How to check if docker image hosted on Windows system is actually using the GPU?

[tuelle]

I'm using LocalAI on a system with the RTX4070 GPU with 8GB on a ZBOX barebone. I have configured the docker-compose file to pass through access to the GPU. Also configured use of cuBLAS. But it is actually getting slower than using the CPU setup with 14 Cores.

How can I check if the GPU is actually used by LocalAI? I installed the Nvidia driver and docker-desktop on the host. Do I also have to install other libraries on the host? Do I have to configure the docker service?

When I run docker run --rm --gpus all nvidia/cuda:11.0.3-base-ubuntu20.04 nvidia-smi I get this output:

+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 546.01 Driver Version: 546.01 CUDA Version: 12.3 |
|-----------------------------------------+----------------------+----------------------+
| GPU Name TCC/WDDM | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+======================+======================|
| 0 NVIDIA GeForce RTX 4070 ... WDDM | 00000000:01:00.0 On | N/A |
| N/A 36C P8 6W / 115W | 731MiB / 8188MiB | 63% Default |
| | | N/A |
+-----------------------------------------+----------------------+----------------------+

+---------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=======================================================================================|
| 0 N/A N/A 4884 C+G ...\Docker\frontend\Docker Desktop.exe N/A |
| 0 N/A N/A 6752 C+G C:\Windows\explorer.exe N/A |
| 0 N/A N/A 9972 C+G ...2txyewy\StartMenuExperienceHost.exe N/A |
| 0 N/A N/A 10724 C+G ...CBS_cw5n1h2txyewy\TextInputHost.exe N/A |
| 0 N/A N/A 11608 C+G ....Search_cw5n1h2txyewy\SearchApp.exe N/A |
| 0 N/A N/A 12428 C+G ...oogle\Chrome\Application\chrome.exe N/A |
| 0 N/A N/A 14848 C+G ...crosoft\Edge\Application\msedge.exe N/A |
| 0 N/A N/A 16708 C+G ...5n1h2txyewy\ShellExperienceHost.exe N/A |
+---------------------------------------------------------------------------------------+

So its seems to me that the docker-desktop service on the host ist configured correctly (at least for the standard nvidia image)

[dionysius]

12:32AM DBG GRPC(dolphin-2_6-phi-2.Q4_K_M.gguf-127.0.0.1:44467): stderr llm_load_tensors: using CUDA for GPU acceleration
12:32AM DBG GRPC(dolphin-2_6-phi-2.Q4_K_M.gguf-127.0.0.1:44467): stderr llm_load_tensors: system memory used  =   70.44 MiB
12:32AM DBG GRPC(dolphin-2_6-phi-2.Q4_K_M.gguf-127.0.0.1:44467): stderr llm_load_tensors: VRAM used           = 1634.32 MiB
12:32AM DBG GRPC(dolphin-2_6-phi-2.Q4_K_M.gguf-127.0.0.1:44467): stderr llm_load_tensors: offloading 32 repeating layers to GPU
12:32AM DBG GRPC(dolphin-2_6-phi-2.Q4_K_M.gguf-127.0.0.1:44467): stderr llm_load_tensors: offloading non-repeating layers to GPU
12:32AM DBG GRPC(dolphin-2_6-phi-2.Q4_K_M.gguf-127.0.0.1:44467): stderr llm_load_tensors: offloaded 33/33 layers to GPU

When you see offloaded X/Y layers to GPU for LLMs using llama.cpp. Remember the model config itself has gpu_layers config option that needs to be set. https://localai.io/features/gpu-acceleration/index.html#model-configuration

[atljoseph]

Use the nvtop command