enable cpu paged cache

[jiqing-feng]

CPU can also use paged cache with eager or sdpa:
python continuous_batching_simple.py --attn sdpa

Without this change, the previous command error would be like:

Error in generation loop: unsupported operand type(s) for -: 'NoneType' and 'int'
Traceback (most recent call last):
  File "/home/jiqing/transformers/src/transformers/generation/continuous_batching/continuous_api.py", line 1017, in _run_generation_loop
    paged_attention_cache = PagedAttentionCache(
                            ^^^^^^^^^^^^^^^^^^^^
  File "/home/jiqing/transformers/src/transformers/generation/continuous_batching/cache.py", line 191, in __init__
    num_blocks, max_batch_tokens = memory_handler.infer_num_blocks_and_max_batch_tokens(
                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/jiqing/transformers/src/transformers/generation/continuous_batching/cache.py", line 481, in infer_num_blocks_and_max_batch_tokens
    num_blocks, max_batch_tokens = self.compute_num_blocks_and_max_batch_tokens(
                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/jiqing/transformers/src/transformers/generation/continuous_batching/cache.py", line 522, in compute_num_blocks_and_max_batch_tokens
    cache_memory = self.get_available_memory(max_memory_percent)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/jiqing/transformers/src/transformers/generation/continuous_batching/cache.py", line 456, in get_available_memory
    available_memory = total - max(allocated, reserved)
                       ~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~
TypeError: unsupported operand type(s) for -: 'NoneType' and 'int'

[remi-or]

Hi @jiqing-feng , thanks for the contribution! Just letting you know that CPU-compatible continuous batching is not a priority right now, so even though this PR is small, it will not be reviewed right away. I am cautious about two things:

1. How device map "auto" behaves and how it affects the model's repartition
2. The lack of tests / benchmarks. We have a small template for continuous batching PRs, as in [CB] Easy optimizations for continuous batching #42839 if you can follow it, that would be great.

Will get to review this as soon as I have the bandwidth, thanks you!

[jiqing-feng]

1. device_map="auto" will assign accelerator like cuda/xpu is exists, otherwise will use cpu. But I reverted this change and only added cpu as an option when cuda is not available, it will keep it as original if cuda exists.
2. OK, I will add tests and benchmarks for it.

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=42869&sha=2a5e94