[Bug] tvm.tir.schedule.schedule.ScheduleError

[Cookiee235]

The provided TIR module can be successfully compiled using tvm.build with target='llvm'. However, when using ms.tir_integration.tune_tir for Meta-Schedule tuning, the script crashes unexpectedly with a tvm.tir.schedule.ScheduleError.
It seems a bug in the Meta-Schedule component.

Actual behavior

Traceback (most recent call last):
  File "/share_container/LLMFuzz/TirFuzz/bug_tp/topi.max_0_M2.py", line 29, in <module>
    database = ms.tir_integration.tune_tir(mod=mod, target='llvm --num-cores=16', work_dir='./tune_tmp', max_trials_global=1, num_trials_per_iter=1)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/software/tvm-latest/python/tvm/meta_schedule/tir_integration.py", line 146, in tune_tir
    return tune_tasks(
           ^^^^^^^^^^^
  File "/software/tvm-latest/python/tvm/meta_schedule/tune.py", line 122, in tune_tasks
    task_scheduler.tune(
  File "/software/tvm-latest/python/tvm/meta_schedule/task_scheduler/task_scheduler.py", line 132, in tune
    _ffi_api.TaskSchedulerTune(  # type: ignore # pylint: disable=no-member
  File "python/tvm_ffi/cython/function.pxi", line 758, in core.Function.__call__
  File "<unknown>", line 0, in tvm::meta_schedule::GradientBasedNode::Tune(tvm::ffi::Array<tvm::meta_schedule::TuneContext, void>, tvm::ffi::Array<tvm::FloatImm, void>, int, int, int, tvm::meta_schedule::Builder, tvm::meta_schedule::Runner, tvm::ffi::Array<tvm::meta_schedule::MeasureCallback, void>, tvm::ffi::Optional<tvm::meta_schedule::Database, void>, tvm::ffi::Optional<tvm::meta_schedule::CostModel, void>)
  File "<unknown>", line 0, in tvm::meta_schedule::TaskSchedulerNode::Tune(tvm::ffi::Array<tvm::meta_schedule::TuneContext, void>, tvm::ffi::Array<tvm::FloatImm, void>, int, int, int, tvm::meta_schedule::Builder, tvm::meta_schedule::Runner, tvm::ffi::Array<tvm::meta_schedule::MeasureCallback, void>, tvm::ffi::Optional<tvm::meta_schedule::Database, void>, tvm::ffi::Optional<tvm::meta_schedule::CostModel, void>)
  File "<unknown>", line 0, in tvm::meta_schedule::PostOrderApplyNode::GenerateDesignSpace(tvm::IRModule const&)
  File "<unknown>", line 0, in tvm::meta_schedule::AutoInlineNode::Apply(tvm::tir::Schedule const&, tvm::tir::BlockRV const&)
  File "<unknown>", line 0, in tvm::meta_schedule::AutoInlineNode::CheckInline(tvm::tir::Schedule const&, tvm::tir::BlockRV const&)
  File "<unknown>", line 0, in tvm::tir::GetScopeRoot(tvm::tir::ScheduleState const&, tvm::tir::StmtSRef const&, bool) [clone .cold]
  File "<unknown>", line 0, in tvm::tir::ScheduleError::ScheduleError()
tvm.tir.schedule.schedule.ScheduleError

Environment

tvm-0.22.dev0

Steps to reproduce

import tvm
from tvm import meta_schedule as ms


tir_str = """# from tvm.script import ir as I
# from tvm.script import tir as T

@I.ir_module
class Module:
    @T.prim_func
    def main(data: T.Buffer((8, 8), "float32"), data_red: T.Buffer((), "float32")):
        T.func_attr({"target": T.target({"keys": ["cpu"], "kind": "llvm", "mtriple": "x86_64-unknown-linux-gnu", "tag": ""}), "tir.noalias": T.bool(True)})
        with T.block("data_red"):
            T.reads(data[0:8, 0:8])
            T.writes(data_red[()])
            with T.init():
                data_red[()] = T.float32(-3400000.0)
            for i, j in T.grid(8, 8):
                with T.block("update"):
                    T.reads(data_red[()], data[i, j])
                    T.writes(data_red[()])
                    data_red[()] = T.max(data_red[()], data[i, j])
"""

mod = tvm.script.from_source(tir_str)
mod.show()
#with tvm.transform.PassContext(0):
#    tvm.build(mod, target='llvm')
database = ms.tir_integration.tune_tir(mod=mod, target='llvm --num-cores=16', work_dir='./tune_tmp', max_trials_global=1, num_trials_per_iter=1)

Triage

• needs-triage
• tune:meta_schedule