Question about the 'simulated_matmul()' about stonne HOT 7 CLOSED

I did a little test.
There seems to be a problem when a Tensor composed of integers is input.

import torch
import torch_stonne

print("TPU Test\n")
A = torch.randint(0, 10, (4, 1))
B = torch.randint(0, 10, (1, 4))
R = torch.matmul(A,B)

print("A :", A)
print("B :", B)
print("R : ")
print("PyTorch \n ",R)

print("STONNE\n")
r=torch_stonne.simulated_matmul("", A,  B, "../../simulation_files/maeri_128mses_128_bw.cfg", "test_tile.txt", 0)
print(r)

from stonne.

Hello,

Thanks for using STONNE. The purpose of the PyTorch Frontend is to be able to easily obtain the number of cycles required to compute each operation in the architecture. The number of cycles won't vary despite whether the result is correct or not. For that reason, the PyTorch Frontend actually only takes the main parameters from the tensors (dimensions, mainly) and uses them to initialize a random matrix and compute the operation using it. Thus, with the current implementation, it won't provide the correct output tensor for that operation. We are sorry about the inconvenience.

Hope this comment can help you.

from stonne.

So, is it impossible to compute an integer from STONNE?

from stonne.

Well, the only thing is that the PyTorch Frontend does not map directly the values of the input tensors into STONNE. With the current implementation, it's not possible then. We didn't consider this strictly necessary before, but if needed we can open a ticket to try to implement this feature in the future (but we can't promise since we are currently working on other projects)

from stonne.

If possible later, please.

I know this question is rude.
Do you happen to know a simulator that can also simulate the function of NPU (With cycle level simulate)?

from stonne.

Hello,

I am not aware of any other simulator. You could use the workaround of transforming the tensor from integer to float and process it on the simulator. Then you can convert again from float to integer. The simulation cycles will be the same as the size of the tensors will be identical. Does it make sense?

from stonne.

Thank you!

from stonne.