EasyDeL

EasyDeL 🔮

EasyDeL is an open-source framework designed to enhance and streamline the training process of machine learning models. With a primary focus on Jax/Flax, EasyDeL aims to provide convenient and effective solutions for training Flax/Jax models on TPU/GPU for both serving and training purposes.

Key Features

Trainers: EasyDeL offers a range of trainers, including DPOTrainer, ORPOTrainer, SFTTrainer, and VideoCLM Trainer, tailored for specific training requirements.
Serving and API Engines: EasyDeL provides serving and API engines for efficiently using and serving large language models (LLMs) in JAX, enabling seamless integration into various applications.
Quantization Support: EasyDeL supports quantization methods for all models, allowing for efficient inference and training.
Bit Operation Support: EasyDeL supports 8, 6, and 4-bit operations for inference and training in JAX, optimizing performance and resource utilization.
Diverse Model Support: EasyDeL offers a wide range of models in JAX that have never been implemented before, such as Falcon, Qwen2, Phi2, Mixtral, Qwen2Moe, Cohere, Dbrx, Phi3, and MPT.
FlashAttention Integration: EasyDeL integrates FlashAttention in JAX for GPUs and TPUs, enhancing performance and efficiency.
Automatic LLM Serving: EasyDeL enables automatic serving of LLMs with mid and high-level APIs in both JAX and PyTorch, simplifying deployment and integration.
LLM Training and Fine-tuning: EasyDeL provides LLM trainer and fine-tuner capabilities in JAX, allowing for efficient training and customization of language models.
Video CLM Training and Fine-tuning: EasyDeL supports Video CLM trainer and fine-tuner for models such as Falcon, Qwen2, Phi2, MPT, Mixtral, Grok-1, and Qwen2Moe, enabling advanced video-related applications.
Performance Optimization: EasyDeL provides various features to enhance the training process and optimize performance, such as LoRA (Low-Rank Adaptation of Large Language Models), RingAttention, FlashAttention, BlockWise FFN, and Efficient Attention support (through the FJFormer backbone).
Model Conversion: EasyDeL supports automatic conversion of models from JAX-EasyDeL to PyTorch-HF and vice versa, facilitating seamless integration with different frameworks.

With its comprehensive set of features and tools, EasyDeL aims to streamline and accelerate the training and deployment of machine learning models, particularly in the domain of large language models and video-related applications.

Latest News 🔥

Aya Model is Supported.
Falcon Model is Updated and now Falcon11B is supported wih flash attention support.
pallas_flash is now available for CPU/GPU/TPU with custom pallas kernel.
DeepseekV2 Model is Added (beta mood).
OpenELM Model is Added.
EasyDeL project structure has changed now you have to import EasyDel as easydel.
ORPOTrainer is Added
Phi3 Model bugs are fixed, Arctic Model is added.

[!TIP]

use ed.AttentionModule.test_attentions() to find the best attention mechanism that works for you
import easydel as ed
ed.AttentionModule.test_attentions()

Documentation 💫

[!IMPORTANT] Documents and Examples are ready at Here Please have that in mind that EasyDeL is in the loop of fast-development so we might have API changes.

Hands on Code Kaggle Examples

script for mindset of using EasyDeL CausalLanguageModelTrainer on kaggle, but you can do much more.
script SuperVised Finetuning with EasyDeL.

Serving

APIs are changing …

[!NOTE] you can use EasyDeLServeEngine which is a Serve API Engine for production purpose sice that’s more stable provide versioned API and efficient.

EasyDeLState A Snapshot of Your EasyDeL Model

The EasyDeLState class acts like a comprehensive container that holds all the essential information about your EasyDeL model at a given point in time. Think of it as a snapshot of your model. It includes:

Training Progress:
- step: Tracks the current training step.
Model Itself:
- module: Holds the actual instance of your EasyDeL model.
- module_config: Stores the model’s configuration settings.
- module_config_args: Keeps track of arguments used to create the configuration (useful for reloading).
- apply_fn: References the core function that applies your model to data.
Learned Parameters:
- params: Contains the trained weights and biases of your model.
Optimizer Information:
- tx: Stores the optimizer you’re using to update the model’s parameters (e.g., AdamW).
- opt_state: Keeps track of the optimizer’s internal state (this is important for things like momentum in optimizers).
- tx_init: Remembers the initial settings used to create the optimizer (again, for reloading purposes).
Additional Settings:
- hyperparameters: Provides a flexible place to store other hyperparameters related to your model or training process.

Key Capabilities of EasyDeLState:

Initialization (create): Lets you create a brand new EasyDeLState to start training.
Loading (load, load_state, from_pretrained): Enables you to reload a saved model from a checkpoint file or even a pre-trained model from a repository like Hugging Face Hub.
Saving (save_state): Allows you to save your model’s current state, including its parameters and optimizer state.
Optimizer Management (apply_gradients, free_opt_state, init_opt_state): Provides methods for updating the model’s parameters using gradients, releasing optimizer memory, and re-initializing the optimizer if needed.
Sharding (shard_params): Helps you distribute your model’s parameters efficiently across multiple devices ( important for training large models).
PyTorch Conversion (to_pytorch): Gives you a way to convert your EasyDeL model to its PyTorch equivalent.

In Essence:

EasyDeLState streamlines the process of managing, saving, loading, and even converting your EasyDeL models. It ensures that you can easily work with your models and maintain consistency throughout your machine learning workflow.

Supervised Fine-Tuning with EasyDeL

EasyDeL supports both DPO and SFT Trainers, so dealing with LLMs in jax is a lot easier right now let have an example of using Supervised Fine-Tuner in JAX with EasyDeL

from easydel import (
    TrainArguments,
    AutoEasyDeLModelForCausalLM,
    EasyDeLOptimizers,
    EasyDeLSchedulers,
    EasyDeLGradientCheckPointers,
    SFTTrainer,
    conversations_formatting_function  # i have added this one for newcomers so if they 
    # don't know what's going on they can use this pre created prompter
)
from datasets import load_dataset
import flax
from jax import numpy as jnp
from transformers import AutoTokenizer

huggingface_repo_id_or_path = "mistralai/Mistral-7B-Instruct-v0.2"

model, params = AutoEasyDeLModelForCausalLM.from_pretrained(huggingface_repo_id_or_path, )

max_length = 4096
tokenizer = AutoTokenizer.from_pretrained(
    huggingface_repo_id_or_path,
    trust_remote_code=True
)
tokenizer.pad_token = tokenizer.eos_token
configs_to_initialize_model_class = {
    "config": model.config,
    "dtype": jnp.bfloat16,
    "param_dtype": jnp.bfloat16,
    "input_shape": (1, 1)
}

train_arguments = TrainArguments(
    model_class=type(model),
    model_name="SFT-EasyDeL",
    num_train_epochs=3,
    configs_to_initialize_model_class=configs_to_initialize_model_class,
    learning_rate=5e-5,
    learning_rate_end=1e-6,
    optimizer=EasyDeLOptimizers.ADAMW,
    scheduler=EasyDeLSchedulers.WARM_UP_COSINE,
    weight_decay=0.01,
    total_batch_size=32,
    max_training_steps=None,  # None to let trainer Decide
    do_train=True,
    do_eval=False,  # it's optional but supported 
    backend="tpu",  # default backed is set to cpu, so you must define you want to use tpu cpu or gpu
    max_length=max_length,  # Note that you have to change this in the model config too
    gradient_checkpointing=EasyDeLGradientCheckPointers.NOTHING_SAVEABLE,
    sharding_array=(1, -1, 1, 1),  # the way to shard model across gpu,cpu or TPUs using sharding array (1, -1, 1, 1)
    # everything training will be in sequence and model parallel automatic and share data between devices
    remove_ckpt_after_load=True,
    gradient_accumulation_steps=8,
    loss_re_mat="",
    dtype=jnp.bfloat16
)


def prompter(sample):
    return [conversations_formatting_function(tokenizer, messages_field="messages")(sample)]


train_dataset = load_dataset("HuggingFaceH4/deita-10k-v0-sft", split="train_sft")
trainer = SFTTrainer(
    arguments=train_arguments,
    train_dataset=train_dataset,
    eval_dataset=None,  # we don't have eval dataset rn :)
    tokenizer=tokenizer,
    dataset_text_field=None,
    formatting_func=prompter,
    packing=True,
    num_of_sequences=max_length,
)

output = trainer.train(flax.core.FrozenDict({"params": params}))
print(f"Hey ! , here's where your model saved {output.checkpoint_path}")

[!NOTE] You Can use Lora too, for DPO, ORPO and SFT Trainers.

FineTuning

with using EasyDeL FineTuning LLM (CausalLanguageModels) are easy as much as possible with using Jax and Flax and having the benefit of TPUs for the best speed here’s a simple code to use in order to finetune your own Model

Days Has Been Passed and now using easydel in Jax is way more similar to HF/PyTorch Style now it’s time to finetune our model

from easydel import (
    TrainArguments,
    CausalLanguageModelTrainer,
    AutoEasyDeLModelForCausalLM,
    EasyDeLOptimizers,
    EasyDeLSchedulers,
    EasyDeLGradientCheckPointers
)
from datasets import load_dataset
import flax
from jax import numpy as jnp
from transformers import AutoTokenizer

huggingface_repo_id_or_path = "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"

model, params = AutoEasyDeLModelForCausalLM.from_pretrained(huggingface_repo_id_or_path, )

max_length = 2048
tokenizer = AutoTokenizer.from_pretrained(
    huggingface_repo_id_or_path,
    trust_remote_code=True
)
tokenizer.pad_token = tokenizer.eos_token
configs_to_initialize_model_class = {
    "config": model.config,
    "dtype": jnp.bfloat16,
    "param_dtype": jnp.bfloat16,
    "input_shape": (1, 1)
}

train_arguments = TrainArguments(
    model_class=type(model),
    model_name="my_first_model_to_train_using_easydel",
    num_train_epochs=3,
    configs_to_initialize_model_class=configs_to_initialize_model_class,
    learning_rate=5e-5,
    learning_rate_end=1e-6,
    optimizer=EasyDeLOptimizers.ADAMW,  # "adamw", "lion", "adafactor" are supported
    scheduler=EasyDeLSchedulers.LINEAR,
    # "linear","cosine", "none" ,"warm_up_cosine" and "warm_up_linear"  are supported
    weight_decay=0.01,
    total_batch_size=64,
    max_training_steps=None,  # None to let trainer Decide
    do_train=True,
    do_eval=False,  # it's optional but supported 
    backend="tpu",  # default backed is set to cpu, so you must define you want to use tpu cpu or gpu
    max_length=max_length,  # Note that you have to change this in the model config too
    gradient_checkpointing=EasyDeLGradientCheckPointers.NOTHING_SAVEABLE,
    sharding_array=(1, -1, 1, 1),  # the way to shard model across gpu,cpu or TPUs using sharding array (1, -1, 1, 1)
    # everything training will be in sequence and model parallel automatic and share data between devices
    remove_ckpt_after_load=True,
    gradient_accumulation_steps=8,
    loss_re_mat="",
    dtype=jnp.bfloat16
)


def ultra_chat_prompting_process(
        data_chunk
):
    user_part = [
        chunk["content"] for chunk in data_chunk["messages"] if chunk["role"] == "user"
    ]
    assistant_part = [
        chunk["content"] for chunk in data_chunk["messages"] if chunk["role"] == "assistant"
    ]

    prompt = ""

    for uc, ac in zip(user_part, assistant_part):
        prompt += f"<|user|>\n{uc}</s>\n<|assistant|>\n{ac}</s>\n"

    return {"prompt": prompt}


tokenization_process = lambda data_chunk: tokenizer(
    data_chunk["prompt"],
    add_special_tokens=False,
    max_length=max_length,
    padding="max_length"
)

dataset = load_dataset("HuggingFaceH4/ultrachat_200k")
dataset_train = dataset["train_gen"].map(ultra_chat_prompting_process, num_proc=12)
dataset_train = dataset_train.map(
    tokenization_process,
    num_proc=12,
    remove_columns=dataset_train.column_names
)

# you can do the same for evaluation process dataset

trainer = CausalLanguageModelTrainer(
    train_arguments,
    dataset_train,
    checkpoint_path=None
)

output = trainer.train(flax.core.FrozenDict({"params": params}))
print(f"Hey ! , here's where your model saved {output.checkpoint_path}")

[!TIP] you can then convert it to pytorch for better use I don’t recommend jax/flax for hosting models since pytorch is better option for gpus

DPO Fine-tuning

DPOTrainer is the new Trainer in EasyDeL, so you might have except some bugs in process but as far as i have tested everything works just fine, and you can consider it the first DPO Trainer in JAX/Flax let have an example and see how you can fine-tune your own model with DPOTrainer

[!TIP] In case that you want a better script to learn about DPOTrainer you can see examples at here which contain DPO Tuning a Mixtral model with Intel DPO dataset.

from easydel import (
    TrainArguments,
    EasyDeLOptimizers,
    EasyDeLSchedulers,
    EasyDeLGradientCheckPointers,
    DPOTrainer,
    EasyDeLState,
    easystate_to_huggingface_model
)

from datasets import load_dataset
from huggingface_hub import HfApi
from transformers import AutoTokenizer, LlamaForCausalLM as module_pt
from jax import numpy as jnp
import jax
from jax.sharding import PartitionSpec
from fjformer import GenerateRNG
from typing import Optional, Dict
from datasets import Dataset

rng_g = GenerateRNG()
api = HfApi()

max_length = 512  # Overall maximum length
max_target_length = 1024  # Maximum Length for target column in Dataset
max_prompt_length = 1024  # Maximum Length for prompt column in Dataset

model_name_or_path = "erfanzar/LinguaMatic-Tiny"
ref_model_name_or_path = "teknium/OpenHermes-2.5-Mistral-7B"
dtype = jnp.bfloat16

sharding_axis_dims = (1, -1, 1, 1)
sharding_axis_names = ("dp", "fsdp", "tp", "sp")
query_partition_spec = PartitionSpec(
    ("dp", "fsdp"), "sp", "tp", None
)  # Query Partition Spec for Model
key_partition_spec = PartitionSpec(
    ("dp", "fsdp"), "sp", "tp", None
)  # Key Partition Spec for Model
value_partition_spec = PartitionSpec(
    ("dp", "fsdp"), "sp", "tp", None
)  # Value Partition Spec for Model
bias_partition_spec = PartitionSpec(
    ("dp", "fsdp"), None, None, None
)  # Attention Mask / Bias Partition Spec for Model
attention_partition_spec = PartitionSpec(
    ("dp", "fsdp"), "sp", "tp", None
)  # Attention Score / Weight Partition Spec for Model

ref_model_query_partition_spec = PartitionSpec(
    ("dp", "fsdp"), "sp", "tp", None
)  # Query Partition Spec for Ref Model
ref_model_key_partition_spec = PartitionSpec(
    ("dp", "fsdp"), "sp", "tp", None
)  # Key Partition Spec for Ref Model
ref_model_value_partition_spec = PartitionSpec(
    ("dp", "fsdp"), "sp", "tp", None
)  # Value Partition Spec for Ref Model
ref_model_bias_partition_spec = PartitionSpec(
    ("dp", "fsdp"), None, None, None
)  # Attention Mask / Bias Partition Spec for Ref Model
ref_model_attention_partition_spec = PartitionSpec(
    ("dp", "fsdp"), "sp", "tp", None
)  # Attention Score / Weight Partition Spec for Ref Model


def extract_anthropic_prompt(prompt_and_response):
    """
    Extract the anthropic prompt from a prompt and response pair.
    """
    search_term = "\n\nAssistant:"
    search_term_idx = prompt_and_response.rfind(search_term)
    assert search_term_idx != -1, f"Prompt and response does not contain '{search_term}'"
    return prompt_and_response[: search_term_idx + len(search_term)]


def get_hh(split: str, sanity_check: bool = False, silent: bool = False, cache_dir: Optional[str] = None) -> Dataset:
    """
    Load the Anthropic Helpful-Harmless dataset from Hugging Face and convert it to the necessary format.

    The dataset is converted to a dictionary with the following structure:
    {
        'prompt': List[str],
        'chosen': List[str],
        'rejected': List[str],
    }

    Prompts should be structured as follows:
      \n\nHuman: <prompt>\n\nAssistant:
    Multiple turns are allowed, but the prompt should always start with \n\nHuman: and end with \n\nAssistant:.
    """
    dataset = load_dataset("Anthropic/hh-rlhf", split=split, cache_dir=cache_dir)
    if sanity_check:
        dataset = dataset.select(range(min(len(dataset), 1000)))

    def split_prompt_and_responses(sample) -> Dict[str, str]:
        prompt = extract_anthropic_prompt(sample["chosen"])
        return {
            "prompt": prompt,
            "chosen": sample["chosen"][len(prompt):],
            "rejected": sample["rejected"][len(prompt):],
        }

    return dataset.map(split_prompt_and_responses)


arguments = TrainArguments(
    model_name="EasyDeL-DPO",
    num_train_epochs=5,
    learning_rate=1e-4,
    learning_rate_end=3e-5,
    warmup_steps=200,
    optimizer=EasyDeLOptimizers.ADAMW,
    scheduler=EasyDeLSchedulers.LINEAR,
    weight_decay=0.02,
    total_batch_size=128,
    gradient_checkpointing=EasyDeLGradientCheckPointers.NOTHING_SAVEABLE,
    sharding_array=sharding_axis_dims,
    fully_sharded_data_parallel=True,
    gradient_accumulation_steps=2,
    dtype=dtype,
    param_dtype=dtype,
    step_start_point=0,
    training_time="7H",
    do_train=True,
    do_eval=True,
    track_memory=False  # Performance boost.
    # You can set other options too or play with them but for now I just stick with these arguments.
)

tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

if tokenizer.pad_token_id is None:
    tokenizer.pad_token_id = tokenizer.eos_token_id

train_dataset = get_hh("train", sanity_check=True)
eval_dataset = get_hh("test", sanity_check=True)

state = EasyDeLState.from_pretrained(
    pretrained_model_name_or_path=model_name_or_path,
    dtype=dtype,
    param_dtype=dtype,
    init_optimizer_state=False,
    free_optimizer_state=True,
    sharding_axis_dims=sharding_axis_dims,
    sharding_axis_names=sharding_axis_names,
    query_partition_spec=query_partition_spec,
    key_partition_spec=key_partition_spec,
    value_partition_spec=value_partition_spec,
    bias_partition_spec=bias_partition_spec,
    attention_partition_spec=attention_partition_spec,
)

ref_state = EasyDeLState.from_pretrained(
    pretrained_model_name_or_path=ref_model_name_or_path,
    dtype=dtype,
    param_dtype=dtype,
    init_optimizer_state=False,
    free_optimizer_state=True,
    sharding_axis_dims=sharding_axis_dims,
    sharding_axis_names=sharding_axis_names,
    query_partition_spec=ref_model_query_partition_spec,
    key_partition_spec=ref_model_key_partition_spec,
    value_partition_spec=ref_model_value_partition_spec,
    bias_partition_spec=ref_model_bias_partition_spec,
    attention_partition_spec=ref_model_attention_partition_spec,
)

dpo_trainer = DPOTrainer(
    model_state=state,
    ref_model_state=ref_state,
    beta=0.1,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    tokenizer=tokenizer,
    arguments=arguments,
    max_length=max_length,
    max_target_length=max_target_length,
    max_prompt_length=max_prompt_length,
    ref_model_init_kwargs=None,  # In case that you pass the ref_model_state a string you have to pass this one too
    model_init_kwargs=None,  # In case that you pass the model_state a string you have to pass this one too
    dataset_map_arguments={
        "num_proc": 8,
        "batched": True,
        "batch_size": 100,
    },
    auto_shard_model_state=True,
    auto_shard_ref_model_state=True,
    loss_type="sigmoid",
    data_collator=None,  # Pass None in order to use default data_collector (you can create your own)
)

output = dpo_trainer.train()

easydel_jax_model = output.state  # Here's you EasyDeL Model

with jax.default_device(jax.devices("cpu")[0]):
    model = easystate_to_huggingface_model(
        state=EasyDeLState.load_state(
            output.checkpoint_path
        ),
        base_huggingface_module=module_pt,
        config=dpo_trainer.model_state.module.config
    )  # Here's you PyTorch Model

model.push_to_hub("<REPO_ID>", private=False)  # Hope you love open-source too :)
tokenizer.push_to_hub("<REPO_ID>", private=False)  # Hope you love open-source too :)

now you have trained your first model Using DPOTrainer in JAX with EasyDeL.

[!TIP] The API of EasyDeL DPO Trainer is similar to DPO Trainer in TRL from HuggingFace so that means you have freedom and have access to a hackable and changeable code.

EasyDeLState

EasyDeLState is new and cool feature in EasyDeL and have a lot of options like storing Model Parameters, Optimizer State, Model Config, Model Type, Optimizer and Scheduler Configs

Let see and examples of using EasyDeLState

Fine-tuning

Fine-tuning from a previous State or a new state

from easydel import (
    AutoEasyDeLConfig,
    EasyDeLState
)
from transformers import AutoTokenizer
from jax import numpy as jnp, lax
import jax

huggingface_model_repo_id = "REPO_ID"
checkpoint_name = "CKPT_NAME"

state = EasyDeLState.from_pretrained(
    pretrained_model_name_or_path=huggingface_model_repo_id,
    filename=checkpoint_name,
    optimizer="adamw",
    scheduler="none",
    tx_init=None,
    device=jax.devices('cpu')[0],  # Offload Device
    dtype=jnp.bfloat16,
    param_dtype=jnp.bfloat16,
    precision=lax.Precision("fastest"),
    sharding_axis_dims=(1, -1, 1, 1),
    # the way to shard model across gpu,cpu or TPUs using sharding array (1, -1, 1, 1)
    # everything training will be in sequence and model parallel automatic and share data between devices
    sharding_axis_names=("dp", "fsdp", "tp", "sp"),
    query_partition_spec=jax.sharding.PartitionSpec(("dp", "fsdp"), "sp", "tp", None),
    key_partition_spec=jax.sharding.PartitionSpec(("dp", "fsdp"), "sp", "tp", None),
    value_partition_spec=jax.sharding.PartitionSpec(("dp", "fsdp"), "sp", "tp", None),
    bias_partition_spec=jax.sharding.PartitionSpec(("dp", "fsdp"), None, None, None),
    attention_partition_spec=jax.sharding.PartitionSpec(("dp", "fsdp"), "sp", "tp", None),
    shard_attention_computation=True,
    input_shape=(1, 1),
    backend=None,
    init_optimizer_state=False,
    free_optimizer_state=True,
    verbose=True,
    state_shard_fns=None,
)

config = AutoEasyDeLConfig.from_pretrained(
    huggingface_model_repo_id
)

tokenizer = AutoTokenizer.from_pretrained(
    huggingface_model_repo_id,
    trust_remote_code=True
)

max_length = config.max_position_embeddings

configs_to_initialize_model_class = {
    'config': config,
    'dtype': jnp.bfloat16,
    'param_dtype': jnp.bfloat16,
    'input_shape': (8, 8)
}

EasyDeLState also has .load_state() and .save_state() with some other usable options like .free_opt_state() which free optimizer state or .shard_params() which shard parameters you can read docs in order to find out more about these options.

Converting to Huggingface and Pytorch

Let see how you can convert a EasyDeLMistral Model to Huggingface Pytorch Mistral Model from a trained State

from transformers import MistralForCausalLM
from easydel import (
    AutoEasyDeLConfig,
    EasyDeLState,
    easystate_to_huggingface_model
)
import jax

huggingface_model_repo_id = "REPO_ID"

config = AutoEasyDeLConfig.from_pretrained(
    huggingface_model_repo_id
)
with jax.default_device(jax.devices("cpu")[0]):
    model = easystate_to_huggingface_model(
        state=EasyDeLState.load_state(
            "PATH_TO_CKPT"
        ),  # You can Pass EasyDeLState here
        base_huggingface_module=MistralForCausalLM,  # type: ignore
        config=config
    )

model = model.half()  # it's a huggingface model now

Other Use Cases

EasyDeLState have a general use you can use it everywhere in easydel for example for a stand-alone model , serve, fine-tuning and many other features, it’s up to you to test how creative you are 😇.

AttentionModule: A Versatile Attention Mechanism Factory

The AttentionModule class is designed to simplify the creation and execution of different attention mechanisms within your EasyDeL models. It provides a unified interface for working with various attention types, allowing you to easily switch between them and experiment with different configurations.

Key Features:

Mechanism Selection: The attn_mechanism argument lets you choose the specific attention algorithm you want to use (e.g., “vanilla,” “flash,” “splash,” “ring,” “cudnn”).
Sharding and Partitioning: The class supports advanced JAX sharding techniques to distribute attention computations across multiple devices for efficient processing of large models. It handles partitioning of query, key, value, bias, and attention weight matrices using PartitionSpec.
Blockwise Attention: Enables the use of blockwise attention for increased memory efficiency, especially with long sequences.
Caching Support: Facilitates the use of attention caching to speed up inference and generation tasks.
Dropout and Determinism: Allows for applying dropout to attention weights and controlling the deterministic behavior of the attention computation.
Testing Utility: Provides a test_attentions method to compare different attention mechanisms in terms of accuracy, gradient stability, and computation time.

How it Works:

Initialization:
- During initialization, you provide the desired attn_mechanism, JAX mesh for sharding, scaling factor (sm_scale), number of attention heads, head dimensions, and other configuration parameters.
- The class automatically sets default values for many parameters based on the chosen attention mechanism and the provided EasyDeL configuration (base_module_class).
Calling the Module:
- When you call the AttentionModule object, you pass in the query, key, and value states, along with optional parameters like attention masks, biases, and causal flags.
- The module internally selects the appropriate attention function based on the specified attn_mechanism.
- It performs any necessary sharding and partitioning based on the configured partition specifications.
- The attention computation is executed, and the attention outputs (and optionally attention weights) are returned.

Advantages:

Flexibility: Allows you to easily switch between different attention mechanisms without major code changes.
Efficiency: Supports advanced JAX sharding for distributed computation, enabling the handling of large models.

Flash Attention works on TPU with ease but for gpu there are still some improvements in process.

[!TIP] use these partition specs in case of not using custom sharding_axis_names and using sequence sharding with flash flash attention

query_partition_spec=PartitionSpec(("dp", "fsdp"), None, "sp", "tp"),
generation_query_partition_spec=PartitionSpec(("dp", "fsdp"), None, None, "tp"),
key_partition_spec=PartitionSpec(("dp", "fsdp"), None, "sp", "tp"),
value_partition_spec=PartitionSpec(("dp", "fsdp"), None, "sp", "tp"),
attention_partition_spec=PartitionSpec(("dp", "fsdp"), None,"sp", "tp"),

EasyDeLXRapTure for layer tuning and LoRA

in case of using LoRA and applying that on the EasyDeL models there are some other things that you might need to config on your own but a lot of things being handled by EasyDeL so let just jump into an example for LoRA fine-tuning section and use EasyDeLXRapTure in for mistral models with flash attention example

from flax.core import FrozenDict
from easydel import (
    TrainArguments,
    CausalLanguageModelTrainer,
    AutoEasyDeLModelForCausalLM,
    EasyDeLOptimizers,
    EasyDeLSchedulers,
    EasyDeLGradientCheckPointers,
    EasyDeLXRapTureConfig
)
from datasets import load_dataset
import flax
from jax import numpy as jnp
from transformers import AutoTokenizer

huggingface_repo_id_or_path = "mistralai/Mistral-7B-Instruct-v0.1"

model, params = AutoEasyDeLModelForCausalLM.from_pretrained(huggingface_repo_id_or_path, )

max_length = 8196
model_parameters = FrozenDict({"params": params})

dtype = jnp.bfloat16
param_dtype = jnp.bfloat16  # you can change that if you want 

tokenizer = AutoTokenizer.from_pretrained(
    huggingface_repo_id_or_path,
    trust_remote_code=True
)

model.config.add_basic_configurations(
    attn_mechanism="flash",  # Using FlashAttention
    block_b=1,
    block_q=1024,
    block_k=1024,
    block_k_major=1024,
)

tokenizer.pad_token = tokenizer.eos_token
configs_to_initialize_model_class = {
    "config": model.config,
    "dtype": dtype,
    "param_dtype": param_dtype,
    "input_shape": (1, 1)
}

rapture = EasyDeLXRapTureConfig(
    parameters=model_parameters,
    lora_dim=64,
    fully_fine_tune_parameters=["embed_tokens"],  # Model layer to be fully fine tuned
    lora_fine_tune_parameters=["q_proj", "v_proj", "k_proj", "o_proj"],  # LoRA Layer Targets you can pass this to none
    # For only Layer Tuning or transfer learning
    verbose=True
)

train_arguments = TrainArguments(
    model_class=type(model),
    model_name="EasyDeL-Lora-Example",
    num_train_epochs=3,
    configs_to_initialize_model_class=configs_to_initialize_model_class,
    learning_rate=1e-4,  # Using higher learning rate is recommended
    learning_rate_end=8e-5,
    optimizer=EasyDeLOptimizers.ADAMW,  # "adamw", "lion", "adafactor" are supported
    scheduler=EasyDeLSchedulers.LINEAR,
    # "linear","cosine", "none" ,"warm_up_cosine" and "warm_up_linear"  are supported
    weight_decay=0.01,
    total_batch_size=512,
    max_training_steps=None,  # None to let trainer Decide
    do_train=True,
    do_eval=False,  # it's optional but supported 
    backend="tpu",  # default backed is set to cpu, so you must define you want to use tpu cpu or gpu
    max_length=max_length,  # Note that you have to change this in the model config too
    gradient_checkpointing=EasyDeLGradientCheckPointers.NOTHING_SAVEABLE,
    sharding_array=(1, -1, 1, 1),  # the way to shard model across gpu,cpu or TPUs using sharding array (1, -1, 1, 1)
    # everything training will be in sequence and model parallel automatic and share data between devices
    remove_ckpt_after_load=True,
    gradient_accumulation_steps=1,
    loss_re_mat="",
    dtype=dtype,
    param_dtype=param_dtype,
    rapture_config=rapture,
    merge_lora_rapture_parameters=True  # turning this off is still not supported and not recommended to do so
    # What this does ? this will merge the lora parameters with the original model parameters and the end of training
)


def ultra_chat_prompting_process(
        data_chunk
):
    user_part = [
        chunk["content"] for chunk in data_chunk["messages"] if chunk["role"] == "user"
    ]
    assistant_part = [
        chunk["content"] for chunk in data_chunk["messages"] if chunk["role"] == "assistant"
    ]

    prompt = ""

    for uc, ac in zip(user_part, assistant_part):
        prompt += f"<|user|>\n{uc}</s>\n<|assistant|>\n{ac}</s>\n"

    return {"prompt": prompt}


tokenization_process = lambda data_chunk: tokenizer(
    data_chunk["prompt"],
    add_special_tokens=False,
    max_length=max_length,
    padding="max_length"
)

dataset = load_dataset("HuggingFaceH4/ultrachat_200k")
dataset_train = dataset["train_gen"].map(ultra_chat_prompting_process, num_proc=12)
dataset_train = dataset_train.map(
    tokenization_process,
    num_proc=12,
    remove_columns=dataset_train.column_names
)

# you can do the same for evaluation process dataset

trainer = CausalLanguageModelTrainer(
    train_arguments,
    dataset_train,
    checkpoint_path=None
)

output = trainer.train()  # you should not pass the parameters in Trainer.train anymore when
# you are using LoRA or transfer Learning
print(f"Hey ! , here's where your model saved {output.checkpoint_path}")

Contributing

EasyDeL is an open-source project, and contributions are welcome. If you would like to contribute to EasyDeL, please fork the repository, make your changes, and submit a pull request. The team behind EasyDeL will review your changes and merge them if they are suitable.

License 📜

EasyDeL is a Fully Open-Source released under the Apache v2 license. Please see the LICENSE file in the root directory of this project for more information.

Contact

If you have any questions or comments about EasyDeL, you can reach out to me

Citing EasyDeL 🥶

To cite this repository:

@misc{Zare Chavoshi_2023,
    title={EasyDeL, an open-source library, is specifically designed to enhance and streamline the training process of machine learning models. It focuses primarily on Jax/Flax and aims to provide convenient and effective solutions for training Flax/Jax Models on TPU/GPU for both Serving and Training purposes.},
    url={https://github.com/erfanzar/EasyDeL},
    journal={EasyDeL Easy and Fast DeepLearning with JAX},
    publisher={Erfan Zare Chavoshi},
    author={Zare Chavoshi, Erfan},
    year={2023}
}

This site is open source. Improve this page.