Ubospica opened a new pull request, #14670:
URL: https://github.com/apache/tvm/pull/14670
This PR adds the optimizer library to support the Relax training workflow.
Optimizer library generates relax functions representing common optimizers,
such as SGD and Adam.
Taking SGD as an example, the optimizer function is like:
```
@R.function
def SGD(
params: R.Tuple(R.Tensor((3, 3), "float32"), R.Tensor((3,), "float32")),
gradients: R.Tuple(R.Tensor((3, 3), "float32"), R.Tensor((3,),
"float32")),
optim_states: R.Tuple(R.Tensor((), "int64")),
) -> R.Tuple(
R.Tuple(R.Tensor((3, 3), "float32"), R.Tensor((3,), "float32")),
R.Tuple(R.Tensor((), "int64")),
):
with R.dataflow():
num_steps: R.Tensor((), "int64") = optim_states[0]
num_steps_new: R.Tensor((), "int64") = R.add(num_steps, R.const(1,
"int64"))
x: R.Tensor((3, 3), "float32") = params[0]
x_grad: R.Tensor((3, 3), "float32") = gradients[0]
lv: R.Tensor((3, 3), "float32") = R.multiply(R.const(0.01,
"float32"), x_grad)
x_new: R.Tensor((3, 3), "float32") = R.subtract(x, lv)
y: R.Tensor((3,), "float32") = params[1]
y_grad: R.Tensor((3,), "float32") = gradients[1]
lv1: R.Tensor((3,), "float32") = R.multiply(R.const(0.01,
"float32"), y_grad)
y_new: R.Tensor((3,), "float32") = R.subtract(y, lv1)
params_new: R.Tuple(R.Tensor((3, 3), "float32"), R.Tensor((3,),
"float32")) = (
x_new,
y_new,
)
optim_states_new: R.Tuple(R.Tensor((), "int64")) = (num_steps_new,)
R.output(params_new, optim_states_new)
return (params_new, optim_states_new)
```
Co-authored-by: Chaofan Lin
[[email protected]](mailto:[email protected])
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
