- Highlights
- Features
- Improvements
- Validated Hardware
- Validated Configurations
Highlights
- INT8 quantization for ViT model for JAX framework
Features
- Support INT8 quantization for ViT model for Keras/JAX
- Support Out-of-place quantization for Keras/JAX
- Support quantization of BFloat16 models for Keras/JAX
- Support quantization for Conv2D, dot_product_attention for Keras/JAX
Improvements
- Improving performance with constant scales and weights support for Keras/JAX
- Handling uncalibrated model layers for Keras/JAX
- WAN series MXFP8 PTQ example
- DeepSeek V4 series mixed MXFP4/MXFP8 example
Validated Hardware
- Intel Gaudi Al Accelerators (Gaudi 2 and 3)
- Intel Xeon Scalable processor (4th, 5th and 6th Gen)
- Intel® Arc™ B-Series Graphics GPU (B60)
Validated Configurations
- Ubuntu 24.04 & Win 11
- Python 3.11, 3.12, 3.13
- PyTorch 2.10, 2.11, 2.12
- JAX 0.10
Notes
- It is recommended to use version v3.9 or later to mitigate code CVEs.