* 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
* feat: deepseek v1
DeepSeek is still releasing models on the DeepSeek V1 architecture.
```sh
mlx_lm.convert --hf-path deepseek-ai/DeepSeek-Prover-V1.5-RL --mlx-path DeepSeek-Prover-V1.5-RL-8bit --q-bits 8 -q
mlx_lm.generate --model DeepSeek-Prover-V1.5-RL-8bit --ignore-chat-template --max-tokens 512 --prompt 'import Mathlib
import Aesop
set_option maxHeartbeats 0
open BigOperators Real Nat Topology Rat
/-- The second and fourth terms of a geometric sequence are $2$ and $6$. Which of the following is a possible first term?
Show that it is $\frac{2\sqrt{3}}{3}$.-/
theorem amc12b_2003_p6 (a r : ℝ) (u : ℕ → ℝ) (h₀ : ∀ k, u k = a * r ^ k) (h₁ : u 1 = 2)
(h₂ : u 3 = 6) : u 0 = 2 / Real.sqrt 3 ∨ u 0 = -(2 / Real.sqrt 3) := by'
```
* nits
* nits
* nits
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* feature: LoRA adapter for Embeddings
* feature: wire in LoRAEmbedding into the tuner. Allow the embedding and non model.layers Linear layers to be targeted for fine tuning
* feature: DoRA adapter for Embeddings
* feature: wire in DoRAEmbedding
* bugfix: ensure self.m is recalculated when the linear layer is changed in DoRALinear.from_linear
* refactor: prefer from_base over from_linear or from_embedding. prefer fuse over to_linear or to_embedding
* cleanup: remove unused imports in test_dora.py
* refactor: remove unnecessary non_layer_modules
* cleanup: remove wrong comments for lora embedding dropout. remove uncessary parens in dora embedding dropout
* nits
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* Predict stop sequence matches during streaming
Check for overlap of stop sequences and the tokens array for potential sequence matches after more tokens get generated. Generate tokens until we can confirm that the stop sequence is not met.
* fix typo
* Change sequence_overlap logic
* range isn't inclusive, add 1 to max_overlap
* Add test_server.py
Added a test for the sequence_overlap method
* nits
* eos sequence
* finalize
---------
Co-authored-by: Y4hL <43219534+Y4hL@users.noreply.github.com>
Co-authored-by: Awni Hannun <awni@apple.com>
* Added functionality to load in adapters through post-requests so you do not need to restart the server
* ran pre-commit
* nits
* fix test
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* 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>
* 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>
* Su-RoPE
* nits
* Update su_rope.py
* Update su_rope.py
Per GPT4: "The error TypeError: 'type' object is not subscriptable is caused by using the type hint list[float] in a version of Python that does not support it. This syntax is only available in Python 3.9 and later."
* Ran isort
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* GPT-2 model support
* Add test for gpt2 model
* Fix weight sanitizing for quantization
* use approx gelu
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* LoRA: Extract pre_processing_model function
* LoRA: Extract small functions(train_model,evaluate_model)
* move test case to test_tuner_utils.py
* nits
* nits
* remove extra param, validate at it 0
* version
* fix test
---------
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
* fix: Added dedicated error handling to load and get_model_path
Added proper error handling to load and get_model_path by adding a dedicated exception class, because when the local path is not right, it still throws the huggingface RepositoryNotFoundError
* fix: Changed error message and resolved lack of import
* fix: Removed redundant try-catch block
* nits in message
* nits in message
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* support dora finetune
* solve problems in lora.py and tuner.utils.py
* add use_dora (bool) in functions of load adapters
* delete all unsupported quantization code and fix all the calculate problems in mlx_lm/tuner/dora.py
* Using stop_gradient to prevent gradients from flowing through ‘norm’ during backpropagation
* set DEFAULT_USE_DORA in mlx_lm/generate.py
* add annotation for all the use_dora
* mlx_lm/fuse.py support fuse dora layers and fix a bug of to_linear() in mlx_lm/tuner/dora.py
* simplify code of juding type of a fused layer in mlx_lm/fuse.py
* add use_dora in mlx_lm/fuse.py when apply_lora_layers()
* style + nits
* style + nits
* more updates
---------
Co-authored-by: chenyifei08 <chenyifei08@baidu.com>
Co-authored-by: Awni Hannun <awni@apple.com>
* Add `model_config` parameter to `load()` and `load_model()`
For easy editing of the loaded model configuration (e.g., for changing RoPE theta or scaling of Phi-3 model)
Example:
```python
from mlx_lm import load, generate
model, tokenizer = load("mlx-community/Phi-3-mini-4k-instruct-4bit-no-q-embed", model_config={"rope_theta":50000.0})
response = generate(model, tokenizer, prompt, max_tokens=MAX_TOKENS)
```
* Possible bug (default_loss)
* Revert "Possible bug (default_loss)"
This reverts commit 70a55ace18.
* Fix default_loss for lora
* 1. move load_model's new optional `model_config` arg to the end (fetch_from_hub()'s `model = load_model(model_path, lazy)`) 2. fix indentations (`black` hook)
* 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
* Add support for setting MLX cache limit in GB
* Add support for setting MLX cache limit in GB in mlx_lm.server
* format
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* Add model management functionality for local caches
This commit introduces a set of command-line utilities for managing MLX models downloaded and saved locally in Hugging Face cache. The functionalities include scanning existing models, retrieving detailed information about a specific model, and deleting a model by its name.
* Added mlx_lm.model to setup.py
* nits
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* Update model card describe
- Add full link jump
- Add the address of the model uploader's Hugging Face homepage
* Add user_info to reduce whoami calls
* Remove the -U argument
* remove HF user info
* run pre-commit
* support for phi-3 4bits quantized gguf weights
* Added link to 4 bits quantized model
* removed some prints
* Added correct comment
* Added correct comment
* removed print
Since last condition already prints warning for when quantization is None
* Added support for the MiniCPM architecture
* Added support for the MiniCPM architecture
* Updated utils.py and LORA.md
* Updated utils.py and LORA.md
* Update implementation details for MiniCPM architecture
* Cleaning up
* fixed the missing lm.head layer problem
* Refactor Model class to dynamically handle tied and untied word embeddings
* Quick update
* added a dynamic rope scaling base calucaltion
* Added support for the MiniCPM architecture
* Added support for the MiniCPM architecture
* Updated utils.py and LORA.md
* Updated utils.py and LORA.md
* Update implementation details for MiniCPM architecture
* Cleaning up
* fixed the missing lm.head layer problem
* Refactor Model class to dynamically handle tied and untied word embeddings
* added a dynamic rope scaling base calucaltion
* quick fix and clean up
* clean up again
* removed the MiniCPMNorm class as its not used
* forgot something, sorry
* format
* version bump
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* 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
* chore(mlx-lm): clean up the top p imp
* chore: clean up
* chore: add test
* chore: address comments
* chore: clean up docs string
* chore: clean up test
* wip
* wip
* feat: convert mlx model to gguf f16
* chore: conver norm layer to float32 to avoid overflow issue
* chore: add support for mixtral
* chore: clean up
* chore: remove unused import statement
* chore: clean up weight name mapping
* version and readme
* actual version bump
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* Add dropout parameter to lora configuration
A dropout parameter has been added to the lora configuration settings in lora_config.yaml. The LoRALinear class in utils.py has been updated to take this new parameter. Additionally, a AttributeError: 'types.SimpleNamespace' object has no attribute 'prompt' related to `args.prompt` has been removed from lora.py.
* Update lora_config.yaml
Set dropout to 0.0 in the sample config file
* format
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* chore(mlx-lm): fix print_trainable_parameters for quant models
* chore: clean up
* refactor: use layer type to check quant bits
* chore: address comment
* Add --lora-all-linear option to apply LoRa to all linear transfer block layers
* Moved to YAML config and added specification of rank & alpha
* nits in conifg, more tests
* nit
* run tests for prs
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* Convert mlx_lm.lora to use YAML configuration
* pre-commit run fixes
* Fix loading of config file
* Remove invalid YAML from doc
* Update command-line options and YAML parameter overriding, per feedback in #503
* Minor wording change
* Positional argument
* Moved config to a (-c/--config) flag
* Removed CLI option defaults (since CLI options take precedence and their defaults are in CONFIG_DEFAULTS)
* pre-commit format updates
* Fix handling of CLI option defaults
* Prevent None values of unspecified CLI options from overwriting values from CONFIG_DEFAULTS
* nits
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* Use named tuple from typing for typehints
* Add type hints
* Simplify expression
* Type hint fix
* Improved do_POST logic
Use a map of endpoints to methods to reduce redundancy in code
* Fix format
* Improve redundancy
Call method dynamically instead of writing out all arguments twice
* Send response instead of returning
* Fix typo
* Revert change
* Make adapter_file as Optional
* Mark formatter as optional
* format
* Create message generator
Store response data that stays static for the duration of the response inside of the object:
system_fingerprint
request_id
object_type
requested_model
Created a message generator, that dynamically creates messages from the metadata stored inside of the object, and the data from the model pipeline
* Remove leftover
* Update parameters to reflect new object structure
No longer pass all arguments between functions, but use the stores values inside of the object
* Parse body before calling request specific methods
* Call super init
* Update server.py
* Fixed outdated documentation parameter name
* Add documentation
* Fix sending headers twice
During testing I found that when using the streaming option, headers have always been sent twice. This should fix that
* Simplify streaming code by using guard clauses
Don't wrap wfile writes in try blocks, the server class has its own try block to prevent crashing
* Bug fix
* Use Content-Length header
Let the completion type specific methods finish sending the headers. This allows us to send the Content-Length header as the model returns a completion.
* Update utils.py
* Add top_p documentation
* Type hint model and tokenizer as required
* Use static system fingerprint
System fingerprint now stays the same across requests
* Make type hint more specific
* Bug Fix
Supplying less than 2 models to merge would raise ValueError and calls len on unbound "models". Should be "model_paths" instead.
Mark upload_repo as optional
* Move more of the shared code into do_POST
Processing stop_id_sequences is done no matter the request endpoint or type, move it into the shared section. handle_ methods now just return the prompt in mx.array form.
* Store stop_id_sequences as lists instead of np
During testing I found that letting the tokenizer return values as python lists and converting them to mlx arrays was around 20% faster than having the tokenizer convert them to np, and from np to mlx. This allows makes it so numpy no longer needs to be imported.
* Update stop_id_sequences docs
* Turn if check to non-inclusive
Only continue if buffer is smaller
* Documentation fix
* Cleared method names
Instead of handle_stream and generate_competion, we should name it handle_completion.
Instead of handle_completions and handle_chat_completions, we should name it handle_text_completions, since both are completions, calling it text completions should make it more descriptive
* Make comment clearer
* fix format
* format
* Add Starcoder2 model and update utils.py
* Refactor model arguments and modules in starcoder2.py
* Refactor FeedForward class to MLP in starcoder2.py
* Fix typo
* pre-commit
* Refactor starcoder2.py: Update model arguments and modules
* Fix LM head and MLP layers
* Rename input layer norm
* Update bias in linear layers
* Refactor token embeddings in Starcoder2Model
* Rename to standard HF attention layer name
* Add LayerNorm
* Add transposed token embeddings (like in Gemma)
* Refactor MLP and TransformerBlock classes
* Add tie_word_embeddings option to ModelArgs and update Model implementation
* Add conditional check for tying word embeddings in Starcoder2Model
* Fix bias in lm_head linear layer
* Remove unused LayerNorm in stablelm
* Update transformers dependency to use GitHub repository
* fix lm head bug, revert transformer req
* Update RoPE initialization in Attention class
---------
Co-authored-by: Awni Hannun <awni@apple.com>
* StableLM now part of Transformers as stablelm rather than stablelm_epoch; changed config to match new changes
* removing old file
* reference new stablelm
* Add metadata when saving safetensors
Add metadata format="pt" for safetensors so that model's are accessible to `transformers` users as well.
* save with metadata format mlx
Save the model weights with metadata format of "mlx".
* Updated llms/mlx_lm/generate.py
* Don't serve local directory
BaseHTTPRequestHandler serves the current directory by default. Definitely not intended behaviour. Remove the "do_HEAD" and "do_GET" methods.
* Fix typo in method name
I assume hanlde_stream was intended to be called handle_stream
* Fix outdated typehint
load_model returns nn.Module, however fetch_from_hub was not updated to reflect the change
* Add some more type hints
* Add warnings for using in prod
Add a warning to README and runtime, discouraging use in production. The warning is the same as on the python docs for HTTPServer https://docs.python.org/3/library/http.server.html
* format
* nits
---------
Co-authored-by: Awni Hannun <awni@apple.com>
