Unverified50% confidenceFactExact time
During PyTorch's first inference, JIT compilation of CUDA operators, automatic cuDNN convolution algorithm tuning, and computation graph construction and optimization must be completed—operations that only occur on the first execution.
1
Sources
50%
Confidence
Long-term
Relevance
10/6/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCUDA kernels require time-consuming operations such as JIT compilation and kernel selection upon first execution, causing abnormally high inference latency for the first batch, which can be mitigated by using dummy inputs to trigger initialization in advance.75% similarUnverifiedWhen benchmark mode is enabled, cuDNN tries all available convolution algorithms upon first encountering a specific input configuration and caches the fastest one; certain fast algorithms use non-deterministic accumulation strategies.71% similarUnverifiedPyTorch的张量设计深受NumPy的ndarray影响，但增加了自动微分（autograd）能力和CUDA支持70% similarUnverifiedPyTorch installation packages include the required CUDA runtime libraries, so no separate CUDA installation is needed for training or simple inference; only deployment scenarios such as exporting to TensorRT require it.69% similarUnverifiedPyTorch的算子内核大量以CUDA C++编写，OpenAI的Triton语言编译后端仍以NVIDIA汇编器PTXAS为主要优化目标68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/979850API
curl https://kongchang.com/api/v1/knowledge/claims/979850MCP
get_claim(id=979850)