I’m working on a project that requires fully deterministic outputs across different machines using Ollama. I’ve ensured the following parameters are identical:

Model quantization (e.g., llama2:7b-q4_0).

Seed and temperature=0.

Ollama version (e.g., v0.1.25).

However, the hardware/software environments differ in:

GPU drivers (e.g., NVIDIA 535 vs. 545).

CPU architecture (e.g., Intel x86 vs. AMD).

OS (e.g., Windows vs. Linux). Question:

Theoretically, should these configurations produce identical outputs, or are there inherent limitations in Ollama (or LLMs generally) that prevent cross-platform determinism?

Are there documented factors (e.g., hardware-specific floating-point precision, driver optimizations, or OS-level threading) that break reproducibility despite identical model settings?

Does Ollama’s documentation or community acknowledge this as a known limitation, and are there workarounds (e.g., CPU-only mode)?

Example code:

import ollama  

response = ollama.generate(  
    model="llama2:7b-q4_0",  
    prompt="Explain quantum entanglement.",  
    options={'temperature': 0, 'seed': 42}  
)  
print(response['response'])  

The Ollama API docs mention seed and temperature but don’t address cross-platform behavior.