test_sglang.py 14 KB
Newer Older
1
# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
# SPDX-License-Identifier: Apache-2.0

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

import pytest

Alec's avatar
Alec committed
11
12
from tests.serve.common import (
    SERVE_TEST_DIR,
13
    WORKSPACE_DIR,
Alec's avatar
Alec committed
14
15
16
    params_with_model_mark,
    run_serve_deployment,
)
17
from tests.utils.constants import DefaultPort
18
from tests.utils.engine_process import EngineConfig
19
from tests.utils.payload_builder import (
20
21
    anthropic_messages_payload_default,
    anthropic_messages_stream_payload_default,
22
23
24
    chat_payload,
    chat_payload_default,
    completion_payload_default,
25
26
    embedding_payload,
    embedding_payload_default,
27
    metric_payload_default,
28
29
    responses_payload_default,
    responses_stream_payload_default,
30
)
31
32
33
34
35

logger = logging.getLogger(__name__)


@dataclass
36
class SGLangConfig(EngineConfig):
37
38
    """Configuration for SGLang test scenarios"""

39
    stragglers: list[str] = field(default_factory=lambda: ["SGLANG:EngineCore"])
40
41


42
sglang_dir = os.environ.get("SGLANG_DIR") or os.path.join(
43
    WORKSPACE_DIR, "examples/backends/sglang"
44
)
45

46
47
48
# SGLang test configurations
# NOTE: pytest.mark.gpu_1 tests take ~167s (2m 47s) total to run sequentially (with models pre-cached)
# TODO: Parallelize these tests to reduce total execution time
49
50
sglang_configs = {
    "aggregated": SGLangConfig(
51
52
        # Uses backend agg.sh (with metrics enabled) for testing standard
        # aggregated deployment with metrics collection
53
        name="aggregated",
54
55
        directory=sglang_dir,
        script_name="agg.sh",
56
57
58
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.pre_merge,
59
            pytest.mark.timeout(240),  # 3x measured time (39s) + download time (120s)
60
        ],
61
        model="Qwen/Qwen3-0.6B",
62
        env={},
63
        frontend_port=DefaultPort.FRONTEND.value,
64
65
66
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
67
68
            responses_payload_default(),
            responses_stream_payload_default(),
69
            metric_payload_default(min_num_requests=6, backend="sglang"),
70
        ],
71
72
    ),
    "disaggregated": SGLangConfig(
73
74
75
        name="disaggregated",
        directory=sglang_dir,
        script_name="disagg.sh",
Dmitry Tokarev's avatar
Dmitry Tokarev committed
76
77
        marks=[
            pytest.mark.gpu_2,
78
            pytest.mark.pre_merge,
Dmitry Tokarev's avatar
Dmitry Tokarev committed
79
        ],
80
81
        model="Qwen/Qwen3-0.6B",
        env={},
82
        frontend_port=DefaultPort.FRONTEND.value,
83
84
85
86
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
        ],
87
    ),
88
89
    "disaggregated_same_gpu": SGLangConfig(
        # Uses disagg_same_gpu.sh for single-GPU disaggregated testing
90
91
        # Validates metrics from both prefill (DefaultPort.SYSTEM1) and decode
        # (DefaultPort.SYSTEM2) workers
92
93
94
        name="disaggregated_same_gpu",
        directory=sglang_dir,
        script_name="disagg_same_gpu.sh",
95
96
97
98
99
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.pre_merge,
            pytest.mark.skip(reason="unstable"),
        ],
100
101
        model="Qwen/Qwen3-0.6B",
        env={},
102
        frontend_port=DefaultPort.FRONTEND.value,
103
104
105
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
106
107
108
109
110
111
112
113
114
115
116
117
118
119
            # Validate dynamo_component_* and sglang:* metrics from prefill worker
            # (DefaultPort.SYSTEM1)
            metric_payload_default(
                min_num_requests=6,
                backend="sglang",
                port=DefaultPort.SYSTEM1.value,
            ),
            # Validate dynamo_component_* and sglang:* metrics from decode worker
            # (DefaultPort.SYSTEM2)
            metric_payload_default(
                min_num_requests=6,
                backend="sglang",
                port=DefaultPort.SYSTEM2.value,
            ),
120
121
        ],
    ),
122
    "kv_events": SGLangConfig(
123
124
125
        name="kv_events",
        directory=sglang_dir,
        script_name="agg_router.sh",
Dmitry Tokarev's avatar
Dmitry Tokarev committed
126
127
        marks=[
            pytest.mark.gpu_2,
128
            pytest.mark.pre_merge,
Dmitry Tokarev's avatar
Dmitry Tokarev committed
129
        ],
130
131
        model="Qwen/Qwen3-0.6B",
        env={
132
            "DYN_LOG": "dynamo_llm::kv_router::publisher=trace,dynamo_kv_router::scheduling::selector=info",
133
        },
134
        frontend_port=DefaultPort.FRONTEND.value,
135
136
137
        request_payloads=[
            chat_payload_default(
                expected_log=[
138
                    r"ZMQ listener .* received batch with \d+ events \(engine_seq=\d+(?:, [^)]*)?\)",
139
                    r"Event processor for worker_id \d+ processing event: Stored\(",
140
                    r"Selected worker: worker_type=\w+, worker_id=\d+ dp_rank=.*?, logit: ",
141
142
143
                ]
            )
        ],
144
    ),
145
146
147
148
149
    "template_verification": SGLangConfig(
        # Tests custom jinja template preprocessing by verifying the template
        # marker 'CUSTOM_TEMPLATE_ACTIVE|' is applied to user messages.
        # The backend (launch/template_verifier.*) checks for this marker
        # and returns "Successfully Applied Chat Template" if found.
150
151
        # Uses SERVE_TEST_DIR (not sglang_dir) because template_verifier.sh/.py
        # are test-specific mock scripts in tests/serve/launch/
152
        name="template_verification",
153
        directory=SERVE_TEST_DIR,  # special directory for test-specific scripts
154
        script_name="template_verifier.sh",
155
156
157
158
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.pre_merge,
            pytest.mark.nightly,
159
            pytest.mark.timeout(240),  # 3x measured time (20s) + download time (180s)
160
        ],
161
162
        model="Qwen/Qwen3-0.6B",
        env={},
163
        frontend_port=DefaultPort.FRONTEND.value,
164
165
166
167
168
169
        request_payloads=[
            chat_payload_default(
                expected_response=["Successfully Applied Chat Template"]
            )
        ],
    ),
170
    # NOTE: Pack all workers on 1 GPU for lower CI resource requirements
171
    "multimodal_epd_qwen": SGLangConfig(
172
        # E/P/D architecture: Encode, Prefill, Decode workers all on GPU 0
173
        name="multimodal_epd_qwen",
174
        directory=sglang_dir,
175
        script_name="multimodal_epd.sh",
176
177
178
        marks=[pytest.mark.gpu_1, pytest.mark.pre_merge],
        model="Qwen/Qwen3-VL-2B-Instruct",
        script_args=["--model", "Qwen/Qwen3-VL-2B-Instruct", "--single-gpu"],
179
        timeout=360,
180
181
182
183
184
        env={
            "DYN_ENCODE_WORKER_GPU": "0",
            "DYN_WORKER_GPU": "0",
            "DYN_ENCODE_GPU_MEM": "0.1",
            "DYN_WORKER_GPU_MEM": "0.4",
185
186
            # FIXME: NIXL Agent Initialization (shared memory interface) causes segfault
            "UCX_TLS": "^mm",
187
        },
188
        frontend_port=DefaultPort.FRONTEND.value,
189
190
191
192
193
194
195
196
197
198
199
200
        request_payloads=[
            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,
201
202
203
                # NOTE: The response text may mention 'bus', 'train', 'streetcar', etc.
                # so we need something consistently found in the response, or a different
                # approach to validation for this test to be stable.
204
                expected_response=["image"],
205
                temperature=0.0,
206
                max_tokens=100,
207
208
209
            )
        ],
    ),
210
211
212
213
214
215
216
217
218
219
220
221
222
    "multimodal_disagg_qwen": SGLangConfig(
        # E/P/D architecture: Encode, Prefill, Decode workers all on GPU 0
        name="multimodal_disagg_qwen",
        directory=sglang_dir,
        script_name="multimodal_disagg.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.pre_merge,
            pytest.mark.timeout(360),
        ],
        model="Qwen/Qwen3-VL-2B-Instruct",
        script_args=["--model", "Qwen/Qwen3-VL-2B-Instruct", "--single-gpu"],
        timeout=360,
223
224
225
226
        env={
            # FIXME: NIXL Agent Initialization (shared memory interface) causes segfault
            "UCX_TLS": "^mm",
        },
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
        frontend_port=DefaultPort.FRONTEND.value,
        request_payloads=[
            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=["image"],
                temperature=0.0,
                max_tokens=100,
            )
        ],
    ),
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
    "multimodal_agg_qwen": SGLangConfig(
        # Tests single-process aggregated multimodal inference using DecodeWorkerHandler
        # with in-process vision encoding (no separate encode worker)
        name="multimodal_agg_qwen",
        directory=sglang_dir,
        script_name="agg.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.pre_merge,
            pytest.mark.nightly,
            pytest.mark.timeout(300),
        ],
        model="Qwen/Qwen2.5-VL-7B-Instruct",
        script_args=[
            "--model-path",
            "Qwen/Qwen2.5-VL-7B-Instruct",
            "--chat-template",
            "qwen2-vl",
        ],
        delayed_start=0,
        timeout=360,
        frontend_port=DefaultPort.FRONTEND.value,
        request_payloads=[
            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=["image"],
                temperature=0.0,
                max_tokens=100,
            )
        ],
    ),
286
287
288
289
    "embedding_agg": SGLangConfig(
        name="embedding_agg",
        directory=sglang_dir,
        script_name="agg_embed.sh",
290
291
292
293
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.pre_merge,
            pytest.mark.nightly,
294
            pytest.mark.timeout(270),  # 3x measured time (29s) + download time (180s)
295
        ],
296
297
        model="Qwen/Qwen3-Embedding-4B",
        delayed_start=0,
298
        frontend_port=DefaultPort.FRONTEND.value,
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
        request_payloads=[
            # Test default payload with multiple inputs
            embedding_payload_default(
                repeat_count=2,
                expected_response=["Generated 2 embeddings with dimension"],
            ),
            # Test single string input
            embedding_payload(
                input_text="Hello, world!",
                repeat_count=1,
                expected_response=["Generated 1 embeddings with dimension"],
            ),
            # Test multiple string inputs
            embedding_payload(
                input_text=[
                    "The quick brown fox jumps over the lazy dog.",
                    "Machine learning is transforming technology.",
                    "Natural language processing enables computers to understand text.",
                ],
                repeat_count=1,
                expected_response=["Generated 3 embeddings with dimension"],
            ),
        ],
    ),
323
324
325
326
    "completions_only": SGLangConfig(
        name="completions_only",
        directory=sglang_dir,
        script_name="agg.sh",
327
328
        marks=[
            pytest.mark.gpu_1,
329
            pytest.mark.post_merge,
330
            pytest.mark.timeout(
331
332
                420
            ),  # Total test timeout: 2x measured average (79.36s) + download time (240s) for 7B model
333
        ],
334
335
336
337
338
339
340
341
342
343
344
        model="deepseek-ai/deepseek-llm-7b-base",
        script_args=[
            "--model-path",
            "deepseek-ai/deepseek-llm-7b-base",
            "--dyn-endpoint-types",
            "completions",
        ],
        request_payloads=[
            completion_payload_default(),
        ],
    ),
345
346
347
348
349
350
351
352
    "anthropic_messages": SGLangConfig(
        name="anthropic_messages",
        directory=sglang_dir,
        script_name="agg.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.post_merge,
            pytest.mark.timeout(240),
Dmitry Tokarev's avatar
Dmitry Tokarev committed
353
            pytest.mark.skip(reason="DYN-2261"),
354
355
356
357
358
359
360
361
362
        ],
        model="Qwen/Qwen3-0.6B",
        env={"DYN_ENABLE_ANTHROPIC_API": "1"},
        frontend_port=DefaultPort.FRONTEND.value,
        request_payloads=[
            anthropic_messages_payload_default(),
            anthropic_messages_stream_payload_default(),
        ],
    ),
363
364
365
}


Alec's avatar
Alec committed
366
@pytest.fixture(params=params_with_model_mark(sglang_configs))
367
368
369
370
371
372
373
def sglang_config_test(request):
    """Fixture that provides different SGLang test configurations"""
    return sglang_configs[request.param]


@pytest.mark.e2e
@pytest.mark.sglang
374
375
376
# Use 2 system ports because some `sglang_configs` validate metrics on multiple ports.
# This test iterates over all configs via `sglang_config_test`.
@pytest.mark.parametrize("num_system_ports", [2], indirect=True)
Alec's avatar
Alec committed
377
def test_sglang_deployment(
378
379
380
381
    sglang_config_test,
    request,
    runtime_services_dynamic_ports,
    dynamo_dynamic_ports,
382
    num_system_ports,
383
    predownload_models,
Alec's avatar
Alec committed
384
):
385
    """Test SGLang deployment scenarios using common helpers"""
386
387
388
    assert (
        num_system_ports >= 2
    ), "serve tests require at least SYSTEM_PORT1 + SYSTEM_PORT2"
389
390
391
392
    config = dataclasses.replace(
        sglang_config_test, frontend_port=dynamo_dynamic_ports.frontend_port
    )
    run_serve_deployment(config, request, ports=dynamo_dynamic_ports)
393
394


395
396
@pytest.mark.e2e
@pytest.mark.sglang
397
@pytest.mark.gpu_2
398
@pytest.mark.nightly
399
400
401
@pytest.mark.skip(
    reason="Requires 4 GPUs - enable when hardware is consistently available"
)
402
403
404
def test_sglang_disagg_dp_attention(
    request, runtime_services_dynamic_ports, dynamo_dynamic_ports, predownload_models
):
405
406
    """Test sglang disaggregated with DP attention (requires 4 GPUs)"""

407
    # Kept for reference; this test uses a different launch path and is skipped