Skip to content

Latest commit

 

History

History
29 lines (21 loc) · 1.05 KB

File metadata and controls

29 lines (21 loc) · 1.05 KB

Qwen-Qwen3.5-2B — CUDA optimization

This folder contains Olive recipes for optimizing Qwen-Qwen3.5-2B targeting the CUDA EP.

What this folder is for

  • Execution Provider: CUDA EP
  • Typical precision: INT4 precision by default
  • Example recipe filename: Qwen-Qwen3.5-2B_cuda_int4.json

Setup

  1. Install the main branch of Olive:
  2. Install the appropriate runtime package for this backend:
    • onnxruntime-genai-cuda (CUDA build)
  3. Run Olive to build/optimize the model
    • olive run --config Qwen-Qwen3.5-2B_cuda_int4.json

Additional notes:

  • Pipeline: SelectiveMixedPrecision (kld_gradient) → GPTQRTN (8-bit lm_head/embeddings) → ModelBuilderTieWordEmbeddings
  • GPTQ group size: 128
  • Uses text-only mode (exclude_embeds=false, prune_lm_head=true) for standalone LLM inference without multimodal pipeline.
  • Requires NVIDIA GPU with CUDA support.
  • Ensure CUDA toolkit and cuDNN are properly installed.

This README was auto-generated for the CUDA EP of Qwen-Qwen3.5-2B.