Summary

torch.clamp / torch.clip on float16 tensors had inconsistent validation
between CPU and GPU. When a scalar bound exceeds the float16 range (~65504),
CPU correctly raises RuntimeError but GPU silently succeeded and returned
incorrect results — the out-of-range bound saturates to inf when stored back
as float16, giving wrong clamp behavior with no warning.

x = torch.zeros(1, dtype=torch.float16)
torch.clip(x, max=65507.0)          # CPU: RuntimeError ✓
torch.clip(x.cuda(), max=65507.0)   # GPU: silent wrong result ✗

Root Cause

CPU kernels (cpu/TensorCompareKernel.cpp) convert the scalar bound via
.to<scalar_t>() where scalar_t = at::Half, which calls c10::check_overflow
and raises. CUDA kernels (cuda/TensorCompare.cu) first promote the scalar to
opmath_t (float for float16) before using it, so a value like 65507.0
fits in float without triggering any overflow check — the error is silently
bypassed.

Fix

Add the overflow check in TORCH_META_FUNC(clamp) in TensorCompare.cpp.
The meta function runs before kernel dispatch on all devices, making the
check device-agnostic in a single change. The existing c10::Scalar::to<T>()
mechanism is reused so the error message is identical to what CPU already
produced — no new error strings introduced.

isReducedFloatingType + AT_DISPATCH_REDUCED_FLOATING_TYPES ensures the
check covers float16, bfloat16, and all float8 variants. bfloat16 with
max=65507 correctly does not raise since 65507 is representable in
bfloat16 (same exponent range as float32).

Prior Attempts and Related Issues

• [CUDA] Add float16 overflow check in clamp_scalar to match CPU behavior #173776 (closed without merge): Fixed only the CUDA kernel path using a
  custom c10::overflows check. Left XPU, ROCm, and any future backends
  unpatched, and produced mangled C++ type names in the error message.
• [ai_generated] torch.clip with float16 overflow scalar does not raise RuntimeError on XPU intel/torch-xpu-ops#3425: Intel XPU team filed the same bug for their
  backend, confirming that a CUDA-only fix is insufficient. Our meta-function
  approach fixes XPU automatically.

Fixes #171356

Checklist

• Passes lint (lintrunner aten/src/ATen/native/TensorCompare.cpp test/test_shape_ops.py)
• Added/updated tests
• Updated documentation (if applicable)

BC-breaking?

No — this turns a silent wrong result into a RuntimeError consistent with
what CPU already raised. No previously correct code is broken.

Developed with AI assistance (Claude). Reviewed and verified by the author.