1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
|
from .base import *
import einx.tracer as tracer
from einx.tracer.tensor import op
import einx, types
from functools import partial
def create():
import mlx.core as mx
tmx = tracer.import_("mlx.core", "mx")
def to_tuple(x):
if isinstance(x, tuple):
return x
elif isinstance(x, list):
return tuple(x)
elif isinstance(x, np.ndarray):
return tuple(x.tolist())
else:
raise ValueError(f"Cannot convert {type(x)} to tuple")
def to_dtype(x):
if isinstance(x, str):
if x == "bool":
return mx.bool_
else:
return vars(mx)[x]
else:
return x
to_dtype2 = to_dtype
class mlx(Backend):
name = "mlx"
tensor_types = [mx.array]
to_dtype = staticmethod(to_dtype2)
@staticmethod
@einx.trace
def to_tensor(tensor, shape):
return einx.tracer.apply(
tmx.array,
args=[tensor],
output=einx.tracer.Tensor(shape),
)
@staticmethod
@einx.trace
def reshape(tensor, shape):
if einx.tracer.is_scalar(tensor):
tensor = tmx.array(tensor)
return einx.tracer.apply(
tmx.reshape, args=[tensor, list(to_tuple(shape))], output=einx.tracer.Tensor(shape)
)
transpose = op.transpose(tmx.transpose)
broadcast_to = op.broadcast_to(tmx.broadcast_to)
@staticmethod
@einx.trace
def einsum(equation, *tensors):
raise NotImplementedError("mlx does not support einsum yet")
@staticmethod
@einx.trace
def arange(start, stop=None, step=None, dtype="int32"):
args = [start]
if stop is not None:
args.append(stop)
if step is not None:
args.append(step)
return op.arange(tmx.arange)(*args, dtype=to_dtype(dtype))
stack = op.stack(tmx.stack)
concatenate = op.concatenate(tmx.concatenate)
add = associative_binary_to_nary(op.elementwise(tmx.add))
subtract = op.elementwise(tmx.subtract)
multiply = associative_binary_to_nary(op.elementwise(tmx.multiply))
true_divide = op.elementwise(tmx.divide)
floor_divide = op.elementwise(tmx.floor_divide)
divide = op.elementwise(tmx.divide)
mod = op.elementwise(tmx.remainder)
logical_and = associative_binary_to_nary(op.elementwise(tmx.logical_and))
logical_or = associative_binary_to_nary(op.elementwise(tmx.logical_or))
where = op.elementwise(tmx.where)
less = op.elementwise(tmx.less)
less_equal = op.elementwise(tmx.less_equal)
greater = op.elementwise(tmx.greater)
greater_equal = op.elementwise(tmx.greater_equal)
equal = op.elementwise(tmx.equal)
not_equal = op.elementwise(tmx.not_equal)
maximum = associative_binary_to_nary(op.elementwise(tmx.maximum))
minimum = associative_binary_to_nary(op.elementwise(tmx.minimum))
sum = op.reduce(tmx.sum)
mean = op.reduce(tmx.mean)
var = op.reduce(tmx.var)
prod = op.reduce(tmx.prod)
count_nonzero = op.reduce(tmx.count_nonzero)
any = op.reduce(tmx.any)
all = op.reduce(tmx.all)
min = op.reduce(tmx.min)
max = op.reduce(tmx.max)
logsumexp = op.reduce(tmx.logsumexp)
log = op.elementwise(tmx.log)
exp = op.elementwise(tmx.exp)
sqrt = op.elementwise(tmx.sqrt)
rsqrt = op.elementwise(tmx.rsqrt)
square = op.elementwise(tmx.square)
@staticmethod
@einx.trace
def get_at(tensor, coordinates):
return tensor[coordinates]
@staticmethod
@einx.trace
def set_at(tensor, coordinates, updates):
return einx.tracer.apply(
tensor.at[coordinates].set, args=[updates], output=einx.tracer.Tensor(tensor.shape)
)
@staticmethod
@einx.trace
def add_at(tensor, coordinates, updates):
return einx.tracer.apply(
tensor.at[coordinates].add, args=[updates], output=einx.tracer.Tensor(tensor.shape)
)
@staticmethod
@einx.trace
def subtract_at(tensor, coordinates, updates):
return einx.tracer.apply(
tensor.at[coordinates].add, args=[-updates], output=einx.tracer.Tensor(tensor.shape)
)
softmax = op.keep_shape(tmx.softmax)
stop_gradient = op.keep_shape(tmx.stop_gradient)
# vmap = op.vmap(tmx.vmap)
@staticmethod
def vmap(op, in_axes, out_axes, input_shapes=None, output_shapes=None):
raise NotImplementedError("mlx does not fully support vmap yet")
sqrt = tmx.sqrt
rsqrt = tmx.rsqrt
square = tmx.square
class random:
@einx.trace
def bernoulli(rng, p, shape):
einx.tracer.apply(
tmx.random.bernoulli,
args=[p, shape, rng],
output=einx.tracer.Tensor(shape),
)
return mlx()
|
