Philipp Moritz

### Your current environment H100 80GB, torch 2.3.0 ### 🐛 Describe the bug Since https://github.com/vllm-project/vllm/pull/4454, the performance on the FP8 MoE kernel has been pretty bad. On FP8, for qps...

fix ParseFromString for empty messages

1

Thanks for the package! ParseFromString doesn't work for parsing empty message strings at the moment; this patch should fix it. Running the following script in the `test` directory triggers the...

- [ ] For the plasma store, if a client gets disconnected and another one connects, reuse the slot in plasma_store_state::waiting and plasma_store_store_state::client_id (possibly get rid of client_id) - [...

fix memory leak during broadcasting

4

found and reported by @nhe150, fixes issue #123

This is an issue for tracking the status of implementing https://github.com/amplab/SparkNet/commit/b73ed9780a1a21c5a1f2068bc6cb017857ee6f49 on top of the current JavaCPP architecture. TODOs for completing this: - [x] Integrate the patch for caffe/include/caffe/parallel.hpp and...