Method forward

deepspeed/runtime/engine.py:2675–2714

r"""Execute forward propagation Arguments: *inputs: Variable length input list **kwargs: variable length keyword arguments

(self, *inputs, **kwargs)

    @instrument_w_nvtx
    def forward(self, *inputs, **kwargs):
        r"""Execute forward propagation
        Arguments:
            *inputs: Variable length input list
            **kwargs: variable length keyword arguments
        """
        # Clear the backward seen flag at the start of each forward pass.
        # This is used to track multiple gradient hook phases with reentrant checkpointing.
        if isinstance(self.optimizer, ZeROOptimizer):
            self.optimizer.clear_backward_seen_flag()

        if self.autotuning_profile_model_info():
            ma = get_ma_status()

        if self.is_deepcompile_enabled() and not self.is_deepcompile_active() and not self.is_compiled:
            log_dist_once(
                "DeepCompile is enabled but engine.compile() has not been called; executing without DeepCompile until compile() runs.",
                ranks=[0])

        if self.is_deepcompile_active() and hasattr(self, "launch_compile_passes"):
            # We can't have this in forward prologue as the compiler compiles hooks including the forward prologue.
            self.launch_compile_passes(self.global_steps)

        with deepcompile_z3_forward_context(self), autocast_if_enabled(self):
            loss = self.module(*inputs, **kwargs)

        # Register output backward hooks
        # preprocess_once_fn is called for preprocessing
        # preprocess_per_tensor_fn scales a tensor for gradient accumulation
        register_output_backward_hooks(loss,
                                       preprocess_once_fn=self._backward_prologue,
                                       preprocess_per_tensor_fn=self._backward_prologue_per_tensor)

        if self.autotuning_profile_model_info():
            activation_mem = get_ma_status() - ma
            self.autotuning_model_info["activation_mem_per_gpu"] = activation_mem
            print_json_dist(self.autotuning_model_info, [0], path=self.autotuning_model_info_path())
            exit()

        return loss

    def _cast_inputs_half(self, inputs):
        if isinstance(inputs, (list, tuple)):