[georgeguimaraes]

The dispatch cache keys a tensor literal by its shape and type only. Two functions that differ only in the values of a literal get the same key, so the second function runs the program of the first:

a = Nx.tensor([1.0, 2.0, 3.0], backend: Nx.BinaryBackend)
b = Nx.tensor([10.0, 20.0, 30.0], backend: Nx.BinaryBackend)
zero = Nx.tensor([0.0, 0.0, 0.0], backend: EMLX.Backend)

Nx.Defn.jit(fn x -> Nx.add(x, a) end, compiler: EMLX).(zero)
# [1.0, 2.0, 3.0]
Nx.Defn.jit(fn x -> Nx.add(x, b) end, compiler: EMLX).(zero)
# [1.0, 2.0, 3.0], expected [10.0, 20.0, 30.0]

I found this in an ONNX interpreter, where graphs with the same structure have different constant tensors (window functions, resize weights, gather indices). There is no error, only wrong values.

With this change, the key also has an md5 of the literal's data. Template tensors have no data and keep the old key.