Yihua Cheng
Hi, I had the same issue when using ubuntu 18.04 with `libx264-dev` version 2:0.152.2854+gite9a5903-2. I fixed this by changing the functions that started with `x264_8_` to `x264_` in ` src/util/ssim.cc`....
cc @maobaolong @YaoJiayi @Shaoting-Feng @hickeyma @KuntaiDu @sammshen Please feel free to leave your thoughts here, thanks!
@tensorflowt Hey, can you share the configuration file and the workload so that we can better understand the problem? Thanks!
@tensorflowt Thanks for sharing! Currently, when the workload is too heavy (i.e., the decode cannot process all the requests in real time), there will be a known issue that causes...
The current communication protocol between P and D is that the P worker will first send the KV cache to the D worker, and the D worker has a temporary...
XpYd support is WIP now. There should be a PR in next week.
@lsyyysky Can you try using this as the proxy: https://github.com/LMCache/LMCache/blob/dev/examples/disagg_prefill/1p1d/disagg_proxy_server_first_token_from_decoder.py The `disagg_proxy_server_first_token_from_prefiller` is an experimental implementation that depends on a recent PR in vLLM (vllm-project/vllm#19441). So you need to install...
@csbo98 Thanks for your interest! This is still an unresolved issue, but it's quite old and many people may not be aware of this. I suggest to create a new...
@YaoJiayi @hickeyma Do you guys think building a docker image with all the dependencies for running the unit tests is a good idea?
@hickeyma nixl requires a bunch of dependencies, such as UCX. It would be better to use a base docker image with all those dependencies installed (like the "cuda_dl_base_xxxx" one we...