Spaces:
Paused
Paused
File size: 3,768 Bytes
2f86d9b | 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 | # Copyright 2026 DeepMind Technologies Limited.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from etils import etree
from etils.enp.typing import ArrayAliasMeta # pylint: disable=g-importing-member
from etils.enp.typing import f32
from flax import linen as nn
from gemma import peft
import jax
from jax import numpy as jnp
import numpy as np
i4 = ArrayAliasMeta(shape=None, dtype=jnp.int4) # pyrefly: ignore[bad-argument-type]
def _dense_to_quantized(module):
if isinstance(module, nn.Dense):
return peft.SimulateQuantizedDense(
wrapped=module, method=peft.QuantizationMethod.INT4
)
if isinstance(module, nn.Einsum):
return peft.SimulateQuantizedEinsum(
wrapped=module, method=peft.QuantizationMethod.INT4
)
else:
return module
def _dense_to_int4(module):
if isinstance(module, nn.Dense):
return peft.IntDense(wrapped=module, dtype=jnp.int4) # pyrefly: ignore[bad-argument-type]
if isinstance(module, nn.Einsum):
return peft.IntEinsum(wrapped=module, dtype=jnp.int4) # pyrefly: ignore[bad-argument-type]
else:
return module
class MyModule(nn.Module):
@nn.compact
def __call__(self, x):
# Call a model twice
model = nn.Dense(2)
y0 = model(x)
y1 = model(x)
# Call another model
y2 = nn.Dense(3)(x)
# Test Einsum
y3 = nn.Einsum(
shape=(4, 2, 3),
einsum_str='bi,imn->bmn',
)(x)
return {
'y0': y0,
'y1': y1,
'y2': y2,
'y3': y3,
}
def test_quantization_simulation():
model = MyModule()
with peft.ModuleInterceptor(_dense_to_quantized):
out, params = model.init_with_output(
jax.random.key(0),
jnp.zeros((1, 4)),
)
assert etree.spec_like(params) == etree.spec_like({
'params': {
'Dense_0': {
'kernel': f32[4, 2],
'bias': f32[2],
},
'Dense_1': {
'kernel': f32[4, 3],
'bias': f32[3],
},
'Einsum_0': {
'kernel': f32[4, 2, 3],
'bias': f32[2, 3],
},
},
})
assert etree.spec_like(out) == etree.spec_like({
'y0': f32[1, 2],
'y1': f32[1, 2],
'y2': f32[1, 3],
'y3': f32[1, 2, 3],
})
def test_quantization():
model = MyModule()
params = model.init(jax.random.key(0), jnp.zeros((1, 4)))
params_q = peft.quantize(
params,
method=peft.QuantizationMethod.INT4,
in_place_keys=True,
checkpoint_kernel_key='kernel',
)
with peft.ModuleInterceptor(_dense_to_int4):
out = model.apply(params_q, jnp.zeros((1, 4)))
assert etree.spec_like(params_q) == etree.spec_like({
'params': {
'Dense_0': {
'kernel': i4[4, 2],
'bias': f32[2],
'scale': f32[1, 2],
},
'Dense_1': {
'kernel': i4[4, 3],
'bias': f32[3],
'scale': f32[1, 3],
},
'Einsum_0': {
'kernel': i4[4, 2, 3],
'bias': f32[2, 3],
'scale': f32[1, 1, 3],
},
},
})
assert etree.spec_like(out) == etree.spec_like({
'y0': f32[1, 2],
'y1': f32[1, 2],
'y2': f32[1, 3],
'y3': f32[1, 2, 3],
})
|