Is there a way to downcast a float32 tensor to a float16 tensor?

[balisujohn]

I'm reverse engineering a pytorch network so I want numbers to exactly match. For some reason pytorch conv1d is automatically turning float32 input tensors into a float16 output tensor, I'm doing analogous transformations with the float32 tensors in ggml, but I'd like to immediately cast the resulting fp32 tensor to fp16 afterward to see if I get the numbers to exactly match. Is there a recommended way to do this, or would I need to add it?

[slaren]

You can use ggml_cpy to convert tensors to a different data type, you would need to create a tensor with the same dimensions and a different type and specify it as the destination of the copy. For example:

ggml_cpy(ctx, src, ggml_new_tensor(ctx, GGML_TYPE_F32, 4, src->ne));

There are also functions defined in ggml.h to convert data between fp16 and fp32, ggml_fp16_to_fp32 and ggml_fp16_to_fp32_row.