test_hybrid.py 14.8 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

Mor Zusman's avatar
Mor Zusman committed
4
5
import pytest

6
from tests.models.registry import HF_EXAMPLE_MODELS
7
from tests.utils import multi_gpu_test
8
from vllm.engine.arg_utils import EngineArgs
9
from vllm.sampling_params import SamplingParams
10

11
12
from ...utils import check_logprobs_close, check_outputs_equal

13
14
15
# Mark all tests as hybrid
pytestmark = pytest.mark.hybrid_model

16
17
18
19
20
21
22
# NOTE: The first model in each list is taken as the primary model,
# meaning that it will be used in all tests in this file
# The rest of the models will only be tested by test_models

SSM_MODELS = [
    "state-spaces/mamba-130m-hf",
    "tiiuae/falcon-mamba-tiny-dev",
23
    "yujiepan/mamba2-codestral-v0.1-tiny-random",
24
]
25

26
27
HYBRID_MODELS = [
    "ai21labs/Jamba-tiny-dev",
28
29
    # skipping until vLLM implementation issues are resolved
    # "pfnet/plamo-2-1b",
30
    "Zyphra/Zamba2-1.2B-instruct",
31
    "hmellor/tiny-random-BambaForCausalLM",
32
33
    "ibm-granite/granite-4.0-tiny-preview",
    "tiiuae/Falcon-H1-0.5B-Base",
34
    "LiquidAI/LFM2-1.2B",
35
36
]

Chen Zhang's avatar
Chen Zhang committed
37
V1_SUPPORTED_MODELS = [
38
39
    "state-spaces/mamba-130m-hf",
    "ai21labs/Jamba-tiny-dev",
40
    "yujiepan/mamba2-codestral-v0.1-tiny-random",
41
    "Zyphra/Zamba2-1.2B-instruct",
42
    "hmellor/tiny-random-BambaForCausalLM",
43
44
    "ibm-granite/granite-4.0-tiny-preview",
    "tiiuae/Falcon-H1-0.5B-Base",
45
    "LiquidAI/LFM2-1.2B",
Chen Zhang's avatar
Chen Zhang committed
46
47
]

48
FULL_CUDA_GRAPH_MODELS = [
49
    "ai21labs/Jamba-tiny-dev",
50
    "Zyphra/Zamba2-1.2B-instruct",
51
52
]

53
54
55
56
V0_UNSUPPORTED_MODELS = [
    "LiquidAI/LFM2-1.2B",
]

57
58
59
60
61
FP32_STATE_MODELS = [
    "state-spaces/mamba-130m-hf",
    "Zyphra/Zamba2-1.2B-instruct",
]

62
63
64
# Avoid OOM
MAX_NUM_SEQS = 4

Mor Zusman's avatar
Mor Zusman committed
65

66
67
68
@pytest.mark.parametrize("model", SSM_MODELS + HYBRID_MODELS)
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
Mor Zusman's avatar
Mor Zusman committed
69
70
71
72
def test_models(
    hf_runner,
    vllm_runner,
    example_prompts,
Chen Zhang's avatar
Chen Zhang committed
73
    monkeypatch,
Mor Zusman's avatar
Mor Zusman committed
74
75
    model: str,
    max_tokens: int,
76
    num_logprobs: int,
Mor Zusman's avatar
Mor Zusman committed
77
) -> None:
78
79
80
81

    try:
        model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
        model_info.check_available_online(on_fail="skip")
82
        model_info.check_transformers_version(on_fail="skip")
83
    except ValueError:
84
        pass
85

86
    with hf_runner(model) as hf_model:
87
88
        hf_outputs = hf_model.generate_greedy_logprobs_limit(
            example_prompts, max_tokens, num_logprobs)
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
89

90
91
92
93
94
95
96
97
    with monkeypatch.context() as m:
        m.setenv("VLLM_USE_V1", "0")
        if model not in V0_UNSUPPORTED_MODELS:
            with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
                vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
                    example_prompts, max_tokens, num_logprobs)
        else:
            vllm_v0_outputs = None
98

Chen Zhang's avatar
Chen Zhang committed
99
    if model in V1_SUPPORTED_MODELS:
100
101
102
        with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
            vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
                example_prompts, max_tokens, num_logprobs)
Chen Zhang's avatar
Chen Zhang committed
103
104
105
    else:
        vllm_v1_outputs = None

106
    if vllm_v0_outputs is not None:
Chen Zhang's avatar
Chen Zhang committed
107
108
109
110
111
112
113
114
115
        check_logprobs_close(
            outputs_0_lst=hf_outputs,
            outputs_1_lst=vllm_v0_outputs,
            name_0="hf",
            name_1="vllm-v0",
        )

    if model in V1_SUPPORTED_MODELS:
        check_logprobs_close(
116
            outputs_0_lst=hf_outputs,
Chen Zhang's avatar
Chen Zhang committed
117
            outputs_1_lst=vllm_v1_outputs,
118
            name_0="hf",
Chen Zhang's avatar
Chen Zhang committed
119
120
            name_1="vllm-v1",
        )
Mor Zusman's avatar
Mor Zusman committed
121
122


123
@pytest.mark.parametrize("model", [SSM_MODELS[0], HYBRID_MODELS[0]])
124
125
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
126
127
128
129
130
def test_batching(
    vllm_runner,
    example_prompts,
    model: str,
    max_tokens: int,
131
    num_logprobs: int,
132
) -> None:
133
134
135
136
137
138
139
    try:
        model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
        model_info.check_available_online(on_fail="skip")
        model_info.check_transformers_version(on_fail="skip")
    except ValueError:
        pass

140
    for_loop_outputs = []
141
    with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
142
        for prompt in example_prompts:
143
144
145
146
            single_output, = vllm_model.generate_greedy_logprobs([prompt],
                                                                 max_tokens,
                                                                 num_logprobs)
            for_loop_outputs.append(single_output)
147

148
149
        batched_outputs = vllm_model.generate_greedy_logprobs(
            example_prompts, max_tokens, num_logprobs)
150

151
    check_logprobs_close(
152
153
154
155
156
157
158
        outputs_0_lst=for_loop_outputs,
        outputs_1_lst=batched_outputs,
        name_0="for_loop_vllm",
        name_1="batched_vllm",
    )


159
160
161
162
163
164
165
166
167
168
169
@pytest.mark.parametrize("model", [SSM_MODELS[0], HYBRID_MODELS[0]])
@pytest.mark.parametrize("max_tokens", [32])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("chunked_prefill_token_size", [1, 4, 16])
def test_chunked_prefill(
    vllm_runner,
    example_prompts,
    model: str,
    max_tokens: int,
    num_logprobs: int,
    chunked_prefill_token_size: int,
170
    monkeypatch,
171
172
173
) -> None:
    max_num_seqs = chunked_prefill_token_size
    max_num_batched_tokens = chunked_prefill_token_size
174

175
176
177
178
179
180
181
182
    with monkeypatch.context() as m:
        m.setenv("VLLM_USE_V1", "0")
        with vllm_runner(model,
                         enable_chunked_prefill=True,
                         max_num_batched_tokens=max_num_batched_tokens,
                         max_num_seqs=max_num_seqs) as vllm_model:
            chunked = vllm_model.generate_greedy_logprobs(
                example_prompts, max_tokens, num_logprobs)
183

184
185
186
187
188
        with vllm_runner(model,
                         enable_chunked_prefill=False,
                         max_num_seqs=max_num_seqs) as vllm_model:
            non_chunked = vllm_model.generate_greedy_logprobs(
                example_prompts, max_tokens, num_logprobs)
189

190
191
192
193
194
195
        check_logprobs_close(
            outputs_0_lst=chunked,
            outputs_1_lst=non_chunked,
            name_0="chunked",
            name_1="non_chunked",
        )
196
197


198
199
200
@pytest.mark.parametrize("model", [SSM_MODELS[0], HYBRID_MODELS[0]])
@pytest.mark.parametrize("max_tokens", [10])
def test_chunked_prefill_with_parallel_sampling(
201
202
203
204
205
    vllm_runner,
    example_prompts,
    model: str,
    max_tokens: int,
) -> None:
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
    """
    Tests chunked prefill in conjunction with n > 1. 
    
    In this case, prefill is populated with decoding tokens and
    we test that it doesn't fail.

    This test might fail if cache is not allocated correctly for n > 1
    decoding steps inside a chunked prefill forward pass
    (where we have both prefill and decode together)
    """
    sampling_params = SamplingParams(n=3,
                                     temperature=1,
                                     seed=0,
                                     max_tokens=max_tokens)
    with vllm_runner(
            model,
            enable_chunked_prefill=True,
            # forces prefill chunks with decoding
            max_num_batched_tokens=MAX_NUM_SEQS * 3,
            max_num_seqs=MAX_NUM_SEQS,
    ) as vllm_model:
        vllm_model.generate(example_prompts, sampling_params)
228
229


230
@pytest.mark.parametrize("model", [SSM_MODELS[0], HYBRID_MODELS[0]])
231
232
233
234
235
236
237
@pytest.mark.parametrize("max_tokens", [20])
def test_mamba_cache_cg_padding(
    vllm_runner,
    example_prompts,
    model: str,
    max_tokens: int,
) -> None:
238
239
240
241
242
    """
    This test is for verifying that mamba cache is padded to CG captured
    batch size. If it's not, a torch RuntimeError will be raised because
    tensor dimensions aren't compatible.
    """
Shinichi Hemmi's avatar
Shinichi Hemmi committed
243
244
    vllm_config = EngineArgs(model=model,
                             trust_remote_code=True).create_engine_config()
245
    while len(example_prompts) == vllm_config.pad_for_cudagraph(
246
            len(example_prompts)):
247
248
249
        example_prompts.append(example_prompts[0])

    try:
250
        with vllm_runner(model) as vllm_model:
251
252
253
254
255
256
257
258
            vllm_model.generate_greedy(example_prompts, max_tokens)
    except RuntimeError:
        pytest.fail(
            "Couldn't run batch size which is not equal to a Cuda Graph "
            "captured batch size. "
            "Could be related to mamba cache not padded correctly")


259
@pytest.mark.parametrize("model", [SSM_MODELS[0], HYBRID_MODELS[0]])
260
261
262
263
264
265
@pytest.mark.parametrize("max_tokens", [20])
def test_models_preemption_recompute(
    vllm_runner,
    example_prompts,
    model: str,
    max_tokens: int,
266
    monkeypatch,
267
) -> None:
268
269
270
    """
    Tests that outputs are identical with and w/o preemptions (recompute).
    """
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
    with monkeypatch.context() as m:
        m.setenv("VLLM_USE_V1", "0")
        with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
            scheduler = vllm_model.llm.llm_engine.scheduler[0]
            scheduler.ENABLE_ARTIFICIAL_PREEMPT = True
            preempt_vllm_outputs = vllm_model.generate_greedy(
                example_prompts, max_tokens)

            scheduler.ENABLE_ARTIFICIAL_PREEMPT = False
            vllm_outputs = vllm_model.generate_greedy(example_prompts,
                                                      max_tokens)

        check_outputs_equal(
            outputs_0_lst=preempt_vllm_outputs,
            outputs_1_lst=vllm_outputs,
            name_0="vllm_preepmtions",
            name_1="vllm",
        )
289
290


291
@pytest.mark.parametrize("model", [SSM_MODELS[0], HYBRID_MODELS[0]])
292
293
294
def test_fail_upon_inc_requests_and_finished_requests_lt_available_blocks(
    vllm_runner,
    example_prompts,
295
    model: str,
296
) -> None:
297
298
299
300
301
302
303
304
305
    """
    This test is for verifying that the hybrid inner state management doesn't
    collapse in case where the number of incoming requests and
    finished_requests_ids is larger than the maximum mamba block capacity.

    This could generally happen due to the fact that hybrid does support
    statelessness mechanism where it can cleanup new incoming requests in
    a single step.
    """
306
    try:
307
        with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
308
309
            vllm_model.generate_greedy([example_prompts[0]] * 100, 10)
    except ValueError:
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
310
        pytest.fail("Hybrid inner state wasn't cleaned up properly between"
311
312
313
                    "steps finished requests registered unnecessarily ")


314
@pytest.mark.parametrize("model", [SSM_MODELS[0], HYBRID_MODELS[0]])
Mor Zusman's avatar
Mor Zusman committed
315
316
317
def test_state_cleanup(
    vllm_runner,
    example_prompts,
318
    model: str,
Mor Zusman's avatar
Mor Zusman committed
319
) -> None:
320
321
322
323
324
325
    """ 
    This test is for verifying that the Hybrid state is cleaned up between
    steps.
    
    If its not cleaned, an error would be expected.
    """
Mor Zusman's avatar
Mor Zusman committed
326
    try:
327
        with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
Mor Zusman's avatar
Mor Zusman committed
328
329
330
            for _ in range(10):
                vllm_model.generate_greedy([example_prompts[0]] * 100, 1)
    except ValueError:
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
331
        pytest.fail("Hybrid inner state wasn't cleaned up between states, "
Mor Zusman's avatar
Mor Zusman committed
332
333
334
                    "could be related to finished_requests_ids")


335
@multi_gpu_test(num_gpus=2)
336
@pytest.mark.parametrize("model", [SSM_MODELS[0], HYBRID_MODELS[0]])
337
@pytest.mark.parametrize("max_tokens", [64])
338
339
@pytest.mark.parametrize("num_logprobs", [5])
def test_distributed_correctness(
340
341
342
343
    vllm_runner,
    example_prompts,
    model: str,
    max_tokens: int,
344
    num_logprobs: int,
345
) -> None:
346
    with vllm_runner(model, tensor_parallel_size=1,
347
                     max_num_seqs=2) as vllm_model:
348
349
        vllm_outputs_tp_1 = vllm_model.generate_greedy_logprobs(
            example_prompts, max_tokens, num_logprobs)
350

351
    with vllm_runner(model, tensor_parallel_size=2,
352
                     max_num_seqs=2) as vllm_model:
353
354
        vllm_outputs_tp_2 = vllm_model.generate_greedy_logprobs(
            example_prompts, max_tokens, num_logprobs)
355

356
    check_logprobs_close(
357
358
359
360
361
        outputs_0_lst=vllm_outputs_tp_1,
        outputs_1_lst=vllm_outputs_tp_2,
        name_0="vllm_tp_1",
        name_1="vllm_tp_2",
    )
362
363


364
@pytest.mark.parametrize("model", FULL_CUDA_GRAPH_MODELS)
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
def test_full_cuda_graph(
    hf_runner,
    vllm_runner,
    example_prompts,
    monkeypatch,
    model: str,
    max_tokens: int,
    num_logprobs: int,
) -> None:

    try:
        model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
        model_info.check_available_online(on_fail="skip")
        model_info.check_transformers_version(on_fail="skip")
    except ValueError:
        pass

    with hf_runner(model) as hf_model:
385
386
        hf_outputs = hf_model.generate_greedy_logprobs_limit(
            example_prompts, max_tokens, num_logprobs)
387
388

    with monkeypatch.context() as m:
389
390
391
392
393
394
395
396
397
398
399
        m.setenv("VLLM_USE_V1", "0")
        if model not in V0_UNSUPPORTED_MODELS:
            with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
                vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
                    example_prompts, max_tokens, num_logprobs)
        else:
            vllm_v0_outputs = None

    with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
        vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
            example_prompts, max_tokens, num_logprobs)
400

401
    if vllm_v0_outputs is not None:
402
403
404
405
406
407
408
409
        check_logprobs_close(
            outputs_0_lst=hf_outputs,
            outputs_1_lst=vllm_v0_outputs,
            name_0="hf",
            name_1="vllm-v0",
        )

    check_logprobs_close(
410
        outputs_0_lst=hf_outputs,
411
        outputs_1_lst=vllm_v1_outputs,
412
        name_0="hf",
413
414
        name_1="vllm-v1",
    )
415
416


417
@pytest.mark.parametrize("model", FP32_STATE_MODELS)
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
def test_fp32_state(
    hf_runner,
    vllm_runner,
    example_prompts,
    monkeypatch,
    model: str,
    max_tokens: int,
    num_logprobs: int,
) -> None:

    try:
        model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
        model_info.check_available_online(on_fail="skip")
        model_info.check_transformers_version(on_fail="skip")
    except ValueError:
        pass

    with hf_runner(model) as hf_model:
438
439
        hf_outputs = hf_model.generate_greedy_logprobs_limit(
            example_prompts, max_tokens, num_logprobs)
440
441

    with monkeypatch.context() as m:
442
        m.setenv("VLLM_USE_V1", "0")
443
444
        with vllm_runner(model,
                         max_num_seqs=MAX_NUM_SEQS,
445
446
                         mamba_ssm_cache_dtype="float32") as vllm_model:
            vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
447
448
                example_prompts, max_tokens, num_logprobs)

449
450
451
452
453
454
    with vllm_runner(model,
                     max_num_seqs=MAX_NUM_SEQS,
                     mamba_ssm_cache_dtype="float32") as vllm_model:
        vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
            example_prompts, max_tokens, num_logprobs)

455
456
457
458
459
460
    check_logprobs_close(
        outputs_0_lst=hf_outputs,
        outputs_1_lst=vllm_v0_outputs,
        name_0="hf",
        name_1="vllm-v0",
    )
461
462

    check_logprobs_close(
463
        outputs_0_lst=hf_outputs,
464
        outputs_1_lst=vllm_v1_outputs,
465
        name_0="hf",
466
467
        name_1="vllm-v1",
    )