Closed anandijain closed 2 years ago
I've experienced a similar, albeit not as drastic, failure on the M1 on v.1.6.3
, where the Michaelis Menten Testset fails. Since there are still some issues with the architecture, I would wait, take a grain of salt and lean back a bit as long as Ubuntu is doing fine.
Since the Cartpole example works right now, I suspect ADM to be the bad guy here. Probably related to the null space computation. But I did not look into it deeper.
TBH, we could simply get rid of ADM, since the ImplicitOptimizer strategy is more effective robust anyways.
Update : Running tests without the ADM
optimiser works. So this is definitely related.
Julia Version 1.6.3
Commit ae8452a9e0 (2021-09-23 17:34 UTC)
Platform Info:
OS: macOS (x86_64-apple-darwin19.5.0)
CPU: Apple M1
WORD_SIZE: 64
LIBM: libopenlibm
LLVM: libLLVM-11.0.1 (ORCJIT, westmere)
Environment:
JULIA_EDITOR = code
JULIA_NUM_THREADS =
Tested on 3 different configs, the two using 1.7-rc1 (fail) one on 1.6 (passes)
Now I kind of understand for error tests that floating point can change things a bit, but the values are too far off for that to be the cause I think.
Pointers on how to dig into this would be helpful. thanks
gave
I tried on a different machine
gave
On this, everything passed