* Unify attention mask creation in LLMs.
Currently, each model implementation in `mlx-examples/llms/models` has ad-hoc
code to create a mask for the attention mechanism. This usually takes the form:
```
mask = None
if h.shape[1] > 1:
mask = nn.MultiHeadAttention.create_additive_causal_mask(h.shape[1])
mask = mask.astype(h.dtype)
```
This correctly creates a mask only if the input consists of more than one token.
But this code assumes the multi-token input is at the beginning of inference.
If, for example, we are evaluating multiple tokens because of speculative
decoding or prompt cache reuse, this mask will not have the correct shape and
and will cause the raising of an exception in the attention computation.
Some of the models correctly implement the mask creation with code like this:
```
mask = None
if h.shape[1] > 1:
mask = create_additive_causal_mask(
h.shape[1], cache[0].offset if cache is not None else 0
)
mask = mask.astype(h.dtype)
```
This commit unifies the attention mask creation for all models with a new
function `create_attention_mask`, reducing code duplication and helping all
models support inference performance enhancements like those mentioned above.
* Allow batches in LLM key-value cache
The current implementation of the LLM key-value cache assumes that
the input batch is of size 1. Input batching (evaluating multiple
alterative inputs at the same time) can be a valuable tool for
speculative sampling and other techniques.
This change removes the hard-coded batch size from the code that
resizes the key-value cache.
* Simplify causal mask creation
Use the same codepath regardless of whether there's an offset or
not. Addresses [this comment](https://github.com/ml-explore/mlx-examples/pull/911#discussion_r1691459717).
* Use old-style type annotation to avoid linter error
* Initial config handler and test
* Added means to run from CLI
* Update lora config loading and tests
* Constrain scheduler config (warmup and minimum LR) for each kind
* Update reference to moved schedule_config module
* Minor fix
* Fix typos
* Moved build_schedule and tests
* nits in schedule config
* flake
* fix path
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* use nn.RMSNorm, use sdpa, cleanup
* bump mlx versions
* minor update
* use fast layer norm
* version bump
* update requirement for whisper
* update requirement for gguf
* lazy model import in mlx_lm
* change lora loading
* fix olmo lora
* remove a bunch of unused stuff from plamo
* move phixtral to mlx-lm and out of llms/
* Convert HF weights of PLaMo and load it to a plamo model in mlx
* Fix model inference part
* Add bos at the beginning of the prompt
* Fix convert.py to copy tokenizer.model into the converted dir
* Use the required insturction format in generate.py when "--instruct" option is specified
* Change filenames and update existing scripts
* Add README
* Add requirements.txt
* Fix plamo.py to stop generation when EOS appears
* Add quantization to convert.py
* Use mlx>=0.0.9 for mx.core.outer() in PLaMo model
* Update acknowledgements.md
* Fix card text in upload_to_hub()
* Not use prompt template when --instruct is not specified
* Ask if you trust_remote_code for loading tokenizer of PLaMo
* Check the user trusts the remote code when converting
* Remove plamo directory
* Update README
* Add PLaMo model file
* Fix the handling of cache in PLaMo and update README
* Ask if trust_remote_code only when the model is PLaMo
* Remove resolve_trust_remote_code from convert.py and use the latest transformers
* Remove code not to add EOS
* Update README to fix an example not to use noncommercial version of the model
* Remove unused imports
* Remove unnecessary description about the instruct model of PLaMo from README
* format, nits in README
* typo
---------
Co-authored-by: Shunta Saito <shunta@mitmul-mbp.local>
Co-authored-by: Awni Hannun <awni@apple.com>
