您好，能提供下具体的调用部分代码吗？

b35d2916b73f43efb555ebecf478723d.zip

# 不支持 n_fft 320、640、4096
import torch

try:
    import torch_npu
    from torch_npu.contrib import transfer_to_npu
except ImportError:
    print("⚠️  torch_npu not found, NPU tests will be skipped.")
# Check if CUDA/NPU is available and set device


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
input = torch.randn(480000).to(device)

for fft_size in [320, 480, 640, 960, 1920, 512, 1024, 2048, 4096]:
    try:
        output = torch.stft(input, n_fft=fft_size, hop_length=fft_size//2, window=torch.hann_window(fft_size).to(device), return_complex=True)
        print(f"✅ torch.stft with n_fft={fft_size} - Output size: {output.size()}, device: {output.device}, output dtype: {output.dtype}")
    except RuntimeError as e:
        print(f"❌ torch.stft with n_fft={fft_size} - Error: {e}")

print()
for fft_size in [320, 480, 640, 960, 1920, 512, 1024, 2048, 4096]:
    try:
        spec = torch.stft(input, n_fft=fft_size, hop_length=fft_size//2, window=torch.hann_window(fft_size).to(device), return_complex=True)
        recon = torch.istft(spec, n_fft=fft_size, hop_length=fft_size//2, window=torch.hann_window(fft_size).to(device), length=input.shape[-1])
        print(f"✅ torch.istft with n_fft={fft_size} - Output size: {recon.size()}, device: {recon.device}, output dtype: {recon.dtype}")
    except RuntimeError as e:
        print(f"❌ torch.istft with n_fft={fft_size} - Error: {e}")

""" 测试结果 
✅ torch.stft with n_fft=320 - Output size: torch.Size([161, 1001]), device: npu:0, output dtype: torch.complex64
✅ torch.stft with n_fft=480 - Output size: torch.Size([241, 667]), device: npu:0, output dtype: torch.complex64
✅ torch.stft with n_fft=640 - Output size: torch.Size([321, 501]), device: npu:0, output dtype: torch.complex64
✅ torch.stft with n_fft=960 - Output size: torch.Size([481, 334]), device: npu:0, output dtype: torch.complex64
✅ torch.stft with n_fft=1920 - Output size: torch.Size([961, 167]), device: npu:0, output dtype: torch.complex64
✅ torch.stft with n_fft=512 - Output size: torch.Size([257, 626]), device: npu:0, output dtype: torch.complex64
✅ torch.stft with n_fft=1024 - Output size: torch.Size([513, 313]), device: npu:0, output dtype: torch.complex64
✅ torch.stft with n_fft=2048 - Output size: torch.Size([1025, 157]), device: npu:0, output dtype: torch.complex64
✅ torch.stft with n_fft=4096 - Output size: torch.Size([2049, 79]), device: npu:0, output dtype: torch.complex64

[W324 07:42:50.139033988 GetCANNInfo.cpp:117] Warning: Driver Version: B is invalid or not supported yet. (function operator())
✅ torch.istft with n_fft=320 - Output size: torch.Size([160000]), device: npu:0, output dtype: torch.float32
✅ torch.istft with n_fft=480 - Output size: torch.Size([160000]), device: npu:0, output dtype: torch.float32
✅ torch.istft with n_fft=640 - Output size: torch.Size([160000]), device: npu:0, output dtype: torch.float32
✅ torch.istft with n_fft=960 - Output size: torch.Size([160000]), device: npu:0, output dtype: torch.float32
❌ torch.istft with n_fft=1920 - Error: unfold_backward:third_party/op-plugin/op_plugin/ops/opapi/UnfoldBackwardLernelNpuOpApi.cpp:38 NPU function error: call aclnnUnfoldGrad failed, error code is 561002
[ERROR] 2026-03-24-07:42:50 (PID:6196, Device:0, RankID:0) ERR00100 PTA call acl api failed.
EZ9999: Inner Error!
EZ9999[PID: 6196] 2026-03-24-07:42:50.518.011 (EZ9999):  TasksOnceMaxPerCore 0 must be greater than 0.[FUNC:Tiling4Block][FILE:unfold_grad_tiling.cpp][LINE:300]
        TraceBack (most recent call last):
       Tiling4Block failed.[FUNC:DoTiling][FILE:unfold_grad_tiling.cpp][LINE:374]
       UnfoldGrad do tiling failed, ret is -1.
       Check NnopbaseExecutorDoTiling(executor) failed
       Check NnopbaseExecutorTilingAndUpdateBinInfo(executor) failed
       Check NnopbaseExecutorMatchCache(executor) failed
       Check NnopbaseRunForWorkspace(*executor, workspaceSize) failed

✅ torch.istft with n_fft=512 - Output size: torch.Size([160000]), device: npu:0, output dtype: torch.float32
✅ torch.istft with n_fft=1024 - Output size: torch.Size([160000]), device: npu:0, output dtype: torch.float32
❌ torch.istft with n_fft=2048 - Error: unfold_backward:third_party/op-plugin/op_plugin/ops/opapi/UnfoldBackwardLernelNpuOpApi.cpp:38 NPU function error: call aclnnUnfoldGrad failed, error code is 561002
[ERROR] 2026-03-24-07:42:50 (PID:6196, Device:0, RankID:0) ERR00100 PTA call acl api failed.
EZ9999: Inner Error!
EZ9999[PID: 6196] 2026-03-24-07:42:50.574.437 (EZ9999):  TasksOnceMaxPerCore 0 must be greater than 0.[FUNC:Tiling4Block][FILE:unfold_grad_tiling.cpp][LINE:300]
        TraceBack (most recent call last):
       Tiling4Block failed.[FUNC:DoTiling][FILE:unfold_grad_tiling.cpp][LINE:374]
       UnfoldGrad do tiling failed, ret is -1.
       Check NnopbaseExecutorDoTiling(executor) failed
       Check NnopbaseExecutorTilingAndUpdateBinInfo(executor) failed
       Check NnopbaseExecutorMatchCache(executor) failed
       Check NnopbaseRunForWorkspace(*executor, workspaceSize) failed

❌ torch.istft with n_fft=4096 - Error: unfold_backward:third_party/op-plugin/op_plugin/ops/opapi/UnfoldBackwardLernelNpuOpApi.cpp:38 NPU function error: call aclnnUnfoldGrad failed, error code is 561002
[ERROR] 2026-03-24-07:42:50 (PID:6196, Device:0, RankID:0) ERR00100 PTA call acl api failed.
EZ9999: Inner Error!
EZ9999[PID: 6196] 2026-03-24-07:42:50.725.019 (EZ9999):  TasksOnceMaxPerCore 0 must be greater than 0.[FUNC:Tiling4Block][FILE:unfold_grad_tiling.cpp][LINE:300]
        TraceBack (most recent call last):
       Tiling4Block failed.[FUNC:DoTiling][FILE:unfold_grad_tiling.cpp][LINE:374]
       UnfoldGrad do tiling failed, ret is -1.
       Check NnopbaseExecutorDoTiling(executor) failed
       Check NnopbaseExecutorTilingAndUpdateBinInfo(executor) failed
       Check NnopbaseExecutorMatchCache(executor) failed
       Check NnopbaseRunForWorkspace(*executor, workspaceSize) failed

"""

您好，从报错信息上分析，是op-plugin在调用UnfoldBackward算子的时候，没有将数据转为连续内存，建议在op-plugin开源仓提相关issue，这个算子调用目前暂不涉及sip仓的算子实现。

op-plugin代码仓链接：https://gitcode.com/Ascend/op-plugin