[Bug]: Float64 Infinity values are clamped to MAX_FLOAT on CPU backend #34138
Description
OpenVINO Version
2025.4.1
Operating System
Windows System
Device used for inference
CPU
Framework
None
Model used
No response
Issue description
I noticed an issue where float64 infinity values (np.inf, -np.inf) are not preserved when round-tripped through OpenVINO tensors on the CPU backend. Instead of remaining as infinity, they are clamped to the maximum finite representation of float64 (approx 1.797e+308). This differs from float32 behavior, where infinity values are correctly preserved.
Similar issue has been reported before in #30264 , but that issue was closed.
Step-by-step reproduction
import numpy as np
import openvino as ov
import openvino.opset15 as ov_opset
def test_inf_preservation():
# Test float32 (Working)
x_f32 = np.array([np.inf, -np.inf, 1.0], dtype=np.float32)
const_f32 = ov_opset.constant(x_f32, dtype=ov.Type.f32)
# Compile and run a simple identity model
model_f32 = ov.Model([const_f32], [])
compiled_f32 = ov.compile_model(model_f32, "CPU")
res_f32 = compiled_f32({})[0]
print(f"FP32 Input: {x_f32}")
print(f"FP32 Output: {res_f32}")
# Test float64 (Failing)
x_f64 = np.array([np.inf, -np.inf, 1.0], dtype=np.float64)
const_f64 = ov_opset.constant(x_f64, dtype=ov.Type.f64)
# Compile and run a simple identity model
model_f64 = ov.Model([const_f64], [])
compiled_f64 = ov.compile_model(model_f64, "CPU")
res_f64 = compiled_f64({})[0]
print(f"FP64 Input: {x_f64}")
print(f"FP64 Output: {res_f64}")
# Check for failure
assert np.isinf(res_f64[0]), "FP64 +inf was lost!"
assert np.isinf(res_f64[1]), "FP64 -inf was lost!"
if __name__ == "__main__":
try:
test_inf_preservation()
print("Test PASSED (Infinities preserved)")
except AssertionError as e:
print(f"Test FAILED as expected: {e}")
Relevant log output
FP32 Input: [ inf -inf 1.]
FP32 Output: [ inf -inf 1.]
FP64 Input: [ inf -inf 1.]
FP64 Output: [ 3.40282347e+38 -3.40282347e+38 1.00000000e+00]
Test FAILED as expected: FP64 +inf was lost!Issue submission checklist
- I'm reporting an issue. It's not a question.
- I checked the problem with the documentation, FAQ, open issues, Stack Overflow, etc., and have not found a solution.
- There is reproducer code and related data files such as images, videos, models, etc.