趋近智
注意力结构
Multi-Head Attention
隐藏维度大小
-
层数
40
注意力头
-
键值头
-
激活函数
-
归一化
RMS Normalization
位置嵌入
Absolute Position Embedding
不同量化方法和上下文大小的显存要求
The GLM-4-9B model, developed by THUDM (Tsinghua University Department of Computer Science and Technology) and Z.ai, represents an open-source iteration within the GLM-4 series of pre-trained language models. This model is engineered for general language tasks, exhibiting capabilities in semantic understanding, mathematical reasoning, code execution, and knowledge retrieval. It is designed to handle multilingual inputs and outputs, supporting 26 languages including Chinese, English, Japanese, Korean, and German. The GLM-4-9B also supports advanced functionalities such as web browsing, code execution, and custom tool calling through a Function Call mechanism.
Architecturally, GLM-4-9B employs a transformer architecture, which is a common deep learning structure for natural language processing. The model incorporates an autoregressive blank infilling approach for its pre-training phase. The architecture includes specific design choices such as the removal of bias terms except for those in the Query, Key, and Value (QKV) components of attention layers, which contributes to improved length extrapolation. It also utilizes RMSNorm for normalization. The base version of GLM-4-9B supports a context length of up to 128,000 tokens, with specialized variants offering an extended context length of up to 1 million tokens.
GLM-4-9B is engineered for a range of applications, including conversational AI assistants, content generation, and question answering systems. Its design facilitates integration with the Hugging Face Transformers library, simplifying deployment and adoption for developers. The model aims to provide a balance between efficiency and effectiveness, making it suitable for scenarios with resource constraints while maintaining performance across diverse tasks.
General Language Models from Z.ai
排名适用于本地LLM。
没有可用的 GLM-4-9B 评估基准。