test_vllm.py 16.8 KB
Newer Older
1
2
3
# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

4
import base64
5
6
import logging
import os
7
import random
8
from dataclasses import dataclass, field
9
10
11

import pytest

12
13
14
15
16
from tests.serve.common import (
    WORKSPACE_DIR,
    params_with_model_mark,
    run_serve_deployment,
)
17
from tests.serve.conftest import MULTIMODAL_IMG_PATH, MULTIMODAL_IMG_URL
18
19
20
21
22
23
from tests.utils.engine_process import EngineConfig
from tests.utils.payload_builder import (
    chat_payload,
    chat_payload_default,
    completion_payload_default,
    metric_payload_default,
24
)
25
from tests.utils.payloads import ToolCallingChatPayload
26
27
28
29
30

logger = logging.getLogger(__name__)


@dataclass
31
class VLLMConfig(EngineConfig):
32
33
    """Configuration for vLLM test scenarios"""

34
    stragglers: list[str] = field(default_factory=lambda: ["VLLM:EngineCore"])
35
36


37
vllm_dir = os.environ.get("VLLM_DIR") or os.path.join(
38
    WORKSPACE_DIR, "examples/backends/vllm"
39
)
40

41

42
43
44
45
# vLLM test configurations
vllm_configs = {
    "aggregated": VLLMConfig(
        name="aggregated",
46
        directory=vllm_dir,
47
        script_name="agg.sh",
48
        marks=[pytest.mark.gpu_1, pytest.mark.pre_merge],
49
        model="Qwen/Qwen3-0.6B",
50
51
52
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
53
            metric_payload_default(min_num_requests=6, backend="vllm"),
54
        ],
55
    ),
56
57
58
59
    "aggregated_lmcache": VLLMConfig(
        name="aggregated_lmcache",
        directory=vllm_dir,
        script_name="agg_lmcache.sh",
60
        marks=[pytest.mark.gpu_1, pytest.mark.pre_merge],
61
62
63
64
65
66
67
68
        model="Qwen/Qwen3-0.6B",
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
            metric_payload_default(min_num_requests=6, backend="vllm"),
            metric_payload_default(min_num_requests=6, backend="lmcache"),
        ],
    ),
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
    "aggregated_lmcache_multiproc": VLLMConfig(
        name="aggregated_lmcache_multiproc",
        directory=vllm_dir,
        script_name="agg_lmcache_multiproc.sh",
        marks=[pytest.mark.gpu_1],
        model="Qwen/Qwen3-0.6B",
        env={
            "PROMETHEUS_MULTIPROC_DIR": f"/tmp/prometheus_multiproc_test_{os.getpid()}_{random.randint(0, 10000)}"
        },
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
            metric_payload_default(min_num_requests=6, backend="vllm"),
            metric_payload_default(min_num_requests=6, backend="lmcache"),
        ],
    ),
85
86
87
88
    "agg-request-plane-tcp": VLLMConfig(
        name="agg-request-plane-tcp",
        directory=vllm_dir,
        script_name="agg_request_planes.sh",
89
        marks=[pytest.mark.gpu_1, pytest.mark.pre_merge],
90
91
92
93
94
95
96
97
98
99
100
        model="Qwen/Qwen3-0.6B",
        script_args=["--tcp"],
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
        ],
    ),
    "agg-request-plane-http": VLLMConfig(
        name="agg-request-plane-http",
        directory=vllm_dir,
        script_name="agg_request_planes.sh",
101
        marks=[pytest.mark.gpu_1, pytest.mark.pre_merge],
102
103
104
105
106
107
108
        model="Qwen/Qwen3-0.6B",
        script_args=["--http"],
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
        ],
    ),
109
110
    "agg-router": VLLMConfig(
        name="agg-router",
111
        directory=vllm_dir,
112
        script_name="agg_router.sh",
113
        marks=[pytest.mark.gpu_2, pytest.mark.post_merge],
114
        model="Qwen/Qwen3-0.6B",
115
116
117
        request_payloads=[
            chat_payload_default(
                expected_log=[
118
                    r"ZMQ listener .* received batch with \d+ events \(seq=\d+(?:, [^)]*)?\)",
119
                    r"Event processor for worker_id \d+ processing event: Stored\(",
Yan Ru Pei's avatar
Yan Ru Pei committed
120
                    r"Selected worker: worker_id=\d+ dp_rank=.*?, logit: ",
121
122
123
124
125
126
                ]
            )
        ],
        env={
            "DYN_LOG": "dynamo_llm::kv_router::publisher=trace,dynamo_llm::kv_router::scheduler=info",
        },
127
    ),
128
129
    "disaggregated": VLLMConfig(
        name="disaggregated",
130
        directory=vllm_dir,
131
        script_name="disagg.sh",
132
        marks=[pytest.mark.gpu_2, pytest.mark.post_merge],
133
        model="Qwen/Qwen3-0.6B",
134
135
136
137
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
        ],
138
    ),
139
140
    "deepep": VLLMConfig(
        name="deepep",
141
        directory=vllm_dir,
142
        script_name="dsr1_dep.sh",
143
144
145
146
        marks=[
            pytest.mark.gpu_2,
            pytest.mark.vllm,
            pytest.mark.h100,
147
            pytest.mark.nightly,
148
        ],
149
        model="deepseek-ai/DeepSeek-V2-Lite",
150
        script_args=[
151
152
153
154
155
156
157
158
159
            "--model",
            "deepseek-ai/DeepSeek-V2-Lite",
            "--num-nodes",
            "1",
            "--node-rank",
            "0",
            "--gpus-per-node",
            "2",
        ],
160
        timeout=700,
161
        request_payloads=[
162
163
            chat_payload_default(),
            completion_payload_default(),
164
        ],
165
    ),
166
167
    "multimodal_agg_llava_epd": VLLMConfig(
        name="multimodal_agg_llava_epd",
168
        directory=vllm_dir,
169
        script_name="agg_multimodal_epd.sh",
170
        marks=[pytest.mark.gpu_2, pytest.mark.nightly],
171
        model="llava-hf/llava-1.5-7b-hf",
172
173
174
175
        script_args=["--model", "llava-hf/llava-1.5-7b-hf"],
        request_payloads=[
            chat_payload(
                [
176
177
178
179
                    {
                        "type": "text",
                        "text": "What colors are in the following image? Respond only with the colors.",
                    },
180
181
                    {
                        "type": "image_url",
182
                        "image_url": {"url": MULTIMODAL_IMG_URL},
183
184
185
                    },
                ],
                repeat_count=1,
186
                expected_response=["purple"],
187
                temperature=0.0,
188
                max_tokens=100,
189
190
            )
        ],
191
    ),
192
193
194
195
    "multimodal_agg_qwen_epd": VLLMConfig(
        name="multimodal_agg_qwen_epd",
        directory=vllm_dir,
        script_name="agg_multimodal_epd.sh",
196
        marks=[pytest.mark.gpu_2, pytest.mark.nightly],
197
198
199
200
201
202
203
        model="Qwen/Qwen2.5-VL-7B-Instruct",
        delayed_start=0,
        script_args=["--model", "Qwen/Qwen2.5-VL-7B-Instruct"],
        timeout=360,
        request_payloads=[
            chat_payload(
                [
204
205
206
207
                    {
                        "type": "text",
                        "text": "What colors are in the following image? Respond only with the colors.",
                    },
208
209
                    {
                        "type": "image_url",
210
                        "image_url": {"url": MULTIMODAL_IMG_URL},
211
212
213
                    },
                ],
                repeat_count=1,
214
215
                expected_response=["purple"],
                max_tokens=100,
216
217
218
            )
        ],
    ),
219
220
    "multimodal_agg_qwen": VLLMConfig(
        name="multimodal_agg_qwen",
221
222
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
223
        marks=[pytest.mark.gpu_2, pytest.mark.nightly],
224
        model="Qwen/Qwen2.5-VL-7B-Instruct",
225
        script_args=["--model", "Qwen/Qwen2.5-VL-7B-Instruct"],
226
        delayed_start=0,
227
        timeout=360,
228
229
230
231
        request_payloads=[
            chat_payload(
                [
                    {
232
233
                        "type": "text",
                        "text": "What colors are in the following image? Respond only with the colors.",
234
                    },
235
236
                    {
                        "type": "image_url",
237
                        "image_url": {"url": MULTIMODAL_IMG_URL},
238
239
240
                    },
                ],
                repeat_count=1,
241
242
                expected_response=["purple"],
                max_tokens=100,
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
    "multimodal_agg_llava": VLLMConfig(
        name="multimodal_agg_llava",
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
        marks=[
            pytest.mark.gpu_2,
            # https://github.com/ai-dynamo/dynamo/issues/4501
            pytest.mark.xfail(strict=False),
        ],
        model="llava-hf/llava-1.5-7b-hf",
        script_args=["--model", "llava-hf/llava-1.5-7b-hf"],
        delayed_start=0,
        timeout=360,
        request_payloads=[
            # HTTP URL test
            chat_payload(
                [
                    {"type": "text", "text": "What is in this image?"},
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": "http://images.cocodataset.org/test2017/000000155781.jpg"
                        },
                    },
                ],
                repeat_count=1,
                expected_response=["bus"],
                temperature=0.0,
            ),
            # String content test - verifies string → array conversion for multimodal templates
            chat_payload_default(
                repeat_count=1,
                expected_response=[],  # Just validate no error
            ),
        ],
    ),
282
    # TODO: Update this test case when we have video multimodal support in vllm official components
283
284
    "multimodal_video_agg": VLLMConfig(
        name="multimodal_video_agg",
285
        directory=os.path.join(WORKSPACE_DIR, "examples/multimodal"),
286
        script_name="video_agg.sh",
287
        marks=[pytest.mark.gpu_2, pytest.mark.nightly],
288
289
        model="llava-hf/LLaVA-NeXT-Video-7B-hf",
        delayed_start=0,
290
        script_args=["--model", "llava-hf/LLaVA-NeXT-Video-7B-hf"],
291
        timeout=360,
292
293
294
295
296
297
298
299
300
301
302
303
304
        request_payloads=[
            chat_payload(
                [
                    {"type": "text", "text": "Describe the video in detail"},
                    {
                        "type": "video_url",
                        "video_url": {
                            "url": "https://storage.googleapis.com/gtv-videos-bucket/sample/BigBuckBunny.mp4"
                        },
                    },
                ],
                repeat_count=1,
                expected_response=["rabbit"],
305
                temperature=0.7,
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
    "multimodal_audio_agg": VLLMConfig(
        name="multimodal_audio_agg",
        directory="/workspace/examples/multimodal",
        script_name="audio_agg.sh",
        marks=[pytest.mark.gpu_2],
        model="Qwen/Qwen2-Audio-7B-Instruct",
        delayed_start=0,
        script_args=["--model", "Qwen/Qwen2-Audio-7B-Instruct"],
        timeout=500,
        request_payloads=[
            chat_payload(
                [
                    {"type": "text", "text": "What is recited in the audio?"},
                    {
                        "type": "audio_url",
                        "audio_url": {
                            "url": "https://raw.githubusercontent.com/yuekaizhang/Triton-ASR-Client/main/datasets/mini_en/wav/1221-135766-0002.wav"
                        },
                    },
                ],
                repeat_count=1,
                expected_response=[
                    "The original content of this audio is:'yet these thoughts affected Hester Pynne less with hope than apprehension.'"
                ],
                temperature=0.8,
            )
        ],
    ),
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
    "aggregated_toolcalling": VLLMConfig(
        name="aggregated_toolcalling",
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
        marks=[pytest.mark.gpu_2, pytest.mark.multimodal],
        model="Qwen/Qwen3-VL-30B-A3B-Instruct-FP8",
        script_args=[
            "--model",
            "Qwen/Qwen3-VL-30B-A3B-Instruct-FP8",
            "--max-model-len",
            "10000",
            "--dyn-tool-call-parser",
            "hermes",
        ],
        delayed_start=0,
        timeout=600,
        request_payloads=[
            ToolCallingChatPayload(
                body={
                    "messages": [
                        {
                            "role": "user",
                            "content": [
                                {
                                    "type": "text",
                                    "text": "Describe what you see in this image in detail.",
                                },
                                {
                                    "type": "image_url",
                                    "image_url": {"url": MULTIMODAL_IMG_URL},
                                },
                            ],
                        }
                    ],
                    "tools": [
                        {
                            "type": "function",
                            "function": {
                                "name": "describe_image",
                                "description": "Provides detailed description of objects and scenes in an image",
                                "parameters": {
                                    "type": "object",
                                    "properties": {
                                        "objects": {
                                            "type": "array",
                                            "items": {"type": "string"},
                                            "description": "List of objects detected in the image",
                                        },
                                        "scene": {
                                            "type": "string",
                                            "description": "Overall scene description",
                                        },
                                    },
                                    "required": ["objects", "scene"],
                                },
                            },
                        }
                    ],
                    "tool_choice": "auto",
                    "max_tokens": 1024,
                },
                repeat_count=1,
                expected_response=["purple"],  # Validate image understanding
                expected_log=[],
                expected_tool_name="describe_image",  # Validate tool call happened
            )
        ],
    ),
405
406
407
    # TODO: Enable this test case when we have 4 GPUs runners.
    # "multimodal_disagg": VLLMConfig(
    #     name="multimodal_disagg",
408
    #     directory=os.path.join(WORKSPACE_DIR, "examples/multimodal"),
409
410
411
412
    #     script_name="disagg.sh",
    #     marks=[pytest.mark.gpu_4, pytest.mark.vllm],
    #     model="llava-hf/llava-1.5-7b-hf",
    #     delayed_start=45,
413
    #     script_args=["--model", "llava-hf/llava-1.5-7b-hf"],
414
    # ),
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
    "completions_only": VLLMConfig(
        name="completions_only",
        directory=vllm_dir,
        script_name="agg.sh",
        marks=[pytest.mark.gpu_1],
        model="deepseek-ai/deepseek-llm-7b-base",
        script_args=[
            "--model",
            "deepseek-ai/deepseek-llm-7b-base",
            "--dyn-endpoint-types",
            "completions",
        ],
        request_payloads=[
            completion_payload_default(),
        ],
    ),
431
432
433
}


Alec's avatar
Alec committed
434
@pytest.fixture(params=params_with_model_mark(vllm_configs))
435
436
437
438
439
def vllm_config_test(request):
    """Fixture that provides different vLLM test configurations"""
    return vllm_configs[request.param]


440
@pytest.mark.vllm
441
@pytest.mark.e2e
442
@pytest.mark.nightly
Alec's avatar
Alec committed
443
def test_serve_deployment(
444
    vllm_config_test, request, runtime_services, predownload_models, image_server
Alec's avatar
Alec committed
445
):
446
447
448
449
    """
    Test dynamo serve deployments with different graph configurations.
    """
    config = vllm_config_test
450
    run_serve_deployment(config, request)
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497


@pytest.mark.vllm
@pytest.mark.e2e
@pytest.mark.gpu_2
def test_multimodal_b64(request, runtime_services, predownload_models):
    """
    Test multimodal inference with base64 url passthrough.

    This test is separate because it loads the required image at runtime
    (not collection time), ensuring it only fails when actually executed.
    """
    # Load B64 image at test execution time
    with open(MULTIMODAL_IMG_PATH, "rb") as f:
        b64_img = base64.b64encode(f.read()).decode()

    # Create payload with B64 image
    b64_payload = chat_payload(
        [
            {
                "type": "text",
                "text": "What colors are in the following image? Respond only with the colors.",
            },
            {
                "type": "image_url",
                "image_url": {"url": f"data:image/png;base64,{b64_img}"},
            },
        ],
        repeat_count=1,
        expected_response=["purple"],
        max_tokens=100,
    )

    # Create test config
    config = VLLMConfig(
        name="test_multimodal_b64",
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
        marks=[],  # markers at function-level
        model="Qwen/Qwen2.5-VL-7B-Instruct",
        script_args=["--model", "Qwen/Qwen2.5-VL-7B-Instruct"],
        delayed_start=0,
        timeout=360,
        request_payloads=[b64_payload],
    )

    run_serve_deployment(config, request)