Open bailuan opened 1 month ago
at the bottom of the ir, there is something like this: {-# dialect_resources: { builtin: { torch_tensor_1_77_torch.int64: "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" } }
I know this means 77 i64 numbers, where it is from when the turbine produce this ir?
I export clip torchIR by turbine, there is a line ir looks like: "%2 = torch.vtensor.literal(dense_resource : tensor<1x77xsi64>) : !torch.vtensor<[1,77],si64>",
I use iree-opt's "torch-to-iree" pass to lower the torchIR to linalgIR.
this line ir looks like:
"%cst_9 = arith.constant dense_resource : tensor<1x77xi64>"
I try to use iree-opt's "demote-i64-to-i32" to lowering the linalgIR from torchIR, but failed. i want to know where the dense_resource come from? Is any way to set the data type to i32, but not i64(i have tried to set up the model's data type to torch.int32, but it didn't work).
Could someone have idea?