Slow Inference on LLAMA 3.1 405B using ollama.generate with Large Code Snippets on multi-H100 GPUs

I'm experiencing very slow inference times when using the ollama.generate function on a multiple H100 GPU machine. Specifically, it is taking up to 5 minutes per inference, even though the hardware should be able to handle this much faster. The input is a large code snippet, and I expected inference to take significantly less time.

Setup:

• Model: LLaMA 405B
• Hardware: multiple H100 GPUs
• Library Version: 0.3.3
• CUDA Version: 12.6
• Driver Version: 560.35.03
• Operating System: Ubuntu 2404

Steps to Reproduce:

1. Use the ollama.generate function with a large code snippet as input.
2. Observe inference time (up to 5 minutes per call).

Code Example:

import ollama

response = ollama.generate(
    model="llama-405b",
    prompt="Explain the following code:\n[Insert large code snippet here]"
)

print(response)

Expected Behavior: I expected the inference time to be significantly faster, especially on a machine with multiple H100 GPUs. Ideally, the inference should take seconds, not minutes.

Actual Behavior: The inference is taking up to 5 minutes per call, which seems excessively slow for this hardware setup.

Additional Information:

• GPU Utilization: Memory usage across all GPUs is around 50%, while compute utilization is around 25%, with occasional spikes. This suggests under-utilization of GPU resources.

• Mixed Precision: Not sure if mixed precision or quantization is being used by default. This could help improve the inference time.

• Parallelism: It's unclear how the model is being distributed across GPUs, or if any model parallelism optimizations are being applied.

Questions:

1. Is there any support for batching inputs, using mixed precision (FP16/BF16), or quantization in the Ollama library to speed up inference?
2. Are there any known optimizations for better multi-GPU inference (e.g., reducing communication overhead) when using this library?
3. Are there configuration settings that can help fully utilize the multiple H100 GPUs and reduce inference time for large code snippets?

Start by reproducing the reported latency through the ollama.generate entry point with the large code snippet and the listed multi-H100 setup. Investigate how the Python library exposes batching, precision, quantization, and multi-GPU configuration; no source files or tests are named. Done would require an agreed optimization or configuration change supported by measurements showing improved inference time.