FrozenGene commented on issue #4828: [QNN][TFLite] TFLite rounding mode support URL: https://github.com/apache/incubator-tvm/pull/4828#issuecomment-599656124 > Serveral updates: > > 1. tflite conversion segfault is resolved by using the tflite whl provided by @FrozenGene in the tvm tutorial and change tf version to 1.13. > 2. qnn_mobilenet_v2 segfault is resolved by setting opt_level explicitly to 2, specifically disabling the FoldScaleAxis pass (this one is causing the problem) > 3. bit-exact execution remains > > But it is strange that when mobilnet_v2 qnn test is passed locally, it still fails ci test. Below is a local screenshot. >  > > I'm listing my local env here to see if anything need to be aligned: > llvm 9.0.1 (this I can downgrade, but maybe there is a specific version used in the ci so I can try?) > gcc 7.4.0 > Ubuntu 18.04 > cmake 3.15.2 > > Please have another look, thanks! @FrozenGene @anijain2305 @tqchen Thanks for the hard work. For FoldScaleAxis, It will be better we could dig what issue it is (I describe some background of this algorithm here: https://github.com/apache/incubator-tvm/pull/1664/files#diff-88bcab09a087491487d7655fa76ba116R527). I checked CI's docker instance just now. Our CI's LLVM version is LLVM 9 too. So your LLVM is OK IMO. However, I think you could run `make cython3` on your tvm root directory (not `build` directory) if you don't do this. Sometimes, cython will expose problem (CI has it).
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
