bug: nitro cuda windows low performance on machine has multiple GPUs - tested using Jan App

[hiento09]

Describe the bug
My windows machine has 3 GPUs, when I enabled all 3 GPUs, the token speed was slow (6-9/s) and it even not able to load tinyllama 1B. When I disabled 2 GPUs, 1 active only, the performance was back to normal

Screenshots

• 3 GPUs active

  • Low performance
    
  • Load tinyllama error
    

• 1 GPU active only, then the performance was back to normal
  

Desktop (please complete the following information):

• OS: Windows 11
• Nvidia driver: 531.18
• cuda version: 12.3
• Nitro version: 0.1.27
• GPU:
• 1 RTX 4070ti
• 2 RTX 1660ti

[KossBoii]

@hiento09 I have a feeling that this problem coming from the communication between different GPUs. I'll look out for this while reading the codebase right now.

[hiro-v]

@KossBoii that's the exact problem of multiple GPU problem.
I tested again on that machine:

• If using only 4070ti => 55tok/sec
• If using either 1 out of 2 2 1660ti => 28tok/sec

The distributed inference requires:

• Good bandwidth between GPUs
• The discrepancies between multiple GPUs should be not too much (e.g in this case 4070ti have to wait for 1660ti to compute). And also this case uses PCIe3 and 4, not NVlink => The data have to transmitted via CPU to get to another GPU.
• Explicitly set the value for TP (tensor parallel) in nitro.

It depends but I think the option to use 1 model on a single GPU with the help of CUDA_VISIBLE_DEVICES makes sense in this case (i.e hardware sensing feature)

[hiro-v]

This should be properly supported with this instead: ggml-org/llama.cpp#6017