return errors::Unimplemented(
        "BroadcastMul is not supported in tracing mode yet.");
  }
  auto imm_ctx = dyn_cast<ImmediateExecutionContext>(ctx);
  AbstractTensorPtr minus_1(imm_ctx->CreateInt32Scalar(-1));
  ImmediateTensorHandlePtr dim(imm_ctx->CreateLocalHandle(minus_1.get()));
  AbstractTensorHandle* expand_dims_outputs;
  TF_RETURN_IF_ERROR(
      ops::ExpandDims(ctx, vec, dim.get(), &expand_dims_outputs, "ExpandDims"));

    msys 'pacman --noconfirm -Syy coreutils curl dash file filesystem findutils'; \
    msys 'pacman --noconfirm -Syy flex gawk gcc-libs grep gzip inetutils info'; \
    msys 'pacman --noconfirm -Syy less lndir mintty ncurses pactoys-git patch'; \
    msys 'pacman --noconfirm -Syy pax-git pkgfile rebase sed tar tftp-hpa time tzcode util-linux which'; \
    $old_path = [Environment]::GetEnvironmentVariable(\"PATH\", \"Machine\"); \

    name = "Neg_Div_Grad_Y";
    TF_RETURN_IF_ERROR(Neg(ctx, upstream_grad, &temp_output,
                           name.c_str()));  // -U
    AbstractTensorHandlePtr MinusU(temp_output);

    name = "Mul_Div_Grad_Y";
    TF_RETURN_IF_ERROR(Mul(ctx, MinusU.get(), Z, &temp_output,
                           name.c_str()));  // -U*Z
    AbstractTensorHandlePtr UZ(temp_output);

    name = "Div_Grad_Y";

        pipeline deterministic. This may lead to increased model loading time,
        because time spent in graph optimizations is now unbounded (was 20
        minutes).
*   Deterministic Op Functionality (enabled by setting `TF_DETERMINISTIC_OPS` to
    `"true"` or `"1"`):
    *   Add a deterministic GPU implementation of
        `tf.nn.softmax_cross_entropy_with_logits`. See PR