Skip to navigation
DocumentationFine-tune ModelsModels

Phi Models

View as Markdown

This page provides detailed technical specifications for the Phi model family supported by NeMo Customizer. For information about supported features and capabilities, refer to Tested Models.

Deployment configurations on this page that use the multi-LLM NIM (nvcr.io/nim/nvidia/llm-nim:1.15.5) cannot serve merged LoRA checkpoints (finetuning_type: lora_merged). Deploy merged LoRA models with vLLM or the model-free NIM; see Choose an Inference Engine.

Microsoft Phi-4

PropertyValue
CreatorMicrosoft
ArchitectureDecoder-only Transformer
DescriptionPhi-4 is Microsoft’s most advanced small language model, designed to deliver strong reasoning capabilities while being efficient to deploy.
Max I/O Tokens16K
Parameters14 billion
Training DataHigh-quality data with emphasis on reasoning and code
Default Namemicrosoft/phi-4
Hugging Facemicrosoft/phi-4

Training Options

  • LoRA: 2x 80GB GPU, tensor parallel size 2, micro batch size 2
  • Full SFT: 4x 80GB GPU, tensor parallel size 2

Deployment Configuration

  • LoRA:
  • NIM Image: nvcr.io/nim/nvidia/llm-nim:1.15.5
  • Full SFT:
  • NIM Image: nvcr.io/nim/nvidia/llm-nim:1.15.5
  • Additional Environment Variables:
  • NIM_MODEL_PROFILE: vllm