Open
Description
Steps to reproduce:
- Access https://tensorflow.github.io/tfjs/e2e/benchmarks/local-benchmark/index.html
- Use the default backend
wasm
and clickRun benchmark
. - See the left table
Peak memory | 19.63 MB |
---|
- Switch to
webgl
backend. Change any flags, for examplewebgl pack
. And then switch back towasm
. - Click
Run benchmark
again. The peak memory increases as below:
Peak memory | 33.51 MB |
---|
- If we repeat 4 and 5, you will find the peak memory gets bigger and bigger. However, if you don't change the flag, the peak memory won't change. It happens for all backends that the flag will affect the peak memory. It seems like a bug.