Unverified50% confidenceFactExact time
Even with fixed random seeds, non-deterministic operators on GPUs (such as certain cuDNN convolution algorithms and atomic additions) may still introduce tiny differences, making bitwise reproduction of results impossible.
1
Sources
50%
Confidence
Long-term
Relevance
11/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSetting a random seed can only guarantee consistency in initialization, data order, and explicit random sampling; it cannot eliminate non-determinism caused by floating-point parallel computation during backpropagation.75% similarUnverified即使固定PyTorch或TensorFlow的随机种子，若不设置torch.backends.cudnn.deterministic = True，计算结果仍可能因硬件调度差异无法完全复现74% similarVerifiedEven when Temperature is set to 0 (greedy decoding), the order of parallel floating-point operations on modern GPUs can still lead to minor differences in results.74% similarUnverifiedIn large-scale GPU matrix operations, the associative law of floating-point addition does not strictly hold, and different parallel reduction orders, batch sizes, and memory scheduling can cause minor discrepancies in logits.73% similarUnverifiedNon-determinism refers to the phenomenon where the same input may produce slightly different outputs across different runs, which is especially common in models that use GPU parallel computation, dropout layers, or stochastic sampling strategies.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1125900API
curl https://kongchang.com/api/v1/knowledge/claims/1125900MCP
get_claim(id=1125900)