Worked on the huggingface/diffusers repository, focusing on improving the reliability of stochastic sampling within the FlowMatchEulerDiscreteScheduler. Addressed a bug by ensuring that a random noise generator is always used during sampling, which enhanced both fidelity and reproducibility in training and inference workflows. This fix resolved nondeterministic behavior and brought the diffusion process in line with expected real-world dynamics. The work involved deep understanding of Python, data science, and machine learning principles, applying them to refine core scheduler logic. The contribution prioritized correctness and reliability, resulting in more consistent outcomes for users leveraging diffusion models in this library.