Possible bug RPC Server - Kimi K2 - Apple Silicon

[alejandrods]

I'm encountering an issue with distributed inference using Kimi-K2-Instruct-0905 IQ4_NL quantization across 2 Mac Studios (512GB each) via RPC. The model loads successfully but only generates repeated zeros ("0000000...").

Command I ran in the host:

./llama.cpp/llama-server -hf unsloth/Kimi-K2-Instruct-0905-GGUF:IQ4_NL --host 0.0.0.0 --port 11300 --ctx-size 130736 --temp 0.7 --min-p 0.0 --top-p 0.80 --top-k 20 --parallel 1 --presence-penalty 1.0 --rpc host_ip:11400

Command I ran in the RPC:

./llama.cpp/rpc-server -p 11400 --host 0.0.0.0 -c

What do you think it may be the problem? Thank you!