Crash when opt_level = 0

Questions
#1

Hi all, I met a segmentation fault after running the following script:

import tvm
from tvm import relay


var_0 = relay.var("var_0", shape=(), dtype="float64")
var_1 = relay.var("var_1", shape=(), dtype="float64")
var_2 = relay.add(var_0, var_1)
tuple = relay.Tuple([var_2])
func = relay.Function([var_0, var_1], tuple)
mod = tvm.IRModule.from_expr(func)

with tvm.transform.PassContext(opt_level=0):
    exe = relay.backend.vm.compile(mod, "llvm")

The crash still exists after changing target from llvm to cuda. And I have run it on macOS M1 and centos with x86 architecture and rtx 3090

#2

Hi @Haoyang I think it is as same as [Bug][VM] Segmentation fault triggered if opt level set to 0 · Issue #10324 · apache/tvm (github.com)

#3

Wow, thanks. I’m not sure assigning opt_level to 0 is legal or not. BTW, has this problem been solved?

#4

Not yet, it is not appropriate to modify the pass level cause it may impact other cases. I reported a bug and am still waiting for suggestions from the community.