Unverified50% confidenceFactExact time
即使温度设为0，分布式推理环境中不同GPU的浮点运算顺序差异和CUDA非确定性算子（如atomicAdd）仍会引入数值差异
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
VerifiedEven 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.84% similarVerifiedEven when the temperature is set to 0, some inference frameworks still have non-determinism in underlying floating-point computations, causing output differences across hardware and across batches.74% similarUnverifiedEven when temperature is set to 0, some models still exhibit slight output fluctuations due to floating-point errors in parallel computation.73% similarUnverified加载MTP层后额外浮点运算会导致数值产生微小变化，可能导致模型走不同推理路径产生不同结果，即使温度设为0也会出现70% similarUnverifiedAt Temperature=0, greedy decoding is used, but due to the non-determinism of GPU floating-point operations (FP16/BF16 precision), completely consistent output still cannot be guaranteed in actual engineering.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/527782API
curl https://kongchang.com/api/v1/knowledge/claims/527782MCP
get_claim(id=527782)