erfanzar/xerxes-pro

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, EasyDeL aims to provide convenient and effective solutions for training Flax/Jax models on TPU/GPU for both serving and training purposes.

Using Example

Using From EasyDeLState (*.easy files)

from easydel import EasyDeLState, AutoShardAndGatherFunctions
from jax import numpy as jnp, lax

shard_fns, gather_fns = AutoShardAndGatherFunctions.from_pretrained(
    "REPO_ID", # Pytorch State should be saved to in order to find shard gather fns with no effort, otherwise read docs.
    backend="gpu",
    depth_target=["params", "params"],
    flatten=False
)

state = EasyDeLState.load_state(
    "*.easy",
    dtype=jnp.float16,
    param_dtype=jnp.float16,
    precision=lax.Precision("fastest"),
    verbose=True,
    state_shard_fns=shard_fns
)
# State file Ready to use ...

Using From AutoEasyDeLModelForCausalLM (from PyTorch)

from easydel import AutoEasyDeLModelForCausalLM
from jax import numpy as jnp, lax


model, params = AutoEasyDeLModelForCausalLM.from_pretrained(
    "REPO_ID",
    dtype=jnp.float16,
    param_dtype=jnp.float16,
    precision=lax.Precision("fastest"),
    auto_shard_params=True,
)
# Model and Parameters Ready to use ...

Using From AutoEasyDeLModelForCausalLM (from EasyDeL)

from easydel import AutoEasyDeLModelForCausalLM
from jax import numpy as jnp, lax


model, params = AutoEasyDeLModelForCausalLM.from_pretrained(
    "REPO_ID/",
    dtype=jnp.float16,
    param_dtype=jnp.float16,
    precision=lax.Precision("fastest"),
    auto_shard_params=True,
    from_torch=False
)
# Model and Parameters Ready to use ...