d5604eaba321c11c1b9616c283262c4ddea55049,rllib/models/torch/modules/relative_multi_head_attention.py,RelativeMultiHeadAttention,forward,#RelativeMultiHeadAttention#Any#Any#,70

Before Change

        // Add previous memory chunk (as const, w/o gradient) to input.
        // Tau (number of (prev) time slices in each memory chunk).
        Tau = list(memory.shape)[1] if memory is not None else 0
        if memory is not None:
            memory.requires_grad_(False)
            inputs = torch.cat((memory, inputs), dim=1)

After Change

        // Add previous memory chunk (as const, w/o gradient) to input.
        // Tau (number of (prev) time slices in each memory chunk).
        Tau = list(memory.shape)[1]
        inputs = torch.cat((memory.detach(), inputs), dim=1)

        // Apply the Layer-Norm.
        if self._input_layernorm is not None:
            inputs = self._input_layernorm(inputs)

Instances