Hu Dong

> Same comment as in #370, can you please do a quick test with `0.4.0-SNAPSHOT`? NP. will give it a try.

It seems the memory leak problem still exists. The memory usage of `TString.tensorOf` calls grows as the number of iterations increases. **1024 iterations**  **65536 iterations** 

I did another test by setting JVM max heap size after upgrading TF java to `0.4.0-SNAPSHOT`. With `-Xmx47m`, 4 iterations can pass while 65536 iterations would cause OOM. 

Hi, guys! You could retry compiling the project with following suggestions: 1. rename the IDEA sdk from something like "IDEA IU-135.123" to "IDEA". 2. delete line 13 ~ line 15...

Just run into the same problem with `transport='rest'` and `send_message_async`.

Disabling the custom all-reduce functionality with the `--disable-custom-all-reduce` flag resolves the issue. We've tested this successfully with both the `vllm:0.4.1` and `vllm:0.4.2` Docker images, without needing to enable eager mode...

This plugin is not in active development. But PRs are welcome.

Just encountered the same problem after updating it to 0.64.0. However, downgrading to 0.63.0 works for me, without any configuration change.