Unverified50% confidenceFactExact time
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
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
McNemar检验对比机器学习模型：随机种子下的正确用法
redditr/learnmachinelearning7/11/2026
Related Claims
Unverified实际部署中因并行计算的浮点非确定性可能产生细微差异并随链式生成放大67% similarUnverified即便固定所有随机种子，在不同硬件或并行度下，训练结果仍可能出现被大模型放大为显著能力差异的细微偏差67% similarUnverifiedThe developer systematically ruled out the possibility of evolution exploiting physics engine loopholes, and after increasing the velocity solver iteration count from 8 to 24, the error dropped to ±0.24 meters per seed, but a diagonal gait still did not emerge.66% similarUnverifiedAn intermittently faulty memory module may alternately produce correct and incorrect computation results, causing sporadic deviations in gradient calculations during neural network training, affecting model convergence without triggering hardware alerts.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/503586API
curl https://kongchang.com/api/v1/knowledge/claims/503586MCP
get_claim(id=503586)