Most floating point implementations have deterministic rounding. The popular LLM inference engine llama.cpp is deterministic when using the same sampler seed, hardware, and cache configuration.