mindspore.ops.cdist with broadcastable batch dimensions throws Framework Unexpected Exception on CPU

[tingPetty]

mindspore.ops.cdist fails when input batch dimensions are broadcastable but not exactly equal (example: (2, 5, 8) vs (1, 7, 8)). The runtime reports a Framework Unexpected Exception and kernel resize failure on CPU. In PyTorch, this case succeeds and returns output shape (2, 5, 7). If broadcasting is unsupported by design, the error should still be a clear user-facing validation message instead of an internal unexpected exception.

Environment

Hardware Environment(Ascend/GPU/CPU):

/device cpu

Software Environment:

• MindSpore version (source or binary): 2.8.0 (binary, pip in conda env)
• Python version (e.g., Python 3.7.5): 3.10.18
• OS platform and distribution (e.g., Linux Ubuntu 16.04): Windows 10 (10.0.26200)
• GCC/Compiler version (if compiled from source): N/A (not compiled from source)

Describe the current behavior

ops.cdist(x1, x2, p=0.0) where:

• x1.shape = (2, 5, 8)
• x2.shape = (1, 7, 8)
  throws RuntimeError with:
• Framework Unexpected Exception Raised
• CPU kernel op [Default/Cdist-op0] resize failed

Describe the expected behavior

Expected behavior should be one of:

1. Support broadcastable batch dimensions and return output (similar to PyTorch behavior), or
2. Raise a stable and explicit user-facing ValueError stating unsupported shape rules.

In either case, internal "Framework Unexpected Exception" should be avoided.

Steps to reproduce the issue

1. Run:

import numpy as np
import torch
import mindspore as ms
from mindspore import ops

print("mindspore:", ms.__version__)
print("torch:", torch.__version__)

x1_ms = ms.Tensor(np.ones((2, 5, 8), dtype=np.float32))
x2_ms = ms.Tensor(np.ones((1, 7, 8), dtype=np.float32))
try:
    y_ms = ops.cdist(x1_ms, x2_ms, p=0.0)
    print("MindSpore success:", y_ms.shape)
except Exception as e:
    print("MindSpore failed:", type(e).__name__)
    print(e)

x1_t = torch.ones((2, 5, 8), dtype=torch.float32)
x2_t = torch.ones((1, 7, 8), dtype=torch.float32)
y_t = torch.cdist(x1_t, x2_t, p=0.0, compute_mode="donot_use_mm_for_euclid_dist")
print("PyTorch success:", tuple(y_t.shape), y_t.dtype)

2. Observe MindSpore reports kernel resize failure and unexpected exception.
3. Observe PyTorch succeeds with shape (2, 5, 7).

Related log / screenshot

Key logs:

[ERROR] ... CdistCpuKernelMod::Resize] invalid input shape, the batch shape of input0 must be the same as the shape of input1
RuntimeError:
- Framework Unexpected Exception Raised:
- Kernel build failed:
CPU kernel op [Default/Cdist-op0] resize failed.
- C++ Call Stack:
mindspore\ccsrc\plugin\cpu\cpu_device_context.cc:547

Special notes for this issue

• This issue is marked as worth reporting mainly because of internal unexpected-exception behavior.