Spaces:
Runtime error
Runtime error
File size: 12,375 Bytes
c7a96cd |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 |
import pytest
import torch
from copy import copy
from transformers import AutoTokenizer
from text_generation_server.pb import generate_pb2
from text_generation_server.models.causal_lm import CausalLMBatch
from text_generation_server.models.bloom import BloomCausalLMBatch, BLOOM
@pytest.fixture(scope="session")
def default_bloom():
return BLOOM("bigscience/bloom-560m")
@pytest.fixture(scope="session")
def bloom_560m_tokenizer():
return AutoTokenizer.from_pretrained("bigscience/bloom-560m", padding_side="left")
@pytest.fixture
def default_pb_request(default_pb_parameters, default_pb_stop_parameters):
return generate_pb2.Request(
id=0,
inputs="Test",
truncate=100,
parameters=default_pb_parameters,
stopping_parameters=default_pb_stop_parameters,
)
@pytest.fixture
def default_pb_batch(default_pb_request):
return generate_pb2.Batch(id=0, requests=[default_pb_request], size=1)
@pytest.fixture
def default_bloom_batch(default_pb_batch, bloom_560m_tokenizer):
return BloomCausalLMBatch.from_pb(
default_pb_batch, bloom_560m_tokenizer, torch.device("cpu")
)
@pytest.fixture
def default_multi_requests_bloom_batch(default_pb_request, bloom_560m_tokenizer):
req_0 = copy(default_pb_request)
req_0.id = 1
req_1 = default_pb_request
req_1.id = 2
req_1.stopping_parameters.max_new_tokens = 5
batch_pb = generate_pb2.Batch(id=0, requests=[req_0, req_1], size=2)
return BloomCausalLMBatch.from_pb(
batch_pb, bloom_560m_tokenizer, torch.device("cpu")
)
def test_batch_from_pb(default_pb_batch, default_bloom_batch):
batch = default_bloom_batch
assert batch.batch_id == default_pb_batch.id
assert batch.requests == default_pb_batch.requests
assert len(batch.input_ids) == default_pb_batch.size
assert batch.input_ids[0][-1] == 10264
assert torch.all(batch.input_ids[0][:-1] == 3)
assert batch.attention_mask[0][0] == 1
assert torch.all(batch.attention_mask[0][1:] == 0)
assert batch.past_key_values is None
assert all(
[
torch.equal(input_ids, all_input_ids[:, 0])
for input_ids, all_input_ids in zip(batch.input_ids, batch.all_input_ids)
]
)
assert batch.input_lengths == [1]
assert len(batch) == default_pb_batch.size
assert len(batch.next_token_choosers) == len(batch.stopping_criterias) == len(batch)
assert batch.max_input_length == batch.input_lengths[0]
def test_batch_concatenate_no_prefill(default_bloom_batch):
with pytest.raises(ValueError):
BloomCausalLMBatch.concatenate([default_bloom_batch, default_bloom_batch])
def test_causal_lm_batch_type(default_bloom):
assert default_bloom.batch_type == BloomCausalLMBatch
def test_causal_lm_generate_token(default_bloom, default_bloom_batch):
sequence_length = len(default_bloom_batch.all_input_ids[0])
generations, next_batch = default_bloom.generate_token(default_bloom_batch)
assert len(generations) == len(default_bloom_batch)
assert isinstance(next_batch, CausalLMBatch)
assert not next_batch.keys_head_dim_last
assert len(next_batch.all_input_ids) == len(next_batch)
assert len(next_batch.all_input_ids[0]) == sequence_length + 1
assert len(next_batch.attention_mask[0]) == 11
assert torch.all(next_batch.all_input_ids[0][-2:] == 10264)
assert torch.all(next_batch.all_input_ids[0][:-2] == 3)
assert torch.all(next_batch.attention_mask[0][:2] == 1)
assert torch.all(next_batch.attention_mask[0][2:] == 0)
assert next_batch.input_ids.shape == (len(next_batch), 1)
assert next_batch.input_ids[0, 0] == 10264
assert next_batch.input_lengths == [2]
assert next_batch.max_input_length == next_batch.input_lengths[0]
assert next_batch.past_key_values is not None
assert all(
[p[0].shape == (16, 64, sequence_length) for p in next_batch.past_key_values]
)
assert all(
[p[1].shape == (16, sequence_length, 64) for p in next_batch.past_key_values]
)
assert all([generation.generated_text is None for generation in generations])
assert all([len(generation.prefill_tokens) == 1 for generation in generations])
assert all([generation.token_id.item() == 10264 for generation in generations])
assert all([generation.token_text == "Test" for generation in generations])
assert generations[0].request_id == 0
def test_causal_lm_generate_token_completion(default_bloom, default_bloom_batch):
next_batch = default_bloom_batch
for _ in range(default_bloom_batch.stopping_criterias[0].max_new_tokens - 1):
generations, next_batch = default_bloom.generate_token(next_batch)
assert len(generations) == len(default_bloom_batch)
generations, next_batch = default_bloom.generate_token(next_batch)
assert next_batch is None
assert len(generations) == 1
assert (
generations[0].generated_text.text == "TestTestTestTestTestTestTestTestTestTest"
)
assert generations[0].request_id == default_bloom_batch.requests[0].id
assert (
generations[0].generated_text.generated_tokens
== default_bloom_batch.stopping_criterias[0].max_new_tokens
)
def test_causal_lm_generate_token_completion_multi(
default_bloom, default_multi_requests_bloom_batch
):
next_batch = default_multi_requests_bloom_batch
for i in range(
default_multi_requests_bloom_batch.stopping_criterias[1].max_new_tokens - 1
):
generations, next_batch = default_bloom.generate_token(next_batch)
assert len(generations) == len(default_multi_requests_bloom_batch)
generations, next_batch = default_bloom.generate_token(next_batch)
assert next_batch is not None
assert len(generations) == 2
assert generations[1].generated_text.text == "TestTestTestTestTest"
assert (
generations[1].request_id == default_multi_requests_bloom_batch.requests[1].id
)
assert (
generations[1].generated_text.generated_tokens
== default_multi_requests_bloom_batch.stopping_criterias[1].max_new_tokens
)
# Copy stopping_criterias before filtering
stopping_criterias = default_multi_requests_bloom_batch.stopping_criterias.copy()
next_batch = next_batch.filter([next_batch.requests[0]])
for _ in range(
stopping_criterias[0].max_new_tokens - stopping_criterias[1].max_new_tokens - 1
):
generations, next_batch = default_bloom.generate_token(next_batch)
assert len(generations) == len(next_batch)
generations, next_batch = default_bloom.generate_token(next_batch)
assert next_batch is None
assert len(generations) == 1
assert (
generations[0].generated_text.text == "TestTestTestTestTestTestTestTestTestTest"
)
assert (
generations[0].request_id == default_multi_requests_bloom_batch.requests[0].id
)
assert (
generations[0].generated_text.generated_tokens
== default_multi_requests_bloom_batch.stopping_criterias[0].max_new_tokens
)
def test_batch_concatenate(
default_bloom, default_bloom_batch, default_multi_requests_bloom_batch
):
next_batch_0 = default_bloom_batch
_, next_batch_0 = default_bloom.generate_token(next_batch_0)
_, next_batch_0 = default_bloom.generate_token(next_batch_0)
next_batch_1 = default_multi_requests_bloom_batch
_, next_batch_1 = default_bloom.generate_token(next_batch_1)
# Clone past_key_values before concatenating to compare after,
# because they are removed from the concatenated batches
next_batch_0_past_key_values = [
(k.clone(), v.clone()) for (k, v) in next_batch_0.past_key_values
]
next_batch_1_past_key_values = [
(k.clone(), v.clone()) for (k, v) in next_batch_1.past_key_values
]
next_batch = BloomCausalLMBatch.concatenate([next_batch_0, next_batch_1])
assert torch.equal(next_batch.all_input_ids[0], next_batch_0.all_input_ids[0])
assert torch.equal(next_batch.all_input_ids[1], next_batch_1.all_input_ids[0])
assert torch.equal(next_batch.all_input_ids[2], next_batch_1.all_input_ids[1])
assert torch.all(
next_batch.attention_mask[0, : -next_batch.padding_right_offset] == 1
)
assert torch.all(
next_batch.attention_mask[1:, 1 : -next_batch.padding_right_offset] == 1
)
assert torch.all(next_batch.attention_mask[1:, 3:] == 0)
assert next_batch.batch_id == 0
assert torch.all(next_batch.input_ids == 10264)
assert next_batch.input_lengths == [3, 2, 2]
assert next_batch.max_input_length == 3
assert next_batch.requests[0] == next_batch_0.requests[0]
assert next_batch.requests[1:] == next_batch_1.requests
assert next_batch.next_token_choosers[0] == next_batch_0.next_token_choosers[0]
assert next_batch.next_token_choosers[1:] == next_batch_1.next_token_choosers
assert next_batch.stopping_criterias[0] == next_batch_0.stopping_criterias[0]
assert next_batch.stopping_criterias[1:] == next_batch_1.stopping_criterias
assert next_batch.past_key_values is not None
assert all([p[0].shape == (3, 16, 64, 2) for p in next_batch.past_key_values])
assert all([p[1].shape == (3, 16, 2, 64) for p in next_batch.past_key_values])
for i, past in enumerate(next_batch.past_key_values):
assert torch.equal(next_batch_0_past_key_values[i][0][:, :, -2:], past[0][0])
assert torch.equal(
next_batch_1_past_key_values[i][0][:, :, -1:],
past[0][1:, :, :, -1].reshape(-1, 64, 1),
)
assert torch.equal(next_batch_0_past_key_values[i][1][:, -2:, :], past[1][0])
assert torch.equal(
next_batch_1_past_key_values[i][1][:, -1:, :],
past[1][1:, :, -1, :].reshape(-1, 1, 64),
)
for _ in range(
default_multi_requests_bloom_batch.stopping_criterias[1].max_new_tokens - 2
):
generations, next_batch = default_bloom.generate_token(next_batch)
assert len(generations) == len(next_batch)
generations, next_batch = default_bloom.generate_token(next_batch)
assert next_batch is not None
assert len(generations) == 3
assert generations[2].generated_text.text == "TestTestTestTestTest"
assert (
generations[2].request_id == default_multi_requests_bloom_batch.requests[1].id
)
assert (
generations[2].generated_text.generated_tokens
== default_multi_requests_bloom_batch.stopping_criterias[1].max_new_tokens
)
next_batch = next_batch.filter([next_batch.requests[0], next_batch.requests[1]])
for _ in range(
default_bloom_batch.stopping_criterias[0].max_new_tokens
- default_multi_requests_bloom_batch.stopping_criterias[1].max_new_tokens
- 2
):
generations, next_batch = default_bloom.generate_token(next_batch)
assert len(generations) == len(next_batch)
generations, next_batch = default_bloom.generate_token(next_batch)
assert next_batch is not None
assert len(generations) == 2
assert (
generations[0].generated_text.text == "TestTestTestTestTestTestTestTestTestTest"
)
assert generations[0].request_id == default_bloom_batch.requests[0].id
assert (
generations[0].generated_text.generated_tokens
== default_bloom_batch.stopping_criterias[0].max_new_tokens
)
next_batch = next_batch.filter([next_batch.requests[1]])
for _ in range(
default_multi_requests_bloom_batch.stopping_criterias[0].max_new_tokens
- default_bloom_batch.stopping_criterias[0].max_new_tokens
- default_multi_requests_bloom_batch.stopping_criterias[1].max_new_tokens
- 4
):
generations, next_batch = default_bloom.generate_token(next_batch)
assert len(generations) == len(next_batch)
generations, next_batch = default_bloom.generate_token(next_batch)
assert next_batch is None
assert len(generations) == 1
assert (
generations[0].generated_text.text == "TestTestTestTestTestTestTestTestTestTest"
)
assert (
generations[0].request_id == default_multi_requests_bloom_batch.requests[0].id
)
assert (
generations[0].generated_text.generated_tokens
== default_multi_requests_bloom_batch.stopping_criterias[0].max_new_tokens
)
|