6585887db30feab4c87f81b6ffedc3ba2dcb13bb,thinc/neural/_classes/multiheaded_attention.py,MultiHeadedAttention,_apply_attn,#MultiHeadedAttention#,398

Before Change

        nD = V0.shape[-1]
        V1 = V0.reshape((nB * nH, nL, nD))
        S1 = S0.reshape((nB * nH, nL, nL))
        S2 = self.ops.xp.matmul(S1, V1)

        S3 = S2.reshape((nB, nH, nL, nD))

        def backprop_attn4(dS3):
            dS2 = dS3.reshape((nB * nH, nL, nD))

After Change

        // S3: (nB, nH, nL, nD)
        nB, nH, nL, nL = S0.shape
        nD = V0.shape[-1]
        V1 = V0.reshape((nB * nH, nL, nD))
        S1 = S0.reshape((nB * nH, nL, nL))
        S2 = self.ops.matmul(self.ops.xp.ascontiguousarray(S1), self.ops.xp.ascontiguousarray(V1))

        S3 = S2.reshape((nB, nH, nL, nD))

        def backprop_attn4(dS3):
            dS2 = self.ops.xp.ascontiguousarray(dS3.reshape((nB * nH, nL, nD)))

Instances