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

Phi-3-small

Microsoft's Phi-3-small is a member of the Phi family of small language models (SLMs), engineered to deliver high performance within a compact computational footprint. This model variant, with 7 billion parameters, is positioned for broad commercial and research applications where resource efficiency and responsiveness are critical. It addresses scenarios demanding robust language understanding, logical reasoning, and efficient processing on constrained hardware environments, including on-device deployments.

The underlying architecture of Phi-3-small is a dense, decoder-only Transformer. It incorporates several design choices aimed at optimizing performance and memory efficiency, notably leveraging Grouped Query Attention (GQA) where four query heads share a single key-value head, thereby reducing the KV cache footprint. Additionally, the model utilizes alternating layers of dense and blocksparse attention mechanisms, which further contribute to efficient memory management while preserving long-context retrieval capabilities. The training methodology includes a meticulous process of Supervised Fine-tuning (SFT) and Direct Preference Optimization (DPO), ensuring the model's alignment with human preferences and safety guidelines.

Phi-3-small is designed to operate with a default context length of 8,192 tokens (8K), with a further extended variant supporting up to 128,000 tokens through the application of LongRope technology. The model's training regimen involved an extensive dataset comprising 4.8 trillion tokens, derived from a combination of rigorously filtered public documents, high-quality educational content, and synthetically generated data, emphasizing data quality and reasoning density. This enables the model to excel in tasks such as complex language understanding, mathematical problem-solving, and code generation, making it suitable for deployment across various hardware platforms, from cloud-based inference to edge devices and mobile platforms.