TensorRT icon indicating copy to clipboard operation
TensorRT copied to clipboard

[stable diffusion] [unet] [compiling] faild to find implemtation ForeignNode[/up_blocks.0/resnets.0/time_mixer/Constant_1_output_0.../conv_act/Mul

Open focusunsink opened this issue 11 months ago • 7 comments

focusunsink avatar Feb 19 '25 06:02 focusunsink

GPU: 4090 or H800. tensorrt 10.6.0 torch 2.4.0

when I compile unet using tensorrt. it raise this error. Even I use H800 to make more memory available. This error still happens.

focusunsink avatar Feb 19 '25 06:02 focusunsink

This is because image is to large or frames is too many. when I change frames from 72 to 16. it success.

focusunsink avatar Feb 19 '25 06:02 focusunsink

@asfiyab-nvidia perhaps it'll be useful to provide a table denoting VRAM requirements for different pipelines in the README

kevinch-nv avatar Mar 07 '25 19:03 kevinch-nv

This could be avoided when I set config.num_tactics = 2 or other small number. So search tactics may consume much VRAM

focusunsink avatar Mar 08 '25 02:03 focusunsink

@kevinch-nv sounds good. We can release VRAM documentation for pipelines in the next release. @focusunsink have you tried lowering the builder optimization level using the --optimization_level flag? The default value is 3 for most pipelines. You can reduce it for further memory savings

asfiyab-nvidia avatar Mar 11 '25 20:03 asfiyab-nvidia

I set config.builder_optimization_level = 3, still not work.

focusunsink avatar Mar 12 '25 03:03 focusunsink

I set workspace=8 config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, workspace << 30) then it works

focusunsink avatar Mar 12 '25 06:03 focusunsink