Feat (minifloat): add support for user specified minifloat format#821

Merged

Giuseppe5 merged 15 commits intoXilinx:devfrom fabianandresgrob:fp8

Mar 7, 2024

+453-38

Commits on Mar 7, 2024

Feat (FP8): implements conversion from fp32 to user specified fp8 format
fabianandresgrob
committed
Feat (minifloat): calculates max_value based on user format
fabianandresgrob
committed
Tests (minifloat): adds tests and fixes bugs
fabianandresgrob
committed
Fix (minifloat): restructure FloatClamp and max val calculation
fabianandresgrob
committed
Fix (minifloat): restructure MaxValue and clamp computation
fabianandresgrob
committed
remove second max val class
fabianandresgrob
committed
Fix (minifloat): make MaxFloatInfNaN jit compatible
fabianandresgrob
committed
Fix (minifloat): fix private attr for jit and make hypothesis tests work with jit
fabianandresgrob
committed
Fix (minifloat): fix jit issues with FloatClamp
fabianandresgrob
committed
Feat (minifloat): use tensor_clamp_impl for minifloat conversion and clean up
fabianandresgrob
committed
Fix (minifloat): compute max_value during dependency injection
fabianandresgrob
committed
Fix (minifloat): add check for saturating and clean up
fabianandresgrob
committed
Fix (minifloat): restructuring quantizers
fabianandresgrob
committed
Fix (minifloat): restructure OCP format quantizers
fabianandresgrob
committed
Fix (minifloat): rename clamping test
fabianandresgrob
committed