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-medium

Phi-3-medium is a compact, high-performance large language model developed by Microsoft, belonging to the Phi-3 family of models. With 14 billion parameters, it is designed for a broad array of commercial and research applications, particularly those operating within memory or compute-constrained environments and latency-sensitive scenarios. This model aims to provide strong reasoning capabilities, notably in mathematics, logic, and code generation, positioning it as a foundational component for developing generative artificial intelligence features.

The training methodology for Phi-3-medium leverages a high-quality, reasoning-dense dataset, which is a refined and scaled version of the data utilized for its predecessor, Phi-2. This dataset incorporates both meticulously filtered publicly available web content and synthetically generated data, ensuring a robust and instruction-adherent model. The training process includes supervised fine-tuning (SFT) and direct preference optimization (DPO) to enhance its ability to follow instructions precisely and to reinforce safety measures.

The model employs a dense decoder-only Transformer architecture, a common and effective structure for autoregressive language modeling tasks. Its internal mechanisms include Grouped Query Attention (GQA) for efficient memory utilization and processing, Root Mean Square (RMS) normalization for stable training, and Rotary Positional Embeddings (RoPE) to handle positional information within sequences. A specific variant of RoPE, known as LongRope, facilitates the model's capacity to process extended context lengths up to 128,000 tokens. Phi-3-medium is optimized for deployment across diverse hardware, including graphics processing units (GPUs), central processing units (CPUs), and mobile devices, often leveraging technologies like ONNX Runtime and DirectML for cross-platform compatibility and efficient inference.