* 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>
* CLI arguments may set num_batches to -1
The CLI arguments allow you to validate with the entire dataset by passing a negative one value, but this quickly results in a division by zero `NaN` to appear as the validation score!
* Must properly assemble the mini batches when validating with entire dataset.
Tested locally, a validation of a novel took about an hour, with a loss of 0.928. Thanks @awni for the correction!
* Set up the pre-commit hooks and run them so that black may format lora.py.
* Generate response with optional arguments
* Reference response generation example
* Include transformers and sentencepiece
* Update example to run Mistral-7B-Instruct-v0.3
* Link to generation example
* Style changes from pre-commit
* Add logit soft capping to gemma, and fix precision issues
Gemma was babbling nonsense - so I figured out it was due to not having logit softcapping and precision issues causing NaNs (so I implemented the softcapping and added more float32 inference). gemma-27b-it-4bit now works flawlessly (or near-flawlessly, no sliding-window attention).
* get rid of comments
* get rid of last comments (sry lol)
* nits
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* Add hf_dataset configuration for using HF hub-hosted datasets for (Q)LoRA training
* Pre-commit formatting
* Fix YAML config example
* Print DS info
* Include name
* Add hf_dataset parameter default
* Remove TextHFDataset and CompletionsHFDataset and use Dataset and CompletionsDataset instead, adding a text_key constructor argument to the former (and changing it to work with a provided data structure instead of just from a JSON file), and prompt_key and completion_key arguments to the latter with defaults for backwards compatibility.
* nits
* update docs
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* Initial implementation
* Fix handling of return_step_logits in return
* Fixed OpenAI parameter expectations and logprob structure and datatypes
* pre-commit black formatting
* Remove unused parameter
* fix log probs
* fix colorize
* nits in server
* nits in server
* Fix top_logprobs structure (a dict) and include tokens in logprobs response
* nits
* fix types
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* Tweaks to run dspy-produced calls to the server, with gemma template.
following comment https://github.com/stanfordnlp/dspy/issues/385#issuecomment-1998939936
can try it out with:
```sh
python -m server --model mlx-community/gemma-1.1-7b-it-4bit --port 1143
```
modulo patching the relative imports in server.py
```
-from .tokenizer_utils import TokenizerWrapper
-from .utils import generate_step, load
+from mlx_lm.tokenizer_utils import TokenizerWrapper
+from mlx_lm.utils import generate_step, load
```
and then, ont the dspy side:
```python
import dspy
lm = dspy.OpenAI(model_type="chat", api_base="http://localhost:11434/v1/", api_key="not_needed", max_tokens=250)
lm("hello")
```
* simpler way to validate float or int
* remove logic that works around incompatible templates, too gemma specific
* tweak messages for common denominator
* use generate.py workaround for DBXR
* put behind flag
* oops
* Solution to chat template issue: pass in a custom template!
The template should likely adhere to the OpenAI chat model.
Here is such a template for Gemma.
--chat-template "{{ bos_token }}{% set extra_system = '' %}{% for message in messages %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{% if role == 'system' %}{% set extra_system = extra_system + message['content'] %}{% else %}{% if role == 'user' and extra_system %}{% set message_system = 'System: ' + extra_system %}{% else %}{% set message_system = '' %}{% endif %}{{ '<start_of_turn>' + role + '\n' + message_system + message['content'] | trim + '<end_of_turn>\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\n'}}{% endif %}"
* remove convoluted solution
* Tweak for when None is provided explicitly, and must be set to [] too.
For example, the outlines library provides None explicitly.
* style
---------
Co-authored-by: Awni Hannun <awni@apple.com>
