🐛 Describe the bug Hi, I'm trying to support while_loop with DispatchKey.XLA; when I try linear and MNIST with torch, code would be dispatched to DispatchKey.CompositeExplicitAutograd to use pure python while, and finish; my local exampl...
(torch.autograd.Function): @staticmethod def forward(ctx, input, normalized_shape, eps): global fused_layer_norm_cuda if fused_layer_norm_cuda is None: fused_layer_norm_cuda = importlib.import_module("fused_layer_norm_cuda") ctx.normalized_shape = normalized_shape ctx.eps ...
(torch.autograd.Function): @staticmethod def forward(ctx, input, weight, normalized_shape, eps): global fused_layer_norm_cuda if fused_layer_norm_cuda is None: fused_layer_norm_cuda = importlib.import_module("fused_layer_norm_cuda") ctx.normalized_shape = normalized_shape...