Matthias Langer

Hi, sorry for the delay. I am currently travelling overseas. I get back to you later this week with some benchmark results.

Well, that is a now a big question. Very hard to answer. I admit, I may have been chasing a ghost here. I had a use-case, and there it was...

Actually, This could be the root cause of the issue. With 64 byte, any unaligned memory access will require hitting 2 cache lines. With a 32 byte lookup like SSE,...

> following up with this and the hard work. I think the group lookup starts at the index determined by the hash, so I'm not sure how it can be...

Closing this for now. @AtekiRyu Please don't hesitate to reopen this ticket if you have trouble using or face performance issues with the FTRL or any other optimizer.

**Regarding 2:** Using the `parallel_hash_map` as your `volatile_db` is the suggested approach, if you cannot put the entire embedding table directly into the GPU. **Regarding 3:** For performance reasons (avoid...

To be honest, I am aware of the problem for quite some time. I kind of ignored it to avoid compatibility problems. I think it should be fixed. @yingcanw Are...

@rhdong : How come, I never can see the output of these checks? On my side it always says: ``` Error: We are currently unable to download the log. Please...

@rhdong Well that explains it (see error below). It seems that logs are discarded more or less just less than 3 days after they were generated. Can you increase the...

@rhdong What could case this? @MoFHeka ``` root@13beb51e47e9:/# python Python 3.7.7 (default, Jan 19 2022, 04:18:49) [GCC 7.5.0] on linux Type "help", "copyright", "credits" or "license" for more information. >>>...