✨[Feature] Automatically remove output repacking for convert_method_to_trt_engine
calls
#1938
Labels
feature request
New feature or request
Context
For fully-supported models having multiple output Tensors, Torch often appends a Tuple/List packing function call to the scripted IR. This can cause
convert_method_to_trt_engine
, which would otherwise succeed due to converter coverage, to fail.Details
This proposed feature would automatically detect and remove unnecessary output packing schemes, such as
prim::TupleConstruct
andprim::ListConstruct
, which are applied to the outputs of existing TRT engines, thereby allowing more models to be fully converted to their TRT engine counterparts directly.The text was updated successfully, but these errors were encountered: