-
Notifications
You must be signed in to change notification settings - Fork 17
/
Copy pathtest_page_attention.py
207 lines (172 loc) · 6.64 KB
/
test_page_attention.py
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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
import unittest
import jax
import numpy as np
import jax.numpy as jnp
import torch
from jetstream_pt.third_party.llama import model_args
from jetstream_pt import environment
from jetstream_pt.page_attention_manager import PageAttentionManager
from jetstream_pt.cache_manager import PageKVCacheGenerate, KVCachePrefill
from absl.testing import parameterized
P = jax.sharding.PartitionSpec
class PageAttentionTest(parameterized.TestCase):
def _make_env(self, bf16_enable=True):
torch_dtype = torch.bfloat16 if bf16_enable else torch.float32
torch.set_default_dtype(torch_dtype)
jax.config.update("jax_dynamic_shapes", False)
jax.config.update("jax_traceback_filtering", "off")
jax.config.update("jax_platform_name", "cpu")
jax.config.update("jax_enable_x64", False)
mesh = jax.sharding.Mesh(np.array(jax.devices()), axis_names=("x",))
replicated = jax.sharding.NamedSharding(mesh, P())
config = model_args.get_model_args("tiny", 128, 1, True)
environment_data = environment.JetEngineEnvironmentData()
environment_data.max_input_sequence_length = 128
environment_data.max_input_sequence_length = 128
environment_data.cache_sequence_length = 128
environment_data.bf16_enable = bf16_enable
environment_data.model_type = "llama-2-tiny"
environment_data.batch_size = 3
environment_data.num_layers = config.n_layers
environment_data.cache_shape = (
1,
config.n_kv_heads,
environment_data.cache_sequence_length,
config.dim // config.n_heads,
)
env = environment.JetEngineEnvironment(environment_data)
env.apply_sharding = lambda *args, **kwargs: None # don't shard on cpu
env.sharding = replicated
return env, config
def test_prefill_insert(self):
env, _ = self._make_env()
pam = PageAttentionManager(
batch_size=3,
paged_attention_total_num_pages=20,
paged_attention_page_size=4,
max_pages_per_sequence=4,
)
shape = (1, 6, 4, 2)
decode_caches = []
decode_caches.append(
PageKVCacheGenerate.empty(shape=shape, device=None, env=env)
)
decode_caches = [c.state() for c in decode_caches]
prefill_chache = KVCachePrefill()
k, v = jnp.arange(6), jnp.arange(6)
k, v = jnp.reshape(k, (1, 1, 3, 2)), jnp.reshape(k, (1, 1, 3, 2))
prefill_chache.update(k, v, 0)
prefill_caches = [prefill_chache]
prefill_caches = [c.state() for c in prefill_caches]
num_pages, update_indexes = pam.reserve_pages_insert(0, 3)
_, kv_heads, _, dim = prefill_caches[0][0].shape
tep_kv = jnp.zeros((kv_heads, num_pages * 4, dim), dtype=jnp.bfloat16)
caches = pam.insert_prefill_cache(
prefill_caches=prefill_caches,
decode_caches=decode_caches,
update_indexes=update_indexes,
tep_kv=tep_kv,
sharding=env.sharding,
)
expected_kv = jnp.arange(6).reshape(3, 2)
padding = jnp.asarray([[0, 0]])
expected_kv = jnp.concatenate((expected_kv, padding))
self.assertTrue(
jnp.array_equal(
caches[0][0][0, 0, 0:4, 0:2], expected_kv.astype(jnp.bfloat16)
)
)
self.assertTrue(
jnp.array_equal(
caches[0][1][0, 0, 0:4, 0:2], expected_kv.astype(jnp.bfloat16)
)
)
def test_prefill_insert_multiple_pages(self):
jax.config.update("jax_platform_name", "cpu")
print(f"---------> {jax.devices()}")
env, _ = self._make_env()
pam = PageAttentionManager(
batch_size=3,
paged_attention_total_num_pages=20,
paged_attention_page_size=4,
max_pages_per_sequence=4,
)
shape = (1, 20, 4, 2)
decode_caches = []
decode_caches.append(
PageKVCacheGenerate.empty(shape=shape, device=None, env=env)
)
decode_caches = [c.state() for c in decode_caches]
self.cache_sharding = env.cache_sharding
prefill_chache = KVCachePrefill()
k, v = jnp.arange(12), jnp.arange(12)
k, v = jnp.reshape(k, (1, 1, 6, 2)), jnp.reshape(k, (1, 1, 6, 2))
prefill_chache.update(k, v, 0)
prefill_caches = [prefill_chache]
prefill_caches = [c.state() for c in prefill_caches]
num_pages, update_indexes = pam.reserve_pages_insert(0, 6)
_, kv_heads, _, dim = prefill_caches[0][0].shape
tep_kv = jnp.zeros((kv_heads, num_pages * 4, dim), dtype=jnp.bfloat16)
decode_caches = pam.insert_prefill_cache(
prefill_caches=prefill_caches,
decode_caches=decode_caches,
update_indexes=update_indexes,
tep_kv=tep_kv,
sharding=env.sharding,
)
self.assertEqual(len(decode_caches), 1)
expected = jnp.arange(16).at[12:16].set([0, 0, 0, 0]).reshape(1, 2, 4, 2)
updated_k = jax.lax.slice_in_dim(decode_caches[0][0], 0, 2, axis=1)
self.assertTrue(jnp.array_equal(updated_k, expected))
noupdated_k = jax.lax.slice_in_dim(decode_caches[0][0], 2, 20, axis=1)
self.assertTrue(jnp.array_equal(noupdated_k, jnp.zeros_like(noupdated_k)))
def test_reserve_pages_decode(self):
env, _ = self._make_env()
pam = PageAttentionManager(
batch_size=3,
paged_attention_total_num_pages=20,
paged_attention_page_size=4,
max_pages_per_sequence=4,
)
slot = 1
seq_len = 8
pam.reserve_pages_insert(slot, seq_len)
expected_slot_page_indices = np.asarray([0, 1])
slot_page_indices = pam.page_indices[slot][0:2]
self.assertTrue(
np.array_equal(slot_page_indices, expected_slot_page_indices)
)
lens = np.asarray([0, seq_len, 0])
pam.fill_new_pages(lens)
expected_slot_page_indices = np.asarray([0, 1, 2, 19])
slot_page_indices = pam.page_indices[slot]
self.assertTrue(
np.array_equal(slot_page_indices, expected_slot_page_indices)
)
expected_0_page_indices = np.asarray([19, 19, 19, 19])
zer0_page_indices = pam.page_indices[0][0:4]
self.assertTrue(np.array_equal(zer0_page_indices, expected_0_page_indices))
def test_get_page_token_indices(self):
env, _ = self._make_env()
pam = PageAttentionManager(
batch_size=5,
paged_attention_total_num_pages=20,
paged_attention_page_size=4,
max_pages_per_sequence=4,
)
pam.reserve_pages_insert(1, 8)
pam.reserve_pages_insert(3, 13)
pam.reserve_pages_insert(0, 3)
lens = np.asarray([3, 8, 0, 13, 0])
pam.fill_new_pages(lens)
page_token_indices = pam.get_page_token_indices(lens)
expected_page_indices = np.asarray([6, 7, 5])
expected_token_indices = np.asarray([3, 4, 9])
self.assertTrue(
np.array_equal(page_token_indices[0], expected_page_indices)
)
self.assertTrue(
np.array_equal(page_token_indices[1], expected_token_indices)
)
if __name__ == "__main__":
unittest.main()