Do anybody have troubles with running multiple sxm2 tesla v100 (through pci-e adapter) like me?

[pl752]

I used one 3090 previously, but now I have purschased two aftermarket sxm2 tesla v100 16gb with sxm2 to pci-e adapters, these by themselves seem to work just fine (in llama.cpp (when used for non-moe models and one at a time), comfy.ui, stress tests, torch model training, etc.). But in llama.cpp I have encountered two different problems.

1. When I use UMA (GGML_CUDA_UNIFIED_MEMORY=1) to avoid random OOM errors in situations where memory margins are very tight, it works just fine with one v100 or 3090+v100, but if I select two v100 at the same time (with or without 3090), token generation and processing speeds drop to a crawl and gibberish is generated instead of text, also in nvtop both v100 are reported as fully utilized, while in normal mode they are displayed as 50% used during generation.
2. When I try to use this build with models which use mxfp4 (like gpt-oss-*) I encounter the problem with using -ncmoe flag. When I select 3090 as a main gpu (computes offloaded layers during preprocessing), add any of the v100 or both to the device list, and load at least two layers onto it, then I get token 30 (?) in a loop when trying to work with sequence length > ~32k (sometimes > ~20k)
   Command is ~/llama.cpp/build/bin/llama-server -m /models/llm_models/gpt-oss-120b-mxfp4-00001-of-00003.gguf -ngl 900 -c 65536 --port 5001 --host 0.0.0.0 -fa on -dev CUDA0,CUDA1,CUDA2 -ts 6,2,2 -ncmoe 12 -v -ub 4096 -b 16384 Other flags aside from -ncmoe don't change the behavior (-cmoe works just fine), and if I select any of the v100 as a main gpu or physically unplug one of the v100 from system, it all just starts to work fine.
   Build command used is: CUDA_HOME=/usr/local/cuda CUDACXX=/usr/local/cuda/bin/nvcc cmake -S . -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="70;86" -DCMAKE_BUILD_TYPE=Release && cmake --build build --config Release -t llama-server -- -j 16
   System used is ubuntu 24.04, llama.cpp from master branch (behavior didn't change when I tried building other version (b7400)), cpu 5800x3d, b550 motherboard, 3090 is plugged into cpu pci-e 4x16, two v100 are plugged into chipset pci-e 3x4 and 3x2 slots, driver is 580-server (not open), cuda is 12.9, 128gb of ddr4 @ 3600cl18 in dual channel.
   I will appreciate any hints and ready to try searching for troubles with gdb or other utilities if asked what to look for and where to look.