Method _pre_forward_hook

deepspeed/inference/engine.py:230–235

(self, module, *inputs, **kwargs)

                self.module.transformer.__class__._attn_bias = build_mpt_atten_bias_tensor

    def _pre_forward_hook(self, module, *inputs, **kwargs):
        if self.use_cuda_events:
            self.timers(INFERENCE_MODEL_TIMER).start()
        else:
            get_accelerator().synchronize()
            self._start = time.time()

    def _post_forward_hook(self, module, input, output):
        if self.use_cuda_events: