Introdução ao MLX para Apple Silicon

3:48 - Basics

import mlx.core as mx

# Make an array
a = mx.array([1, 2, 3])

# Make another array
b = mx.array([4, 5, 6])

# Do an operation
c = a + b

# Access information about the array
shape = c.shape
dtype = c.dtype

print(f"Result c: {c}")
print(f"Shape: {shape}")
print(f"Data type: {dtype}")

5:31 - Unified memory

import mlx.core as mx

a = mx.array([1, 2, 3])
b = mx.array([4, 5, 6])

c = mx.add(a, b, stream=mx.gpu)
d = mx.multiply(a, b, stream=mx.cpu)

print(f"c computed on the GPU: {c}")
print(f"d computed on the CPU: {d}")

6:20 - Lazy computation

import mlx.core as mx

# Make an array
a = mx.array([1, 2, 3])

# Make another array
b = mx.array([4, 5, 6])

# Do an operation
c = a + b

# Evaluates c before printing it
print(c)

# Also evaluates c
c_list = c.tolist()

# Also evaluates c
mx.eval(c)

print(f"Evaluate c by converting to list: {c_list}")
print(f"Evaluate c using print: {c}")
print(f"Evaluate c using mx.eval(): {c}")

7:32 - Function transformation

import mlx.core as mx

def sin(x):
    return mx.sin(x)

dfdx = mx.grad(sin)

def sin(x):
    return mx.sin(x)

d2fdx2 = mx.grad(mx.grad(mx.sin))

# Computes the second derivative of sine at 1.0
d2fdx2(mx.array(1.0))

9:16 - Neural Networks in MLX

import mlx.core as mx
import mlx.nn as nn
import mlx.optimizers as optim

class MLP(nn.Module):
    """A simple MLP."""

    def __init__(self, dim, h_dim):
        super().__init__()
        self.linear1 = nn.Linear(dim, h_dim)
        self.linear2 = nn.Linear(h_dim, dim)

    def __call__(self, x):
        x = self.linear1(x)
        x = nn.relu(x)
        x = self.linear2(x)
        return x

9:57 - MLX and PyTorch

# MLX version
import mlx.core as mx
import mlx.nn as nn
import mlx.optimizers as optim

class MLP(nn.Module):
    """A simple MLP."""

    def __init__(self, dim, h_dim):
        super().__init__()
        self.linear1 = nn.Linear(dim, h_dim)
        self.linear2 = nn.Linear(h_dim, dim)

    def __call__(self, x):
        x = self.linear1(x)
        x = nn.relu(x)
        x = self.linear2(x)
        return x

# PyTorch version
import torch
import torch.nn as nn
import torch.optim as optim

class MLP(nn.Module):
    """A simple MLP."""

    def __init__(self, dim, h_dim):
        super().__init__()
        self.linear1 = nn.Linear(dim, h_dim)
        self.linear2 = nn.Linear(h_dim, dim)

    def forward(self, x):
        x = self.linear1(x)
        x = x.relu()
        x = self.linear2(x)
        return x

11:35 - Compiling MLX functions

import mlx.core as mx
import math

def gelu(x):
    return x * (1 + mx.erf(x / math.sqrt(2))) / 2

@mx.compile
def compiled_gelu(x):
    return x * (1 + mx.erf(x / math.sqrt(2))) / 2

x = mx.random.normal(shape=(4,))

out = gelu(x)
compiled_out = compiled_gelu(x)
print(f"gelu:          {out}")
print(f"compiled gelu: {compiled_out}")

12:32 - MLX Fast package

import mlx.core as mx
import time

def rms_norm(x, weight, eps=1e-5):
    y = x.astype(mx.float32)
    y = y * mx.rsqrt(mx.mean(
        mx.square(y), 
        axis=-1, 
        keepdims=True,
    ) + eps)
    return (weight * y).astype(x.dtype)

batch_size = 8192
feature_dim = 4096
iterations = 1000

x = mx.random.normal([batch_size, feature_dim])
weight = mx.ones(feature_dim)
bias = mx.zeros(feature_dim)

start_time = time.perf_counter()
for _ in range(iterations):
    y = rms_norm(x, weight, eps=1e-5)
    mx.eval(y)
rms_norm_time = time.perf_counter() - start_time
print(f"rms_norm execution: {gelu_time:0.4f} sec")

start_time = time.perf_counter()
for _ in range(iterations):
    mx.eval(mx.fast.rms_norm(x, weight, eps=1e-5))
fast_rms_norm_time = time.perf_counter() - start_time
print(f"mx.fast.rms_norm execution: {compiled_gelu_time:0.4f} sec")

print(f"mx.fast.rms_norm speedup: {rms_norm_time/fast_rms_norm_time:0.2f}x")

13:30 - Custom Metal kernel

import mlx.core as mx

# Build the kernel
source = """
    uint elem = thread_position_in_grid.x;
    out[elem] = metal::exp(inp[elem]);
"""
kernel = mx.fast.metal_kernel(
    name="myexp",
    input_names=["inp"],
    output_names=["out"],
    source=source,
)

# Call the kernel on a sample input
x = mx.array([1.0, 2.0, 3.0])
out = kernel(
    inputs=[x],
    grid=(x.size, 1, 1),
    threadgroup=(256, 1, 1),
    output_shapes=[x.shape],
    output_dtypes=[x.dtype],
)[0]
print(out)

14:41 - Quantization

import mlx.core as mx

x = mx.random.normal([1024])
weight = mx.random.normal([1024, 1024])

quantized_weight, scales, biases = mx.quantize(
        weight, bits=4, group_size=32,
)

y = mx.quantized_matmul(
    x,
    quantized_weight,
    scales=scales,
    biases=biases,
    bits=4,
    group_size=32,
)

w_orig = mx.dequantize(
    quantized_weight,
    scales=scales,
    biases=biases,
    bits=4,
    group_size=32,
)

15:23 - Quantized models

import mlx.nn as nn

model = nn.Sequential(
    nn.Embedding(100, 32),
    nn.Linear(32, 32),
    nn.Linear(32, 32),
    nn.Linear(32, 1),
)

print(model)

nn.quantize(
    model,
    bits=4,
    group_size=32,
)

print(model)

16:50 - Distributed

import mlx.core as mx

group = mx.distributed.init()

world_size = group.size()
rank = group.rank()

x = mx.array([1.0])

x_sum = mx.distributed.all_sum(x)

print(x_sum)

17:20 - Distributed launcher

mlx.launch --hosts ip1, ip2, ip3, ip4 my_script.py

18:20 - MLX Swift

// Swift
import MLX

// Make an array
let a = MLXArray([1, 2, 3])

// Make another array
let b = MLXArray([1, 2, 3])

// Do an operation
let c = a + b

// Access information about the array
let shape = c.shape
let dtype = c.dtype

// Print results
print("a: \(a)")
print("b: \(b)")
print("c = a + b: \(c)")
print("shape: \(shape)")
print("dtype: \(dtype)")