Closed sergachev closed 1 month ago
I just saw this PR. We're aware of the problem and I'm actually working on a different solution that also excludes the RAM, but also makes other changes to the cache string. It should be ready today.
FYI: the alternative PR is ready, but we need to wait until Thursday to submit it. Conceptually, it's very similar to this one.
Then what's the reason to prefer it?
Here is the current version of the other PR: https://github.com/openxla/xla/pull/14954
Then what's the reason to prefer it?
The other PR doesn't change the model_str in cuda executor. It touches only the autotuning key and does a cleanup related to deviceless autotuning keys. It also does changes that are needed internally (and not visible on GitHub). At the point I became aware of your PR we already had the internal one nearly done. It's now fully ready but we have to wait for an extra internal approval until tomorrow.
Exact DRAM size can slightly vary for the same GPU model so it shouldn't be used in model identification. The removed part of gpu_compiler_test_autotune_db.textproto illustrates that.