Repository navigation
[Bug] [Relax][TOPI] Max and min reductions replace infinities with finite extrema #20935
Description
Expected behavior
The maximum of a nonempty all-negative-infinity tensor is -inf; the minimum of an all-positive-infinity tensor is +inf. ONNX ReferenceEvaluator agrees.
Actual behavior
On LLVM CPU, ReduceMax([-inf,-inf,-inf]) returns -3.4028234663852886e38 and ReduceMin([inf,inf,inf]) returns +3.4028234663852886e38. Finite-value controls return 7 correctly. This concerns nonempty infinity-only inputs, separately from CUDA NaN propagation in #19518 and ArgMin indices in #20866.
Environment
TVM 0.28.dev0, LLVM 18.1.3, Linux x86_64; Python 3.12.3, NumPy 2.5.4, ONNX 1.22.0. Installed binary build revision is unknown. Current main was inspected statically, not rebuilt.
Steps to reproduce
import numpy as np
import onnx
import tvm
from onnx import TensorProto as T, helper as h
from onnx.reference import ReferenceEvaluator
from tvm import relax
from tvm.relax.frontend.onnx import from_onnx
for operation, extreme in [("ReduceMax", -np.inf), ("ReduceMin", np.inf)]:
model = h.make_model(h.make_graph([
h.make_node(operation, ["x"], ["y"], keepdims=0)], "reduce",
[h.make_tensor_value_info("x", T.FLOAT, [3])],
[h.make_tensor_value_info("y", T.FLOAT, [])]),
opset_imports=[h.make_opsetid("", 13)], ir_version=10)
onnx.checker.check_model(model, full_check=True)
mod = from_onnx(model, keep_params_in_input=False)
vm = relax.VirtualMachine(relax.build(mod, target="llvm"), tvm.cpu())
normal = np.full(3, 7, dtype="float32")
np.testing.assert_array_equal(vm["main"](tvm.runtime.tensor(normal)).numpy(), 7)
x = np.full(3, extreme, dtype="float32")
reference = ReferenceEvaluator(model).run(None, {"x": x})[0]
np.testing.assert_array_equal(reference, np.array(extreme, dtype="float32"))
actual = vm["main"](tvm.runtime.tensor(x)).numpy()
assert np.isfinite(actual) and abs(actual) == np.finfo("float32").max
print(operation, "finite control passed; expected", extreme, "actual", actual.item())Triage
- needs-triage
Activity
GoyalIshaan commentedon Nov 9, 2026
Could I work on the max/min reduction identities here, with LLVM CPU regression tests for all-infinite inputs and finite controls?
This request was generated and posted by an AI assistant on my behalf; I would do the implementation and submit the PR myself.