[QST] [CuTeDSL] Branching: ValueError: unable to convert ... in type <class 'list'> to Numeric

[HanGuo97]

What is your question?

Hi, I'm running into the following case, in which reassigning a list in a runtime branching (and only runtime branching) doesn't work.

import cutlass
import cutlass.cute as cute


@cute.jit
def f0():
    xs = [0, 1, 2]
    xs = [2, 3, 4]
    print(xs)


@cute.jit
def f1():
    xs = [0, 1, 2]
    if cutlass.const_expr(xs[0] == 0):
        xs = [2, 3, 4]
    print(xs)


@cute.jit
def f2():
    xs = [0, 1, 2]
    if xs[0] == 0:
        xs = [2, 3, 4]
    print(xs)

f0()
f1()
f2()  # --> ValueError

[fengxie]

@anakinxc @brandon-yujie-sun I think it's expected but I believe we need better error message (actionable) in DSL.

We changed default behavior of if condition to dynamic to make it closer to reason the code is "runtime" rather than "compile-time".

if xs[0] == 0 is considered as runtime code and DSL doesn't support python list at runtime. cutlass.const_expr force it get evaluated at "compile-time".

[anakinxc]

This is a bug actually when handling list/tuple without dynamic values, we will fix this.

As a workaround for now, you can tweak the code to

@cute.jit
def f2():
    xs = [cutlass.Int32(0), cutlass.Int32(1), cutlass.Int32(2)]
    if xs[0] == 0:
        xs = [cutlass.Int32(2), cutlass.Int32(3), cutlass.Int32(4)]
    print(xs)

f2()

[HanGuo97]

Thanks both, this is very useful! The workaround I used before seeing the latest comment was this.

@cute.jit
def f3():
    xs = [0, 1, 2]
    if xs[0] == 0:
        _xs = [2, 3, 4]
        for i in cutlass.range_constexpr(len(xs)):
            xs[i] = _xs[i]
    print(xs)

Would this be a reasonable workaround (without performance impact) too?

[anakinxc]

I think this will work if all elements in _xs are const values, if there is a dynamic value, this will generates invalid ir eventually

[HanGuo97]

Interesting, that's good to know (and also thanks for responding in the other thread too). I will try this out and get back to you.