report regression in executorch-0.5.0

[guangy10]

ExecuTorch 0.5.0 was released: https://github.com/pytorch/executorch/releases/tag/v0.5.0. However, this repo is new and there is no cross-repo integration tests, so there may be some regressions in 0.5.0. However, since 0.5.0 is out, there is no way to fix it back in 0.5.0, we can only target to get all known integration issues fixed by the upcoming 0.6.0 in Mar, 2025. The good thing is that we start pinning optimum-executorch to ExecuTorch CI in Run Hugging Face models via Optimum on CI pytorch/executorch#8630

If you run into issue using executorch-0.5.0, you can be unblocked by trying either:

• downgrade to 0.4.0 by running pip install executorch==0.4.0
• install executorch from main branch.

[guangy10]

Notice an issue when running Qwen2.5 with executorch 0.5.0. Works fine with 0.4.0

[tensor_util.h:713] Tensors do not match: numel=(2048,  1), dim=(1, 2)
[tensor_util.h:720]     size(0): (2048, 1)
[op_le.cpp:29] Check failed (tensors_have_same_shape(a, b)):
[method.cpp:1072] KernelCall failed at instruction 0:18 in operator aten::le.Tensor_out: 0x12
[method.cpp:1078] arg 0 with type id 1
[method.cpp:1078] arg 1 with type id 1
[method.cpp:1078] arg 2 with type id 1
[method.cpp:1078] arg 3 with type id 1
Traceback (most recent call last):
  File "/Users/guangyang/optimum-executorch/optimum/executorch/modeling.py", line 360, in text_generation
    generated_tokens = self.generate(
                       ^^^^^^^^^^^^^^
  File "/Users/guangyang/optimum-executorch/optimum/executorch/modeling.py", line 301, in generate
    logits = self.forward(
             ^^^^^^^^^^^^^
  File "/Users/guangyang/optimum-executorch/optimum/executorch/modeling.py", line 114, in forward
    return self.model.forward((input_ids, cache_position))[0]
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: method->execute() failed with error 0x12