Skip to content

(EasyDel Former) is a utility library designed to simplify and enhance the development in JAX

License

Notifications You must be signed in to change notification settings

erfanzar/eformer

Repository files navigation

eformer (EasyDel Former)

License Python JAX

eformer (EasyDel Former) is a utility library designed to simplify and enhance the development of machine learning models using JAX. It provides a collection of tools for sharding, custom PyTrees, quantization, mixed precision training, and optimized operations, making it easier to build and scale models efficiently.

Features

  • Mixed Precision Training (mpric): Advanced mixed precision utilities supporting float8, float16, and bfloat16 with dynamic loss scaling.
  • Sharding Utilities (escale): Tools for efficient sharding and distributed computation in JAX.
  • Custom PyTrees (jaximus): Enhanced utilities for creating custom PyTrees and ArrayValue objects, updated from Equinox.
  • Custom Calling (callib): A tool for custom function calls and direct integration with Triton kernels in JAX.
  • Optimizer Factory: A flexible factory for creating and configuring optimizers like AdamW, Adafactor, Lion, and RMSProp.
  • Custom Operations and Kernels:
    • Flash Attention 2 for GPUs/TPUs (via Triton and Pallas).
    • 8-bit and NF4 quantization for efficient model.
    • Many others to be added.
  • Quantization Support: Tools for 8-bit and NF4 quantization, enabling memory-efficient model deployment.

Installation

You can install eformer via pip:

pip install eformer

Quick Start

Mixed Precision Handler with mpric

from eformer.mpric import PrecisionHandler

# Create a handler with float8 compute precision
handler = PrecisionHandler(
    policy="p=f32,c=f8_e4m3,o=f32",  # params in f32, compute in float8, output in f32
    use_dynamic_scale=True
)

Customizing Arrays With ArrayValue

import jax

from eformer.jaximus import ArrayValue, implicit
from eformer.ops.quantization.quantization_functions import (
    dequantize_row_q8_0,
    quantize_row_q8_0,
)

array = jax.random.normal(jax.random.key(0), (256, 64), "f2")


class Array8B(ArrayValue):
    scale: jax.Array
    weight: jax.Array

    def __init__(self, array: jax.Array):
        self.weight, self.scale = quantize_row_q8_0(array)

    def materialize(self):
        return dequantize_row_q8_0(self.weight, self.scale)


qarray = Array8B(array)


@jax.jit
@implicit
def sqrt(x):
    return jax.numpy.sqrt(x)


print(sqrt(qarray))
print(qarray)

Optimizer Factory

from eformer.optimizers import OptimizerFactory, SchedulerConfig, AdamWConfig

# Create an AdamW optimizer with a cosine scheduler
scheduler_config = SchedulerConfig(scheduler_type="cosine", learning_rate=1e-3, steps=1000)
optimizer, scheduler = OptimizerFactory.create("adamw", scheduler_config, AdamWConfig())

Quantization

from eformer.quantization import Array8B, ArrayNF4

# Quantize an array to 8-bit
qarray = Array8B(jax.random.normal(jax.random.key(0), (256, 64), "f2"))

# Quantize an array to NF4
n4array = ArrayNF4(jax.random.normal(jax.random.key(0), (256, 64), "f2"), 64)

Advanced Mixed Precision Configuration

from eformer.mpric import Policy, LossScaleConfig

# Create a custom precision policy
policy = Policy(
    param_dtype=jnp.float32,
    compute_dtype=jnp.bfloat16,
    output_dtype=jnp.float32
)

# Configure loss scaling
loss_config = LossScaleConfig(
    initial_scale=2**15,
    growth_interval=2000,
    scale_factor=2,
    min_scale=1.0
)

# Create handler with custom configuration
handler = PrecisionHandler(
    policy=policy,
    use_dynamic_scale=True,
    loss_scale_config=loss_config
)

Contributing

We welcome contributions! Please read our Contributing Guidelines to get started.

License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

About

(EasyDel Former) is a utility library designed to simplify and enhance the development in JAX

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •  

Languages