* fix rotating kv cache for chat use case
* reorg + fixes to caching, unify prompt caching across types and use cases for e.g. caching during a chat
* nit in chat
* fix tests
* fix tests
* fix tests
* docs
* chat command
* comments + docs
* Define meta_state on all Cache implementations
* fixes + trim_prompt_cache api
* fix default model
---------
Co-authored-by: Angelos Katharopoulos <a_katharopoulos@apple.com>
* use fast rope
* fix llama
* use fast rope for llama3.1
* requires unreleased mlx
* fix su
* fix deepseek v2
* only one of base or freqs
* nit
* fix
* hard code freqs
* 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
* add dynamicNTK scaling rope
* remove unused var
* fix rope base
* llama3.1 fixes
* TODO for rope eval
* vectorise llama3 base freq calculation
* removed the arbitrary 2.0 rope_scale default case
* fix slow llama3.1 generation by evaluating stateless part of DynamicNTKScalingRoPE in init
* nits + format
* use mx.pi
* fix tests and add test for 3.1
---------
Co-authored-by: Prince Canuma <prince.gdt@gmail.com>
Co-authored-by: Awni Hannun <awni@apple.com>
* add support for granite 3-8B config
* add gpt_bigcode
* add positional embedding condition.
* add support for granite 3-8B config
* add gpt_bigcode
* add positional embedding condition.
* remove unused function
* rebase fix
* move position emebedding to mask creation
* add to tuner and format
* add support for granite 3-8B config
* add gpt_bigcode
* add positional embedding condition.
* add support for granite 3-8B config
* add gpt_bigcode
* add positional embedding condition.
* rebase fix
* move position emebedding to mask creation
* add to tuner and format
* refactor mask
* remove dropout layers
* Pad mask with zeros for non-square attention matrices
The current implementation of the mask assumes the attention matrix is square, which is true if there is no cache. However, if one wishes to produce multiple tokens at a time, such as in speculative decoding implementations, a rectangular mask is necessary.
This change pads the bottom of the mask with zeros so multi-token decoding with a cache works correctly.
* Directly create mask instead of padding
* Update llama.py
* 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/
