This PR provides with a pass to decompose StableHLO quantized programs using uniform quantize/dequantize
operations. For example, the following program
Per the docs/spec.md, following is the exhaustive list of ops which can be interpreted using dq-op-q strategy. The current PR handles all these ops except for un-bolded ones (DotGeneralOp, ConvolutionOp, DynamicConvOp, and AddOp) which are already lowered to integer match using --stablehlo-legalize-quant-to-intpass.
This PR provides with a pass to decompose StableHLO quantized programs using uniform quantize/dequantize operations. For example, the following program
Per the docs/spec.md, following is the exhaustive list of ops which can be interpreted
using dq-op-q
strategy. The current PR handles all these ops except for un-bolded ones (DotGeneralOp
,ConvolutionOp
,DynamicConvOp
, andAddOp
) which are already lowered to integer match using--stablehlo-legalize-quant-to-int
pass.