Disable SDPA Attention for gpt-bigcode model

[jiminha]

What does this PR do?

Disable SDPA attention until gpt-bigcode model support FusedSDPA attention.

[vivekgoe]

@jiminha what issue is this PR fixing?

[jiminha]

@jiminha what issue is this PR fixing?
This is fixing the issue SW-176426(TypeError: GPTBigCodeSdpaAttention.forward() got an unexpected keyword argument 'token_idx'). Transformer4.37.2/PT2.2 upgrade enabled sdpd attention for the gpt-bigcode model. This is temporary fix so the model can run with original attention layer without an error until FusedSDPA is enabled for this model.

[vivekgoe]

@jiminha It is ok to disable SDPA attention if it is causing problems. But I see a problem with using "_use_sdpa" flag here

optimum-habana-fork/optimum/habana/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py

Line 238 in 04b0709

if self._use_sdpa and head_mask is None and not output_attentions:

This creates confusion because "_use_sdpa" is used in transformers to refer to torch SDPA which we should differentiate from habana SDPA. We already fixed 1 problem related to this for Llama in this PR #73.

[vivekgoe]

Discussed with @libinta LGTM

[vivekgoe]

Looks good to me. @libinta please go ahead and merge.

[astachowiczhabana]

huggingface#771