TbeQuantConv2DElemWiseFusionPass

Description

Performs UB fusion on the QuantConv2D and elemwise/broadcast nodes. The elemwise trustlist is Add/Div/Realdiv.

Restrictions

• Only the cascade structures involved in the SD2.1 and SDXL networks are supported, that is, QuantConv2D+Add, QuantConv2D+Add Div, and QuantConv2D+Add+Realdiv in the static scenario.
• Only static scenarios of Atlas inference product are supported.
• The QuantConv2D node applies to the following scenarios: offset, group > 1, DMA, and Nx1 non-fusion.
• The elemwise supports only dual inputs and single output, and the static trustlist must be Add/Div/Realdiv. Otherwise, it is not fused.
• The number of elemwise nodes is 1≤ N ≤ 2. When N = 2, the first elemwise cannot be multi-reference output. Otherwise, it is not fused.

Availability

Atlas inference product