search

km2ii / JVP_TG-PLUS

This is enhanced mod for javplayer, enhance the effect after removing mosaic.
319 stars 66 forks source link

cannot use tecoGAN-da-8x-std due to gpu-architecture #100

Closed idl-polun closed 1 year ago

idl-polun commented 1 year ago

Hi I am trying this new model with 4090, however it cannot run and prompt such error at last line nvrtc: error: invalid value for --gpu-architecture (-arch) the detail as attached file da_fail.txt

I also try other model like basicVSR-pp and tecogan, and those are running well. is this because pytorch doesn't support 4090 yet?

tonyyuan1 commented 1 year ago

The same bug occurs with rtx4070ti too. Upgrading to torch 1.13.1 (cuda11.7) didn't help.

idl-polun commented 1 year ago

I upgraded 11.8 + cudnn-windows-x86_64-8.7.0.84 and replace these files in folder TG\TGMAIN\torch\lib\cublas64_11.dll TG\TGMAIN\torch\lib\cublasLt64_11.dll TG\TGMAIN\torch\lib\cudart64_110.dll TG\TGMAIN\torch\lib\cudnn_adv_infer64_8.dll TG\TGMAIN\torch\lib\cudnn_adv_train64_8.dll TG\TGMAIN\torch\lib\cudnn_cnn_infer64_8.dll TG\TGMAIN\torch\lib\cudnn_cnn_train64_8.dll TG\TGMAIN\torch\lib\cudnn_ops_infer64_8.dll TG\TGMAIN\torch\lib\cudnn_ops_train64_8.dll TG\TGMAIN\torch\lib\cudnn64_8.dll TG\TGMAIN\torch\lib\cufft64_10.dll TG\TGMAIN\torch\lib\cufftw64_10.dll TG\TGMAIN\torch\lib\curand64_10.dll TG\TGMAIN\torch\lib\cusolver64_11.dll TG\TGMAIN\torch\lib\cusolverMg64_11.dll TG\TGMAIN\torch\lib\cusparse64_11.dll TG\TGMAIN\torch\lib\nvrtc64_112_0.dll

I didn't chech which one did the magic, but it can run, guess it's nvrtc64_112_0 issue

PhiGo67 commented 1 year ago

I upgraded 11.8 + cudnn-windows-x86_64-8.7.0.84 and replace these files in folder TG\TGMAIN\torch\lib\cublas64_11.dll TG\TGMAIN\torch\lib\cublasLt64_11.dll TG\TGMAIN\torch\lib\cudart64_110.dll TG\TGMAIN\torch\lib\cudnn_adv_infer64_8.dll TG\TGMAIN\torch\lib\cudnn_adv_train64_8.dll TG\TGMAIN\torch\lib\cudnn_cnn_infer64_8.dll TG\TGMAIN\torch\lib\cudnn_cnn_train64_8.dll TG\TGMAIN\torch\lib\cudnn_ops_infer64_8.dll TG\TGMAIN\torch\lib\cudnn_ops_train64_8.dll TG\TGMAIN\torch\lib\cudnn64_8.dll TG\TGMAIN\torch\lib\cufft64_10.dll TG\TGMAIN\torch\lib\cufftw64_10.dll TG\TGMAIN\torch\lib\curand64_10.dll TG\TGMAIN\torch\lib\cusolver64_11.dll TG\TGMAIN\torch\lib\cusolverMg64_11.dll TG\TGMAIN\torch\lib\cusparse64_11.dll TG\TGMAIN\torch\lib\nvrtc64_112_0.dll

I didn't chech which one did the magic, but it can run, guess it's nvrtc64_112_0 issue

I tried the same, but still no luck with my RTX4070TI and Javplayer/TG-Plus :( TecoGAN-DA is unusable for me, but the rest seems to work.

suphamster commented 1 year ago

I upgraded 11.8 + cudnn-windows-x86_64-8.7.0.84 and replace these files in folder TG\TGMAIN\torch\lib\cublas64_11.dll TG\TGMAIN\torch\lib\cublasLt64_11.dll TG\TGMAIN\torch\lib\cudart64_110.dll TG\TGMAIN\torch\lib\cudnn_adv_infer64_8.dll TG\TGMAIN\torch\lib\cudnn_adv_train64_8.dll TG\TGMAIN\torch\lib\cudnn_cnn_infer64_8.dll TG\TGMAIN\torch\lib\cudnn_cnn_train64_8.dll TG\TGMAIN\torch\lib\cudnn_ops_infer64_8.dll TG\TGMAIN\torch\lib\cudnn_ops_train64_8.dll TG\TGMAIN\torch\lib\cudnn64_8.dll TG\TGMAIN\torch\lib\cufft64_10.dll TG\TGMAIN\torch\lib\cufftw64_10.dll TG\TGMAIN\torch\lib\curand64_10.dll TG\TGMAIN\torch\lib\cusolver64_11.dll TG\TGMAIN\torch\lib\cusolverMg64_11.dll TG\TGMAIN\torch\lib\cusparse64_11.dll TG\TGMAIN\torch\lib\nvrtc64_112_0.dll I didn't chech which one did the magic, but it can run, guess it's nvrtc64_112_0 issue

I tried the same, but still no luck with my RTX4070TI and Javplayer/TG-Plus :( TecoGAN-DA is unusable for me, but the rest seems to work.

Maybe u copied mentioned dll files only to "G\TGMAIN" folder but u need its also to copy here "G\TGMAIN\torch\lib" as mentioned in post that u quoted. For me also it didn't worked on RTX 4070 (not Ti) until I've placed dll in both folders. U can take needed libraries from this link but take only those one that names starting with "cu" and "nv" or u'll broke torch! https://download.pytorch.org/libtorch/cu118/libtorch-win-shared-with-deps-2.0.1%2Bcu118.zip