

Fill-in-Middle tokens are automatically detected and supported as of commit 11ac980, see example.

Prompt template: ChatML

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Compatibility

These quantised GGUFv3 files are compatible with llama.cpp from February 27th 2024 onwards, as of commit 0becb22

They are also compatible with many third party UIs and libraries provided they are built using a recent llama.cpp.

Explanation of quantisation methods

Click to see details

The new methods available are:

• GGML_TYPE_IQ1_S - 1-bit quantization in super-blocks with an importance matrix applied, effectively using 1.56 bits per weight (bpw)
• GGML_TYPE_IQ1_M - 1-bit quantization in super-blocks with an importance matrix applied, effectively using 1.75 bpw
• GGML_TYPE_IQ2_XXS - 2-bit quantization in super-blocks with an importance matrix applied, effectively using 2.06 bpw
• GGML_TYPE_IQ2_XS - 2-bit quantization in super-blocks with an importance matrix applied, effectively using 2.31 bpw
• GGML_TYPE_IQ2_S - 2-bit quantization in super-blocks with an importance matrix applied, effectively using 2.5 bpw
• GGML_TYPE_IQ2_M - 2-bit quantization in super-blocks with an importance matrix applied, effectively using 2.7 bpw
• GGML_TYPE_IQ3_XXS - 3-bit quantization in super-blocks with an importance matrix applied, effectively using 3.06 bpw
• GGML_TYPE_IQ3_XS - 3-bit quantization in super-blocks with an importance matrix applied, effectively using 3.3 bpw
• GGML_TYPE_IQ3_S - 3-bit quantization in super-blocks with an importance matrix applied, effectively using 3.44 bpw
• GGML_TYPE_IQ3_M - 3-bit quantization in super-blocks with an importance matrix applied, effectively using 3.66 bpw
• GGML_TYPE_IQ4_XS - 4-bit quantization in super-blocks with an importance matrix applied, effectively using 4.25 bpw
• GGML_TYPE_IQ4_NL - 4-bit non-linearly mapped quantization with an importance matrix applied, effectively using 4.5 bpw

Refer to the Provided Files table below to see what files use which methods, and how.

Provided files

Name	Quant method	Bits	Size	Max RAM required	Use case
Qwen2.5-Coder-32B-Instruct.IQ1_S.gguf	IQ1_S	1	6.8 GB	7.8 GB	smallest, significant quality loss
Qwen2.5-Coder-32B-Instruct.IQ1_M.gguf	IQ1_M	1	7.4 GB	8.4 GB	very small, significant quality loss
Qwen2.5-Coder-32B-Instruct.IQ2_XXS.gguf	IQ2_XXS	2	8.4 GB	9.4 GB	very small, high quality loss
Qwen2.5-Coder-32B-Instruct.IQ2_XS.gguf	IQ2_XS	2	9.3 GB	10.3 GB	very small, high quality loss
Qwen2.5-Coder-32B-Instruct.IQ2_S.gguf	IQ2_S	2	9.7 GB	10.7 GB	small, substantial quality loss
Qwen2.5-Coder-32B-Instruct.IQ2_M.gguf	IQ2_M	2	10.5 GB	11.5 GB	small, greater quality loss
Qwen2.5-Coder-32B-Instruct.IQ3_XXS.gguf	IQ3_XXS	3	11.9 GB	12.9 GB	very small, high quality loss
Qwen2.5-Coder-32B-Instruct.IQ3_XS.gguf	IQ3_XS	3	12.8 GB	13.8 GB	small, substantial quality loss
Qwen2.5-Coder-32B-Instruct.IQ3_S.gguf	IQ3_S	3	13.4 GB	14.4 GB	small, greater quality loss
Qwen2.5-Coder-32B-Instruct.IQ3_M.gguf	IQ3_M	3	13.8 GB	14.8 GB	medium, balanced quality - recommended
Qwen2.5-Coder-32B-Instruct.IQ4_XS.gguf	IQ4_XS	4	16.5 GB	17.5 GB	small, substantial quality loss

Generated importance matrix file: Qwen2.5-Coder-32B-Instruct.imatrix.dat

Note: the above RAM figures assume no GPU offloading with 4K context. If layers are offloaded to the GPU, this will reduce RAM usage and use VRAM instead.

Example llama.cpp command

Make sure you are using llama.cpp from commit 0becb22 or later.

./llama-cli -ngl 65 -m Qwen2.5-Coder-32B-Instruct.IQ4_XS.gguf --color -c 131072 --temp 0.7 --top-p 0.8 --top-k 20 --repeat-penalty 1.05 -p "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>\n{prompt}<|im_end|>\n<|im_start|>assistant\n"

Change -ngl 65 to the number of layers to offload to GPU. Remove it if you don't have GPU acceleration.

Change -c 131072 to the desired sequence length.

If you are low on V/RAM try quantizing the K-cache with -ctk q8_0 or even -ctk q4_0 for big memory savings (depending on context size). There is a similar option for V-cache (-ctv), only available if you enable Flash Attention (-fa) as well.

For other parameters and how to use them, please refer to the llama.cpp documentation

How to run from Python code

You can use GGUF models from Python using the llama-cpp-python module.

How to load this model in Python code, using llama-cpp-python

For full documentation, please see: llama-cpp-python docs.

First install the package

Run one of the following commands, according to your system:

# Prebuilt wheel with basic CPU support
pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
# Prebuilt wheel with NVidia CUDA acceleration
pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu121 (or cu122 etc.)
# Prebuilt wheel with Metal GPU acceleration
pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/metal
# Build base version with no GPU acceleration
pip install llama-cpp-python
# With NVidia CUDA acceleration
CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python
# Or with OpenBLAS acceleration
CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python
# Or with AMD ROCm GPU acceleration (Linux only)
CMAKE_ARGS="-DGGML_HIPBLAS=on" pip install llama-cpp-python
# Or with Metal GPU acceleration for macOS systems only
CMAKE_ARGS="-DGGML_METAL=on" pip install llama-cpp-python
# Or with Vulkan acceleration
CMAKE_ARGS="-DGGML_VULKAN=on" pip install llama-cpp-python
# Or with SYCL acceleration
CMAKE_ARGS="-DGGML_SYCL=on -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx" pip install llama-cpp-python

# In windows, to set the variables CMAKE_ARGS in PowerShell, follow this format; eg for NVidia CUDA:
$env:CMAKE_ARGS = "-DGGML_CUDA=on"
pip install llama-cpp-python

Simple llama-cpp-python example code

from llama_cpp import Llama

# Chat Completion API

llm = Llama(model_path="./Qwen2.5-Coder-32B-Instruct.IQ4_XS.gguf", n_gpu_layers=65, n_ctx=131072)
print(llm.create_chat_completion(
    repeat_penalty = 1.05,
    messages = [
        {
            "role": "user",
            "content": "Pick a LeetCode challenge and solve it in Python."
        }
    ]
))

Simple llama-cpp-python example fill-in-middle code

from llama_cpp import Llama

# Completion API

prompt = "def add("
suffix = "\n    return sum\n\n"

llm = Llama(model_path="./Qwen2.5-Coder-32B-Instruct.IQ4_XS.gguf", n_gpu_layers=65, n_ctx=131072)
output = llm.create_completion(
    temperature = 0.0,
    repeat_penalty = 1.0,
    prompt = prompt,
    suffix = suffix
)

# Models sometimes repeat suffix in response, attempt to filter that
response = output["choices"][0]["text"]
response_stripped = response.rstrip()
unwanted_response_suffix = suffix.rstrip()
unwanted_response_length = len(unwanted_response_suffix)

filtered = False
if unwanted_response_suffix and response_stripped[-unwanted_response_length:] == unwanted_response_suffix:
    response = response_stripped[:-unwanted_response_length]
    filtered = True

print(f"Fill-in-Middle completion{' (filtered)' if filtered else ''}:\n\n{prompt}\033[32m{response}\033[{'33' if filtered else '0'}m{suffix}\033[0m")

Simple llama-cpp-python example function calling code

from llama_cpp import Llama

# Chat Completion API

grammar = LlamaGrammar.from_json_schema(json.dumps({
    "type": "array",
    "items": {
        "type": "object",
        "required": [ "name", "arguments" ],
        "properties": {
            "name": {
                "type": "string"
            },
            "arguments": {
                "type": "object"
            }
        }
    }
}))

llm = Llama(model_path="./Qwen2.5-Coder-32B-Instruct.IQ4_XS.gguf", n_gpu_layers=65, n_ctx=131072)
response = llm.create_chat_completion(
      temperature = 0.0,
      repeat_penalty = 1.05,
      messages = [
        {
          "role": "user",
          "content": "What's the weather like in Oslo and Stockholm?"
        }
      ],
      tools=[{
        "type": "function",
        "function": {
          "name": "get_current_weather",
          "description": "Get the current weather in a given location",
          "parameters": {
            "type": "object",
            "properties": {
              "location": {
                "type": "string",
                "description": "The city and state, e.g. San Francisco, CA"
              },
              "unit": {
                "type": "string",
                "enum": [ "celsius", "fahrenheit" ]
              }
            },
            "required": [ "location" ]
          }
        }
      }],
      grammar = grammar
)
print(json.loads(response["choices"][0]["text"]))

print(llm.create_chat_completion(
      temperature = 0.0,
      repeat_penalty = 1.05,
      messages = [
        {
          "role": "user",
          "content": "What's the weather like in Oslo?"
        },
        { # The tool_calls is from the response to the above with tool_choice active
          "role": "assistant",
          "content": None,
          "tool_calls": [
            {
              "id": "call__0_get_current_weather_cmpl-...",
              "type": "function",
              "function": {
                "name": "get_current_weather",
                "arguments": { "location": "Oslo, Norway" , "unit": "celsius" }
              }
            }
          ]
        },
        { # The tool_call_id is from tool_calls and content is the result from the function call you made
          "role": "tool",
          "content": "20",
          "tool_call_id": "call__0_get_current_weather_cmpl-..."
        }
      ],
      tools=[{
        "type": "function",
        "function": {
          "name": "get_current_weather",
          "description": "Get the current weather in a given location",
          "parameters": {
            "type": "object",
            "properties": {
              "location": {
                "type": "string",
                "description": "The city and state, e.g. San Francisco, CA"
              },
              "unit": {
                "type": "string",
                "enum": [ "celsius", "fahrenheit" ]
              }
            },
            "required": [ "location" ]
          }
        }
      }],
      #tool_choice={
      #  "type": "function",
      #  "function": {
      #    "name": "get_current_weather"
      #  }
      #}
))

Qwen2.5-Coder-32B-Instruct

Introduction

Qwen2.5-Coder is the latest series of Code-Specific Qwen large language models (formerly known as CodeQwen). As of now, Qwen2.5-Coder has covered six mainstream model sizes, 0.5, 1.5, 3, 7, 14, 32 billion parameters, to meet the needs of different developers. Qwen2.5-Coder brings the following improvements upon CodeQwen1.5:

• Significantly improvements in code generation, code reasoning and code fixing. Base on the strong Qwen2.5, we scale up the training tokens into 5.5 trillion including source code, text-code grounding, Synthetic data, etc. Qwen2.5-Coder-32B has become the current state-of-the-art open-source codeLLM, with its coding abilities matching those of GPT-4o.
• A more comprehensive foundation for real-world applications such as Code Agents. Not only enhancing coding capabilities but also maintaining its strengths in mathematics and general competencies.
• Long-context Support up to 128K tokens.

This repo contains the instruction-tuned 32B Qwen2.5-Coder model, which has the following features:

• Type: Causal Language Models
• Training Stage: Pretraining & Post-training
• Architecture: transformers with RoPE, SwiGLU, RMSNorm, and Attention QKV bias
• Number of Parameters: 32.5B
• Number of Paramaters (Non-Embedding): 31.0B
• Number of Layers: 64
• Number of Attention Heads (GQA): 40 for Q and 8 for KV
• Context Length: Full 131,072 tokens
  • Please refer to this section for detailed instructions on how to deploy Qwen2.5 for handling long texts.

For more details, please refer to our blog, GitHub, Documentation, Arxiv.

Requirements

The code of Qwen2.5-Coder has been in the latest Hugging face transformers and we advise you to use the latest version of transformers.

With transformers<4.37.0, you will encounter the following error:

KeyError: 'qwen2'

Quickstart

Here provides a code snippet with apply_chat_template to show you how to load the tokenizer and model and how to generate contents.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen2.5-Coder-32B-Instruct"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "write a quick sort algorithm."
messages = [
    {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

Processing Long Texts

The current config.json is set for context length up to 32,768 tokens. To handle extensive inputs exceeding 32,768 tokens, we utilize YaRN, a technique for enhancing model length extrapolation, ensuring optimal performance on lengthy texts.

For supported frameworks, you could add the following to config.json to enable YaRN:

{
  ...,
  "rope_scaling": {
    "factor": 4.0,
    "original_max_position_embeddings": 32768,
    "type": "yarn"
  }
}

For deployment, we recommend using vLLM. Please refer to our Documentation for usage if you are not familar with vLLM. Presently, vLLM only supports static YARN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts. We advise adding the rope_scaling configuration only when processing long contexts is required.

Evaluation & Performance

Detailed evaluation results are reported in this 📑 blog.

For requirements on GPU memory and the respective throughput, see results here.

Citation

If you find our work helpful, feel free to give us a cite.

@article{hui2024qwen2,
      title={Qwen2. 5-Coder Technical Report},
      author={Hui, Binyuan and Yang, Jian and Cui, Zeyu and Yang, Jiaxi and Liu, Dayiheng and Zhang, Lei and Liu, Tianyu and Zhang, Jiajun and Yu, Bowen and Dang, Kai and others},
      journal={arXiv preprint arXiv:2409.12186},
      year={2024}
}
@article{qwen2,
      title={Qwen2 Technical Report}, 
      author={An Yang and Baosong Yang and Binyuan Hui and Bo Zheng and Bowen Yu and Chang Zhou and Chengpeng Li and Chengyuan Li and Dayiheng Liu and Fei Huang and Guanting Dong and Haoran Wei and Huan Lin and Jialong Tang and Jialin Wang and Jian Yang and Jianhong Tu and Jianwei Zhang and Jianxin Ma and Jin Xu and Jingren Zhou and Jinze Bai and Jinzheng He and Junyang Lin and Kai Dang and Keming Lu and Keqin Chen and Kexin Yang and Mei Li and Mingfeng Xue and Na Ni and Pei Zhang and Peng Wang and Ru Peng and Rui Men and Ruize Gao and Runji Lin and Shijie Wang and Shuai Bai and Sinan Tan and Tianhang Zhu and Tianhao Li and Tianyu Liu and Wenbin Ge and Xiaodong Deng and Xiaohuan Zhou and Xingzhang Ren and Xinyu Zhang and Xipin Wei and Xuancheng Ren and Yang Fan and Yang Yao and Yichang Zhang and Yu Wan and Yunfei Chu and Yuqiong Liu and Zeyu Cui and Zhenru Zhang and Zhihao Fan},
      journal={arXiv preprint arXiv:2407.10671},
      year={2024}
}