v0.1.8

Bug fixes, add SALT attn:

• mask leak in stable softmax impl
• ebops scale factor handling
• allow q/k/v proj fusing for mha (effect is minor - not needed in general)
• beta setter fix: skip layers disabled ebops computation
• DAIS tracer fixes

Full Changelog: v0.1.7...v0.1.8

v0.1.7

• allow affined unary function LUT
• minor bug fixes
• Disable input quantizer for softmax by default
  • in stable impl, the inputs will be effectively quantized twice if not with pre-exp-lookup quantization

v0.1.6

• pf support for pooling layers
• add back dummy pf for einsum dense layer
• allow ebops scaling factor per layer
• test cov of all layers in da4ml
• explicit pf computation in conv for max pf case
• convlut fix
• breaking shuffle is off by default now in datasets.

Full Changelog: v0.1.5...v0.1.6

v0.1.5

• QDenseT/QConvT
• Dataloader shuffle support
• int bits decay fix in WRAP quantizer

Full Changelog: v0.1.4...v0.1.5

v0.1.4

• Allow .h5 or .weights.h5 saving for pareto
• Fix parallelization effect on ebops for dense

v0.1.3

• Many minor bug fixes
• RNN/GRU support
• Pooling signature fix
• DenseT support
• qkeras emu fix
• remove "hls4ml can't synth this" warning

v0.1.2

• Sugar: Minor bug fix for X only dataset sugar (y cannot be None)
• Sugar: Add EBOPs conditioned early stopping callback
• Windows specific: crash due to int32 overflow in seed def

v0.1.1

• Update tests for da4ml api change
• Add linformer grammar sugar

v0.1.0

• First release