hako-mikan
commented
10 months ago TensorRT has a unique structure, making it difficult to modify with CD-Tuner. However, by adjusting the model before converting it to TensorRT using the Adjust feature of SuperMerger, you can achieve similar effects in TensorRT.
https://github.com/NVIDIA/Stable-Diffusion-WebUI-TensorRT
d1,d2,con1,hd1,hd2,hrs The above items are not valid when TensorRT is enabled. st1,st2 are not sure. con2,bri,col1,col2,col3 are valid.
Activating unet: [TRT] cetfusion_v10 Loading TensorRT engine: D:\A1111 Web UI Autoinstaller\stable-diffusion-webui\models\Unet-trt\cetfusion_v10_23cd1759_cc89_sample=1x4x32x32+2x4x128x128+2x4x256x256-timesteps=1+2+2-encoder_hidden_states=1x77x768+2x77x768+2x231x768.trt [W] 'colored' module is not installed, will not use colors when logging. To enable colors, please install the 'colored' module: python3 -m pip install colored [I] Loading bytes from D:\A1111 Web UI Autoinstaller\stable-diffusion-webui\models\Unet-trt\cetfusion_v10_23cd1759_cc89_sample=1x4x32x32+2x4x128x128+2x4x256x256-timesteps=1+2+2-encoder_hidden_states=1x77x768+2x77x768+2x231x768.trt Profile 0: sample = [(1, 4, 32, 32), (2, 4, 128, 128), (2, 4, 256, 256)] timesteps = [(1,), (2,), (2,)] encoder_hidden_states = [(1, 77, 768), (2, 77, 768), (2, 231, 768)] latent = [(-1946141184), (-1946140928), (-1946155776)]
CD Tuner Effective : [10, 10, 20, 0, 0, 0, 0, 0, 10, 10, 1, -1, -1, 0] 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10/10 [00:01<00:00, 7.67it/s] 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 13.55it/s] Total progress: 100%|██████████████████████████████████████████████████████████████████| 30/30 [00:11<00:00, 2.62it/s] Total progress: 100%|██████████████████████████████████████████████████████████████████| 30/30 [00:11<00:00, 12.43it/s]