spikingjelly.activation_based.precision package

precision prepares ordinary PyTorch models before optimizer construction. Model precision and Triton-neuron storage/compute precision are independently configured by PrecisionConfig. MCore language-model precision remains owned by distributed.llm and its native MCore configuration.

Tutorials: 中文 | English.

Transformer Engine delayed-scaling checkpoints may contain pickled extra state. TE 2.18 and later require NVTE_ALLOW_UNSAFE_PICKLE_EXTRA_STATE=1 to load that metadata. Enable it only for checkpoints from a trusted source; SpikingJelly does not set it automatically.