Fix bloom KV cache usage in ORTForCausalLM

[fxmarty]

Fixes #1150

@michaelbenayoun @echarlaix Basically not all models share the exact same _reorder_cache and prepare_inputs_for_generation, in particular here bloom. What do you think of this solution? @echarlaix I would guess there is the same bug in optimum-intel.

I think my solution is very ugly (now ORTBloomForCausalLM and ORTModelForCausalLM need to be in the same file forever). An other approach is to move all shared methods to ORTModelDecoder (effectively making it a mixin class) and having ORTBloomForCausalLM not inherit from ORTModelForCausalLM, but it does not solve the issue of "all classes in one file", and more importantly I believe that changing the inheritance is too breaking of a change (i.e. isinstance(ort_bloom_model, ORTModelForCausalLM) not working anymore).

Other solution: have a single prepare_inputs_for_generation, _reorder_cache, and dispatch to the relevant function from a dictionary. This adds dynamism, which I think is better to avoid.

Note: should add tests for num_beams>1 in this PR as well

[regisss]

This solution looks good enough to me 👍

[michaelbenayoun]

Looks good for now. If this conditioning on the model type grows maybe we can find a nicer way of doing it, but right now it seems acceptable to me.

[echarlaix]

Thanks for the fix @fxmarty !

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.