Qwen2.5-Coder-14B-Instruct

Qwen2.5-Coder-14B-Instruct is an instruction fine-tuning model optimized for code tasks developed by Qwen. It is suitable for code generation, reasoning, debugging and other application scenarios.

Core features

1. Model architecture

• Contains 48 Transformer layers, using rotation position embedding (RoPE), SwiGLU activation function, RMSNorm normalization and attention mechanism with QKV bias.

• Using Grouped Query Attention (GQA), there are 40 query headers and 8 key-value headers, designed for efficient code processing.

2. Parameter quantity

• The total number of parameters is 14.7 billion, of which 13.1 billion are used for the non-embedded part.

3. context length

• Supports context lengths up to 131,072 tokens and supports handling of large code bases and long documents through YaRN technology.

4. Performance

• Significantly superior performance in code generation, inference, and code repair, as well as strong performance in mathematical calculations and general-purpose tasks.

Model variants

• The basic model provides a variety of parameter sizes, including 0.5B, 1.5B, 3B, 7B, 14B and 32B, suitable for code completion and basic tasks.

• Instruction fine-tuning model Optimized for interactive tasks such as code generation and debugging, the 14B-Instruct model is ideal for chat-based application scenarios.

Deployment requirements

• Python version : 3.9 or higher.

• Transformers library : version 4.37.0 or higher, supports the integration of Qwen2 series models.

Quick to use

The sample code for loading a model using Hugging Face's transformers library is as follows:

 from transformers import AutoModelForCausalLM, AutoTokenizer

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

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

This model can efficiently complete tasks such as code generation and debugging.