Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.

Describe the current behavior / 问题描述 (Mandatory / 必填)

910B2 rms_norm算子精度存在问题。

Environment / 环境信息 (Mandatory / 必填)

芯片型号：Ascend910B2
CANN 版本：9.0.0
CANN 版本详细信息:
从 /usr/local/Ascend/cann-9.0.0/opp/version.info 获取：
Version=9.0.0
version_dir=cann
required_package_runtime_version="9.0"
required_package_ge-executor_version="9.0"
required_package_metadef_version="9.0"
required_package_opbase_version="9.0"
required_package_bisheng-compiler_version="9.0"
required_package_tbe-tik_version="9.0"
timestamp=20260428_134817545

Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)

测试脚本：
m = 2**33 // 128
n = 128
x_shape = (m, n)
print("x_shape:", x_shape)
x = torch.randn(x_shape, dtype=torch.bfloat16).npu()
weight = torch.randn((n,), dtype=torch.bfloat16).npu()
y = torch_npu.npu_rms_norm(x, weight, 1e-5)
y = y[0]
pos = 4294969344
print("y[pos - 1]", y.reshape(-1)[pos - 1 :])
print("y[pos]", y.reshape(-1)[pos:])

x_shape: (67108864, 128)
y[pos - 1] tensor([-0.8242, 0.0000, 0.0000, ..., 0.0000, 0.0000, 0.0000],
device='npu:0', dtype=torch.bfloat16)
y[ps] tensor([0., 0., 0., ..., 0., 0., 0.], device='npu:0', dtype=torch.bfloat16)

测试结果：
从位置 4294969344 开始，输出为全0。
预期结果：
所有位置结果正确。

Describe the expected behavior / 预期结果 (Mandatory / 必填)

所有位置结果正确。

Related log / screenshot / 日志 / 截图 (Mandatory / 必填)

y[pos - 1] tensor([-0.8242, 0.0000, 0.0000, ..., 0.0000, 0.0000, 0.0000],
device='npu:0', dtype=torch.bfloat16)
y[pos] tensor([0., 0., 0., ..., 0., 0., 0.], device='npu:0', dtype=torch.bfloat16)

Special notes for this issue/备注 (Optional / 选填)