Mozilla packaged the Qwen 2.5 models into executable weights that we call llamafiles. This gives you the easiest fastest way to use the model on Linux, MacOS, Windows, FreeBSD, OpenBSD and NetBSD systems you control on both AMD64 and ARM64.

Software Last Updated: 2025-03-31

Llamafile Version: 0.9.2

Quickstart

To get started, you need both the Qwen 2.5 weights, and the llamafile software. Both of them are included in a single file, which can be downloaded and run as follows:

wget https://huggingface.co/Mozilla/Qwen2.5-14B-Instruct-1M-llamafile/resolve/main/Qwen2.5-14B-Instruct-1M-Q6_K.llamafile
chmod +x Qwen2.5-14B-Instruct-1M-Q6_K.llamafile
./Qwen2.5-14B-Instruct-1M-Q6_K.llamafile

The default mode of operation for these llamafiles is our new command line chatbot interface.

Usage

You can use triple quotes to ask questions on multiple lines. You can pass commands like /stats and /context to see runtime status information. You can change the system prompt by passing the -p "new system prompt" flag. You can press CTRL-C to interrupt the model. Finally CTRL-D may be used to exit.

If you prefer to use a web GUI, then a --server mode is provided, that will open a tab with a chatbot and completion interface in your browser. For additional help on how it may be used, pass the --help flag. The server also has an OpenAI API compatible completions endpoint that can be accessed via Python using the openai pip package.

./Qwen2.5-14B-Instruct-1M-Q6_K.llamafile --server

An advanced CLI mode is provided that's useful for shell scripting. You can use it by passing the --cli flag. For additional help on how it may be used, pass the --help flag.

./Qwen2.5-14B-Instruct-1M-Q6_K.llamafile --cli -p 'four score and seven' --log-disable

Troubleshooting

Having trouble? See the "Gotchas" section of the README.

On Linux, the way to avoid run-detector errors is to install the APE interpreter.

sudo wget -O /usr/bin/ape https://cosmo.zip/pub/cosmos/bin/ape-$(uname -m).elf
sudo chmod +x /usr/bin/ape
sudo sh -c "echo ':APE:M::MZqFpD::/usr/bin/ape:' >/proc/sys/fs/binfmt_misc/register"
sudo sh -c "echo ':APE-jart:M::jartsr::/usr/bin/ape:' >/proc/sys/fs/binfmt_misc/register"

On Windows there's a 4GB limit on executable sizes.

Context Window

This model has a max context window size of 128k tokens. By default, a context window size of 8192 tokens is used. You can ask llamafile to use the maximum context size by passing the -c 0 flag. That's big enough for a small book. If you want to be able to have a conversation with your book, you can use the -f book.txt flag.

GPU Acceleration

On GPUs with sufficient RAM, the -ngl 999 flag may be passed to use the system's NVIDIA or AMD GPU(s). On Windows, only the graphics card driver needs to be installed if you own an NVIDIA GPU. On Windows, if you have an AMD GPU, you should install the ROCm SDK v6.1 and then pass the flags --recompile --gpu amd the first time you run your llamafile.

On NVIDIA GPUs, by default, the prebuilt tinyBLAS library is used to perform matrix multiplications. This is open source software, but it doesn't go as fast as closed source cuBLAS. If you have the CUDA SDK installed on your system, then you can pass the --recompile flag to build a GGML CUDA library just for your system that uses cuBLAS. This ensures you get maximum performance.

For further information, please see the llamafile README.

About llamafile

llamafile is a new format introduced by Mozilla on Nov 20th 2023. It uses Cosmopolitan Libc to turn LLM weights into runnable llama.cpp binaries that run on the stock installs of six OSes for both ARM64 and AMD64.

Qwen2.5-14B-Instruct-1M

Introduction

Qwen2.5-1M is the long-context version of the Qwen2.5 series models, supporting a context length of up to 1M tokens. Compared to the Qwen2.5 128K version, Qwen2.5-1M demonstrates significantly improved performance in handling long-context tasks while maintaining its capability in short tasks.

The model 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: 14.7B
Number of Paramaters (Non-Embedding): 13.1B
Number of Layers: 48
Number of Attention Heads (GQA): 40 for Q and 8 for KV
Context Length: Full 1,010,000 tokens and generation 8192 tokens
- We recommend deploying with our custom vLLM, which introduces sparse attention and length extrapolation methods to ensure efficiency and accuracy for long-context tasks. For specific guidance, refer to this section.
- You can also use the previous framework that supports Qwen2.5 for inference, but accuracy degradation may occur for sequences exceeding 262,144 tokens.

For more details, please refer to our blog, GitHub, Technical Report, and Documentation.

Requirements

The code of Qwen2.5 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.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen2.5-14B-Instruct-1M"

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

prompt = "Give me a short introduction to large language model."
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 Ultra Long Texts

To enhance processing accuracy and efficiency for long sequences, we have developed an advanced inference framework based on vLLM, incorporating sparse attention and length extrapolation. This approach significantly improves model generation performance for sequences exceeding 256K tokens and achieves a 3 to 7 times speedup for sequences up to 1M tokens.

Here we provide step-by-step instructions for deploying the Qwen2.5-1M models with our framework.

1. System Preparation

To achieve the best performance, we recommend using GPUs with Ampere or Hopper architecture, which support optimized kernels.

Ensure your system meets the following requirements:

CUDA Version: 12.1 or 12.3
Python Version: >=3.9 and <=3.12

VRAM Requirements:

For processing 1 million-token sequences:
- Qwen2.5-7B-Instruct-1M: At least 120GB VRAM (total across GPUs).
- Qwen2.5-14B-Instruct-1M: At least 320GB VRAM (total across GPUs).

If your GPUs do not have sufficient VRAM, you can still use Qwen2.5-1M for shorter tasks.

2. Install Dependencies

For now, you need to clone the vLLM repository from our custom branch and install it manually. We are working on getting our branch merged into the main vLLM project.

bash

git clone -b dev/dual-chunk-attn git@github.com:QwenLM/vllm.git
cd vllm
pip install -e . -v

3. Launch vLLM

vLLM supports offline inference or launch an openai-like server.

Example of Offline Inference

python

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

# Initialize the tokenizer
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-14B-Instruct-1M")

# Pass the default decoding hyperparameters of Qwen2.5-14B-Instruct
# max_tokens is for the maximum length for generation.
sampling_params = SamplingParams(temperature=0.7, top_p=0.8, repetition_penalty=1.05, max_tokens=512)

# Input the model name or path. See below for parameter explanation (after the example of openai-like server).
llm = LLM(model="Qwen/Qwen2.5-14B-Instruct-1M",
    tensor_parallel_size=4,
    max_model_len=1010000,
    enable_chunked_prefill=True,
    max_num_batched_tokens=131072,
    enforce_eager=True,
    # quantization="fp8", # Enabling FP8 quantization for model weights can reduce memory usage.
)

# Prepare your prompts
prompt = "Tell me something about large language models."
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
)

# generate outputs
outputs = llm.generate([text], 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}")

Example of Openai-like Server

bash

vllm serve Qwen/Qwen2.5-14B-Instruct-1M \
  --tensor-parallel-size 4 \
  --max-model-len 1010000 \
  --enable-chunked-prefill --max-num-batched-tokens 131072 \
  --enforce-eager \
  --max-num-seqs 1

# --quantization fp8  # Enabling FP8 quantization for model weights can reduce memory usage.

Then you can use curl or python to interact with the deployed model.

Parameter Explanations:

--tensor-parallel-size
- Set to the number of GPUs you are using. Max 4 GPUs for the 7B model, and 8 GPUs for the 14B model.
--max-model-len
- Defines the maximum input sequence length. Reduce this value if you encounter Out of Memory issues.
--max-num-batched-tokens
- Sets the chunk size in Chunked Prefill. A smaller value reduces activation memory usage but may slow down inference.
- Recommend 131072 for optimal performance.
--max-num-seqs
- Limits concurrent sequences processed.

You can also refer to our Documentation for usage of vLLM.

Troubleshooting:

Encountering the error: "The model's max sequence length (xxxxx) is larger than the maximum number of tokens that can be stored in the KV cache."

The VRAM reserved for the KV cache is insufficient. Consider reducing the max_model_len or increasing the tensor_parallel_size. Alternatively, you can reduce max_num_batched_tokens, although this may significantly slow down inference.
Encountering the error: "torch.OutOfMemoryError: CUDA out of memory."

The VRAM reserved for activation weights is insufficient. You can try setting gpu_memory_utilization to 0.85 or lower, but be aware that this might reduce the VRAM available for the KV cache.
Encountering the error: "Input prompt (xxxxx tokens) + lookahead slots (0) is too long and exceeds the capacity of the block manager."

The input is too lengthy. Consider using a shorter sequence or increasing the max_model_len.

Evaluation & Performance

Detailed evaluation results are reported in this 📑 blog and our technical report.

Citation

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

@misc{qwen2.5-1m,
    title = {Qwen2.5-1M: Deploy Your Own Qwen with Context Length up to 1M Tokens},
    url = {https://qwenlm.github.io/blog/qwen2.5-1m/},
    author = {Qwen Team},
    month = {January},
    year = {2025}
}

@article{qwen2.5,
      title={Qwen2.5-1M Technical Report}, 
      author={An Yang and Bowen Yu and Chengyuan Li and Dayiheng Liu and Fei Huang and Haoyan Huang and Jiandong Jiang and Jianhong Tu and Jianwei Zhang and Jingren Zhou and Junyang Lin and Kai Dang and Kexin Yang and Le Yu and Mei Li and Minmin Sun and Qin Zhu and Rui Men and Tao He and Weijia Xu and Wenbiao Yin and Wenyuan Yu and Xiafei Qiu and Xingzhang Ren and Xinlong Yang and Yong Li and Zhiying Xu and Zipeng Zhang},
      journal={arXiv preprint arXiv:2501.15383},
      year={2025}
}