技术规格

系统要求

不同量化方法和上下文大小的显存要求

Llama 3 8B

Meta Llama 3 is a foundational large language model developed by Meta AI, designed to facilitate advanced text and code generation across a diverse range of applications. It is made available in multiple parameter scales, including an 8 billion parameter variant, and is provided in both pre-trained and instruction-tuned forms. The architecture is engineered for scalability and responsible deployment in artificial intelligence systems, supporting various use cases from assistant-style conversational agents to complex natural language processing research tasks.

The model employs a decoder-only transformer architecture, incorporating several technical enhancements over its predecessors. Key innovations include an optimized tokenizer with a 128,000-token vocabulary, which contributes to increased encoding efficiency for language. Additionally, the model integrates Grouped-Query Attention (GQA) across both its 8 billion and 70 billion parameter versions, a modification aimed at improving inference efficiency. For enhanced training stability, Llama 3 utilizes Root Mean Square Normalization (RMSNorm) applied as pre-normalization and employs the SwiGLU activation function. Positional encodings within the model are handled through Rotary Positional Embeddings (RoPE).

Llama 3 8B has been pre-trained on a vast corpus exceeding 15 trillion tokens sourced from publicly available datasets, representing a substantial increase in training data volume compared to prior Llama iterations. This model supports a context length of 8,192 tokens. It demonstrates capabilities in generating coherent text, assisting with code completion, and engaging in conversational tasks, and its capabilities extend to multiple languages and tool use in later iterations (Llama 3.1).

没有可用的 Llama 3 8B 评估基准。