[Bug]: Interpolate with cubic mode produces NaN when input spatial size is 1x1 on GPU devices #33447
Description
OpenVINO Version
2026.0.0
Operating System
Ubuntu 20.04 (LTS)
Device used for inference
GPU
Framework
None
Model used
No response
Issue description
When using Interpolate operator with mode=cubic on GPU, the output contains NaN values if the input spatial dimensions are 1x1. CPU produces correct results for the same input.
Operator attributes:
mode: cubic
shape_calculation_mode: sizes
coordinate_transformation_mode: half_pixel
nearest_mode: floor
antialias: False
pads_begin: [0, 0, 0, 0]
pads_end: [0, 0, 0, 0]
cube_coeff: -0.75
Input/Output shapes:
Input: [1, 1, 1, 1] (float32)
Output: [1, 1, 2, 2] (float32)
Results:
CPU output: [25.378757, 25.378757, 25.378757, 25.378757]
GPU output: [nan, nan, nan, nan]
Step-by-step reproduction
Download the attached files: resize_bug.xml, resize_bug.bin, resize_bug_inputs.npz
Run the reproducer script
import numpy as np
import openvino as ov
def main():
print("=" * 70)
print("OpenVINO GPU Bug: Interpolate (cubic) NaN on 1x1 input")
print("=" * 70)
core = ov.Core()
print(f"OpenVINO version: {ov.__version__}")
print(f"Available devices: {core.available_devices}")
model = core.read_model("resize_bug.xml")
print("\nModel loaded: resize_bug.xml")
print("\nModel inputs:")
for inp in model.inputs:
print(f" {inp.get_any_name()}: {inp.get_shape()}, {inp.get_element_type()}")
print("Model outputs:")
for out in model.outputs:
print(f" {out.get_any_name()}: {out.get_shape()}, {out.get_element_type()}")
inputs_npz = np.load("resize_bug_inputs.npz")
inputs = {name: inputs_npz[name] for name in inputs_npz.files}
print("\nInputs loaded: resize_bug_inputs.npz")
for name, arr in inputs.items():
print(f" {name}: shape={arr.shape}, dtype={arr.dtype}, values={arr.flatten()[:3]}")
print("\n" + "-" * 70)
print("CPU Inference:")
cpu_compiled = core.compile_model(model, "CPU")
cpu_result = np.array(cpu_compiled(inputs)[0], copy=True)
print(f" Output: {cpu_result.flatten()}")
print(f" Has NaN: {np.any(np.isnan(cpu_result))}")
print("\n" + "-" * 70)
print("GPU Inference:")
gpu_compiled = core.compile_model(model, "GPU")
gpu_result = np.array(gpu_compiled(inputs)[0], copy=True)
print(f" Output: {gpu_result.flatten()}")
print(f" Has NaN: {np.any(np.isnan(gpu_result))}")
print("\n" + "=" * 70)
if np.any(np.isnan(gpu_result)) and not np.any(np.isnan(cpu_result)):
print("❌ BUG CONFIRMED: GPU produces NaN, CPU is correct")
elif not np.allclose(cpu_result, gpu_result, rtol=1e-3, atol=1e-5, equal_nan=True):
print(f"⚠️ Mismatch: max diff = {np.nanmax(np.abs(cpu_result - gpu_result))}")
else:
print("✓ No bug detected")
if __name__ == "__main__":
main()
Relevant log output
======================================================================
OpenVINO GPU Bug: Interpolate (cubic) NaN on 1x1 input
======================================================================
OpenVINO version: 2026.0.0-20742-6ea81c056f0
Available devices: ['CPU', 'GPU']
Model loaded: resize_bug.xml
Model inputs:
v23_0: [1,1,1,1], <Type: 'float32'>
v45_0: [2,1], <Type: 'float32'>
v32_0: [], <Type: 'float32'>
v12_0: [1], <Type: 'float32'>
v3_0: [], <Type: 'char'>
Model outputs:
/Resize_output_0: [1,1,2,2], <Type: 'float32'>
Inputs loaded: resize_bug_inputs.npz
v23_0: shape=(1, 1, 1, 1), dtype=float32, values=[4.6069603]
v45_0: shape=(2, 1), dtype=float32, values=[3.4742732 6.3095818]
v32_0: shape=(), dtype=float32, values=[4.5283375]
v12_0: shape=(1,), dtype=float32, values=[5.6419754]
v3_0: shape=(), dtype=bool, values=[ True]
----------------------------------------------------------------------
CPU Inference:
Output: [25.38357 25.38357 25.38357 25.38357]
Has NaN: False
----------------------------------------------------------------------
GPU Inference:
Output: [nan nan nan nan]
Has NaN: True
======================================================================
❌ BUG CONFIRMED: GPU produces NaN, CPU is correctIssue 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.