[Bug]: Accuracy mismatch in uint8 arithmetic (Sub/Mul/Add) on CPU: OpenVINO saturates while PyTorch wraps

[xwoven]

OpenVINO Version

2025.4.0-20398-7a975177ff4-releases/2025/4 (Python API)

Operating System

Ubuntu 20.04 (LTS)

Device used for inference

CPU

Framework

None

Model used

No response

Issue description

When compiling and running a minimal graph involving uint8 arithmetic operations (Sub, Mul, Add) on CPU, a significant accuracy mismatch is observed compared to PyTorch.

• PyTorch (Oracle): Performs standard wrapping arithmetic for uint8 (e.g., underflow wraps around).
• OpenVINO (CPU): The results appear to be saturated or incorrectly clamped. The output contains almost exclusively 0s and 255s, whereas the expected output (PyTorch) contains a uniform distribution of values.

This suggests that the CPU plugin might be incorrectly applying saturation logic or handling integer overflow/underflow differently from the ONNX/PyTorch standard for this sequence of operations.

Step-by-step reproduction

1. Install OpenVINO 2025.4.0 Python package.
2. Save the provided reproduction script as repro_bug.py.
3. Run python repro_bug.py.
4. Observe the failure: Mismatched elements: 96 / 96 (100.0%).

import torch
import numpy as np
import openvino.runtime as ov
import onnx
import os

# ================= Configuration =================
# 1. Set Random Seed
SEED = 42
np.random.seed(SEED)
torch.manual_seed(SEED)

# 2. Input Parameters
INPUT_DTYPE = np.uint8    
OPSET_VERSION = 14
ONNX_PATH = "bug_repro.onnx"
DEVICE = "CPU"

# Tolerance for uint8 (Exact Match Required)
RTOL = 0
ATOL = 0

# ================= Model Definition =================
class BugModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.register_buffer("v6_0", torch.randint(0, 255, (16, 6), dtype=torch.uint8))

    def forward(self, v0_0, v3_0):
        # Operations susceptible to overflow/underflow
        v7_0 = torch.sub(v0_0, self.v6_0)
        v5_0 = torch.mul(v3_0, v7_0)
        v2_0 = torch.add(v5_0, v5_0)
        return v2_0

# ================= Reporting Function =================
def print_mismatch_report(actual, desired, rtol, atol, location_name="output"):
    actual = np.array(actual)
    desired = np.array(desired)
    
    diff = np.abs(actual.astype(np.float32) - desired.astype(np.float32))
    with np.errstate(divide='ignore', invalid='ignore'):
        rel_diff = diff / np.abs(desired.astype(np.float32))
        rel_diff[np.isnan(rel_diff)] = 0
        
    tol = atol + rtol * np.abs(desired)
    mismatch_mask = diff > tol
    mismatch_count = np.sum(mismatch_mask)
    
    if mismatch_count == 0:
        print(f"✅ Results match within tolerance (rtol={rtol}, atol={atol}).")
        return

    total_count = desired.size
    mismatch_percent = (mismatch_count / total_count) * 100
    
    print(f"Not equal to tolerance rtol={rtol}, atol={atol}")
    print(f"openvino (cpu opt: True) != torch[cpu] eager at {location_name}")
    print(f"Mismatched elements: {mismatch_count} / {total_count} ({mismatch_percent:.1f}%)")
    print(f"Max absolute difference: {np.max(diff)}")
    print(f"Max relative difference: {np.max(rel_diff)}")
    
    np.set_printoptions(threshold=1000, edgeitems=3, precision=8, suppress=True)
    print(f" x: {actual.__repr__()}")
    print(f" y: {desired.__repr__()}")

# ================= Main Process =================
def run_reproduction():
    print(f"[Info] Dtype: {INPUT_DTYPE}")
    v0_0_np = np.random.randint(0, 255, size=()).astype(INPUT_DTYPE)
    v3_0_np = np.random.randint(0, 255, size=(1,)).astype(INPUT_DTYPE)
    
    v0_0_torch = torch.tensor(v0_0_np)
    v3_0_torch = torch.tensor(v3_0_np)

    model = BugModel()
    model.eval()
    try:
        with torch.no_grad():
            y_torch = model(v0_0_torch, v3_0_torch)
    except Exception as e:
        print(f"❌ PyTorch failed: {e}")
        return

    # Export ONNX
    output_name = "v2_0"
    torch.onnx.export(
        model, 
        (v0_0_torch, v3_0_torch), 
        ONNX_PATH, 
        input_names=['v0_0', 'v3_0'], 
        output_names=[output_name], 
        opset_version=OPSET_VERSION
    )

    # Run OpenVINO
    try:
        core = ov.Core()
        ov_model = core.read_model(ONNX_PATH)
        compiled_model = core.compile_model(ov_model, DEVICE)
        infer_request = compiled_model.create_infer_request()
        
        res = infer_request.infer({
            "v0_0": v0_0_np,
            "v3_0": v3_0_np
        })[compiled_model.output(0)]
        y_ov = res
    except Exception as e:
        print(f"❌ OpenVINO failed: {e}")
        return

    # Compare
    print_mismatch_report(
        actual=y_ov, 
        desired=y_torch.numpy(), 
        rtol=RTOL, 
        atol=ATOL, 
        location_name=output_name
    )
    if os.path.exists(ONNX_PATH):
        os.remove(ONNX_PATH)

if __name__ == "__main__":
    run_reproduction()

Relevant log output

(fuzzer) xwoven@DESKTOP-75U6BRL:~/Rosetta-Fuzzer$ python /home/xwoven/1/test2.py
/home/xwoven/miniconda3/envs/fuzzer/lib/python3.10/site-packages/openvino/runtime/__init__.py:10: DeprecationWarning: The `openvino.runtime` module is deprecated and will be removed in the 2026.0 release. Please replace `openvino.runtime` with `openvino`.
  warnings.warn(
[Info] Dtype: <class 'numpy.uint8'>
Not equal to tolerance rtol=0, atol=0
openvino (cpu opt: True) != torch[cpu] eager at v2_0
Mismatched elements: 96 / 96 (100.0%)
Max absolute difference: 246.0
Max relative difference: 17.214284896850586
 x: array([[  0,   0,   0,   0,   0,   0],
       [255,   0, 255, 255, 255,   0],
       [  0, 255,   0, 255,   0, 255],
       ...
       [255,   0, 255, 255, 255,   0],
       [255, 255,   0, 255,   0, 255],
       [  0, 255,   0, 255, 255, 255]], dtype=uint8)
 y: array([[ 12,  26, 134,  90, 112, 214],
       [192,  76, 198, 128, 130,  60],
       [244,  26, 130, 230, 220, 168],
       ...
       [152, 132,  14,  80,  78, 156],
       [236, 196,  84, 150, 220, 118],
       [ 34, 216, 206,  38,  98, 196]], dtype=uint8)

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.

[aviralgarg05]

I would like to solve this issue, please assign me this issue

[mahimairaja]

.take

[mahimairaja]

I observe the same on openvino 2025.4.0, I am currently working on it

[Wan-Intel]

Hi @xwoven

I've run your code snippet with OpenVINO nightly, and I observed the same issue. I'll escalate this to relevant team, and we will provide an update shortly.

[jgespino]

Ref. 180282