[Bug]: float64 constant output returns incorrect value on GPU devices

[zhihaoxu1325]

OpenVINO Version

2025.4.0

Operating System

Ubuntu 20.04 (LTS)

Device used for inference

GPU

Framework

None

Model used

No response

Issue description

When a model has a constant output of float64 (double) type, the GPU plugin returns incorrect values instead of the expected constant values. The bug affects only float64 type; float32 works correctly.

Expected Behavior
Constant outputs should return the same values on both CPU and GPU devices.

Actual Behavior

GPU returns incorrect values for float64 constant outputs:

• Single values often become 0
• Arrays contain corrupted/garbage values

Constant: [1.0]       -> CPU: [1.], GPU: [0.]
Constant: [3.369349]  -> CPU: [3.369349], GPU: [2.]
Constant: [1. 2. 3.]  -> CPU: [1. 2. 3.], GPU: [0. 1.875 0.]

The bug appears to be related to how the GPU plugin handles constant folding or memory allocation for float64 outputs. The corrupted values suggest a possible issue with memory layout or buffer size calculation for 64-bit floating point data.

Root Cause Hypothesis
The GPU plugin may be:

1. Allocating buffer with incorrect size (treating f64 as f32)
2. Reading from uninitialized memory
3. Incorrectly handling constant nodes in the output during graph compilation

Step-by-step reproduction

Run the minimal reproduction scripts:

#!/usr/bin/env python3
"""
Bug: [GPU] float64 constant output returns incorrect value (zero)

When a model has a constant output of float64 type, GPU returns 0 
instead of the correct constant value.
"""

import numpy as np
import openvino as ov
import openvino.runtime.opset13 as ops
from openvino.runtime import Model

print("=" * 70)
print("[GPU] float64 constant output returns zero")
print("=" * 70)

core = ov.Core()
print(f"OpenVINO version: {ov.__version__}")

# Minimal reproduction
const_value = np.array([3.14159], dtype=np.float64)

const_out = ops.constant(const_value)
param = ops.parameter([1], np.float64)
param_out = ops.add(param, ops.constant(np.array([0.0], dtype=np.float64)))

# Model with two outputs: one constant, one parameter-dependent
model = Model([const_out, param_out], [param])

input_data = np.array([1.0], dtype=np.float64)

cpu_result = np.array(core.compile_model(model, "CPU")([input_data])[0])
gpu_result = np.array(core.compile_model(model, "GPU")([input_data])[0])

print(f"\nConstant value: {const_value}")
print(f"CPU output: {cpu_result}")
print(f"GPU output: {gpu_result}")

if np.allclose(cpu_result, gpu_result):
    print("\nStatus: OK")
else:
    print("\n*** BUG CONFIRMED ***")
    print(f"  Expected: {const_value}")
    print(f"  Got (GPU): {gpu_result}")
    print(f"  GPU returns 0 for float64 constant output")

# Additional test: verify float32 works correctly
print("\n" + "-" * 70)
print("Comparison with float32 (works correctly):")
print("-" * 70)

const_value_f32 = np.array([3.14159], dtype=np.float32)
const_out_f32 = ops.constant(const_value_f32)
param_f32 = ops.parameter([1], np.float32)
param_out_f32 = ops.add(param_f32, ops.constant(np.array([0.0], dtype=np.float32)))
model_f32 = Model([const_out_f32, param_out_f32], [param_f32])

input_data_f32 = np.array([1.0], dtype=np.float32)
cpu_result_f32 = np.array(core.compile_model(model_f32, "CPU")([input_data_f32])[0])
gpu_result_f32 = np.array(core.compile_model(model_f32, "GPU")([input_data_f32])[0])

print(f"  float32 - CPU: {cpu_result_f32}, GPU: {gpu_result_f32}")
print(f"  Match: {'OK' if np.allclose(cpu_result_f32, gpu_result_f32) else 'MISMATCH'}")

# Test various float64 values
print("\n" + "-" * 70)
print("Testing various float64 constant values:")
print("-" * 70)

test_values = [
    np.array([0.0], dtype=np.float64),
    np.array([1.0], dtype=np.float64),
    np.array([-1.0], dtype=np.float64),
    np.array([3.369349], dtype=np.float64),
    np.array([1.0, 2.0, 3.0], dtype=np.float64),
]

for val in test_values:
    const_out = ops.constant(val)
    param = ops.parameter([1], np.float64)
    param_out = ops.add(param, ops.constant(np.array([0.0], dtype=np.float64)))
    model = Model([const_out, param_out], [param])
    
    cpu_r = np.array(core.compile_model(model, "CPU")([input_data])[0])
    gpu_r = np.array(core.compile_model(model, "GPU")([input_data])[0])
    
    match = "OK" if np.allclose(cpu_r, gpu_r) else "MISMATCH"
    print(f"  {val} -> CPU: {cpu_r}, GPU: {gpu_r} [{match}]")

Relevant log output

======================================================================
[GPU] float64 constant output returns incorrect value
======================================================================
OpenVINO version: 2025.4.0-20297-5f5229a2009

Constant value: [3.14159]
CPU output: [3.14159012]
GPU output: [0.]

*** BUG CONFIRMED ***
  Expected: [3.14159]
  Got (GPU): [0.]

----------------------------------------------------------------------
Comparison with float32:
----------------------------------------------------------------------
  CPU: [3.14159], GPU: [3.14159]
  Match: OK

----------------------------------------------------------------------
Various float64 constant values:
----------------------------------------------------------------------
  [0.] -> CPU: [0.], GPU: [0.] [OK]
  [1.] -> CPU: [1.], GPU: [0.] [MISMATCH]
  [-1.] -> CPU: [-1.], GPU: [0.] [MISMATCH]
  [3.369349] -> CPU: [3.369349], GPU: [2.] [MISMATCH]
  [1. 2. 3.] -> CPU: [1. 2. 3.], GPU: [0. 1.875 0.] [MISMATCH]

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.

[Wan-Intel]

We will investigate the issue and provide an update shortly.

[Wan-Intel]

Hi @zhihaoxu1325
I've validated your code snippet with OpenVINO nightly version, and the output returns correct value for both CPU and GPU plugin.

You may install the nightly version via the following command:
pip install --pre openvino openvino-tokenizers openvino-genai --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly

[zhihaoxu1325]

Hi @zhihaoxu1325 I've validated your code snippet with OpenVINO nightly version, and the output returns correct value for both CPU and GPU plugin.

You may install the nightly version via the following command: `pip install --pre openvino openvino-tokenizers openvino-genai --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly`

Hi @Wan-Intel ,

Thanks for testing! However, I notice there are still mismatches in your results:

1. float32 mismatch:

float32 - CPU: [3.14159], GPU: [3.140625]

This looks like GPU is using fp16 precision instead of fp32?

2. float64 array mismatch:

[1. 2. 3.] -> CPU: [1. 2. 3.], GPU: [0. 1.875 0.] [MISMATCH]

This is also show Mismatch.

Could you please clarify:

• Are these mismatches expected behavior?
• The main float64 scalar case shows OK, but the array case still fails.

Thanks!

[Wan-Intel]

Hi @zhihaoxu1325

We'll look into this and share an update soon.

[Wan-Intel]

Hi @zhihaoxu1325

I'll escalate this case to relevant team to further investigate the issue as the issue still persists while using OpenVINO nightly. We will provide an update shortly.

[Wan-Intel]

Hi @zhihaoxu1325

GPU plugin does not support f64. It supports the following data types as inference precision of internal primitives:

• Floating-point data types:

  • f32
  • f16

• Quantized data types:

  • u8
  • i8
  • u1

For more information, please refer to Supported Inference Data Types in GPU Device.

[zhihaoxu1325]

Hi @zhihaoxu1325

GPU plugin does not support f64. It supports the following data types as inference precision of internal primitives:

• Floating-point data types:

  • f32
  • f16

• Quantized data types:

  • u8
  • i8
  • u1

For more information, please refer to Supported Inference Data Types in GPU Device.

Hi @Wan-Intel ,

I notice there are still mismatches in your results not about f64

float32 mismatch:
float32 - CPU: [3.14159], GPU: [3.140625]

This looks like GPU is using fp16 precision instead of fp32?

[Wan-Intel]

Hi @zhihaoxu1325

Since this thread is reporting a bug on float64, you may open a new issue on reporting float32, we will take over the case.