test_disaggregation_different_tp.py 9.05 KB
Newer Older
1
2
3
4
5
import os
import subprocess
import time
import unittest
from types import SimpleNamespace
6
from urllib.parse import urlparse
7
8
9
10
11
12
13
14
15
16
17

import requests

from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
    DEFAULT_MODEL_NAME_FOR_TEST_MLA,
    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
    DEFAULT_URL_FOR_TEST,
    CustomTestCase,
    popen_launch_pd_server,
18
    popen_with_error_check,
19
20
21
)


22
class TestDisaggregationMooncakePrefillLargerTP(CustomTestCase):
23
24
25
26
27
28
29
    @classmethod
    def setUpClass(cls):
        # Temporarily disable JIT DeepGEMM
        cls.original_jit_deepgemm = os.environ.get("SGL_ENABLE_JIT_DEEPGEMM")
        os.environ["SGL_ENABLE_JIT_DEEPGEMM"] = "false"

        cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
30
31
32
33
34
35
36
37
38
39
        parsed_url = urlparse(DEFAULT_URL_FOR_TEST)
        cls.base_host = parsed_url.hostname
        base_port = str(parsed_url.port)
        cls.lb_port = base_port
        cls.prefill_port = f"{int(base_port) + 100}"
        cls.decode_port = f"{int(base_port) + 200}"
        cls.prefill_url = f"http://{cls.base_host}:{cls.prefill_port}"
        cls.decode_url = f"http://{cls.base_host}:{cls.decode_port}"
        cls.lb_url = f"http://{cls.base_host}:{cls.lb_port}"
        print(f"{cls.base_host=} {cls.lb_port=} {cls.prefill_port=} {cls.decode_port=}")
40

41
42
43
        # Non blocking start servers
        cls.start_prefill()
        cls.start_decode()
44

45
        # Block until both
46
47
48
49
50
51
        cls.wait_server_ready(cls.prefill_url + "/health")
        cls.wait_server_ready(cls.decode_url + "/health")

        lb_command = [
            "python3",
            "-m",
52
53
54
            "sglang_router.launch_router",
            "--pd-disaggregation",
            "--mini-lb",  # FIXME: remove this
55
56
57
58
59
60
61
            "--prefill",
            cls.prefill_url,
            "--decode",
            cls.decode_url,
            "--host",
            cls.base_host,
            "--port",
62
            cls.lb_port,
63
64
65
        ]

        print("Starting load balancer:", " ".join(lb_command))
66
        cls.process_lb = popen_with_error_check(lb_command)
67
68
69
70
71
72
73
74
75
        cls.wait_server_ready(cls.lb_url + "/health")

    @classmethod
    def start_prefill(cls):
        prefill_args = [
            "--trust-remote-code",
            "--disaggregation-mode",
            "prefill",
            "--tp",
76
77
78
            "2",
            "--disaggregation-ib-device",
            "mlx5_roce0,mlx5_roce1",
79
80
81
82
83
84
85
86
87
88
89
90
91
92
        ]
        cls.process_prefill = popen_launch_pd_server(
            cls.model,
            cls.prefill_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=prefill_args,
        )

    @classmethod
    def start_decode(cls):
        decode_args = [
            "--trust-remote-code",
            "--disaggregation-mode",
            "decode",
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
            "--tp",
            "1",
            "--base-gpu-id",
            "2",
            "--disaggregation-ib-device",
            "mlx5_roce2",
        ]
        cls.process_decode = popen_launch_pd_server(
            cls.model,
            cls.decode_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=decode_args,
        )

    @classmethod
    def wait_server_ready(cls, url, timeout=60):
        start_time = time.perf_counter()
        while True:
            try:
                response = requests.get(url)
                if response.status_code == 200:
                    print(f"Server {url} is ready")
                    return
            except Exception:
                pass

            if time.perf_counter() - start_time > timeout:
                raise RuntimeError(f"Server {url} failed to start in {timeout}s")
            time.sleep(1)

    @classmethod
    def tearDownClass(cls):
        # Restore JIT DeepGEMM environment variable
        if cls.original_jit_deepgemm is not None:
            os.environ["SGL_ENABLE_JIT_DEEPGEMM"] = cls.original_jit_deepgemm
        else:
            os.environ.pop("SGL_ENABLE_JIT_DEEPGEMM", None)

        for process in [cls.process_lb, cls.process_decode, cls.process_prefill]:
            if process:
                try:
                    kill_process_tree(process.pid)
                except Exception as e:
                    print(f"Error killing process {process.pid}: {e}")
        # wait for 5 seconds
        time.sleep(5)

    def test_gsm8k(self):
        args = SimpleNamespace(
            num_shots=5,
            data_path=None,
            num_questions=200,
            max_new_tokens=512,
            parallel=128,
            host=f"http://{self.base_host}",
            port=int(self.lb_port),
        )
        metrics = run_eval_few_shot_gsm8k(args)
        print(f"Evaluation metrics: {metrics}")

        self.assertGreater(metrics["accuracy"], 0.60)


class TestDisaggregationMooncakeDecodeLargerTP(CustomTestCase):
    @classmethod
    def setUpClass(cls):
        # Temporarily disable JIT DeepGEMM
        cls.original_jit_deepgemm = os.environ.get("SGL_ENABLE_JIT_DEEPGEMM")
        os.environ["SGL_ENABLE_JIT_DEEPGEMM"] = "false"

        cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
        parsed_url = urlparse(DEFAULT_URL_FOR_TEST)
        cls.base_host = parsed_url.hostname
        base_port = str(parsed_url.port)
        cls.lb_port = base_port
        cls.prefill_port = f"{int(base_port) + 100}"
        cls.decode_port = f"{int(base_port) + 200}"
        cls.prefill_url = f"http://{cls.base_host}:{cls.prefill_port}"
        cls.decode_url = f"http://{cls.base_host}:{cls.decode_port}"
        cls.lb_url = f"http://{cls.base_host}:{cls.lb_port}"
        print(f"{cls.base_host=} {cls.lb_port=} {cls.prefill_port=} {cls.decode_port=}")

        # Non blocking start servers
        cls.start_prefill()
        cls.start_decode()

        # Block until both
        cls.wait_server_ready(cls.prefill_url + "/health")
        cls.wait_server_ready(cls.decode_url + "/health")

        lb_command = [
            "python3",
            "-m",
186
187
188
            "sglang_router.launch_router",
            "--pd-disaggregation",
            "--mini-lb",  # FIXME: remove this
189
190
191
192
            "--prefill",
            cls.prefill_url,
            "--decode",
            cls.decode_url,
193
194
195
            "--host",
            cls.base_host,
            "--port",
196
197
198
199
            cls.lb_port,
        ]

        print("Starting load balancer:", " ".join(lb_command))
200
        cls.process_lb = popen_with_error_check(lb_command)
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
        cls.wait_server_ready(cls.lb_url + "/health")

    @classmethod
    def start_prefill(cls):
        prefill_args = [
            "--trust-remote-code",
            "--disaggregation-mode",
            "prefill",
            "--tp",
            "1",
            "--disaggregation-ib-device",
            "mlx5_roce0",
        ]
        cls.process_prefill = popen_launch_pd_server(
            cls.model,
            cls.prefill_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=prefill_args,
        )

    @classmethod
    def start_decode(cls):
        decode_args = [
            "--trust-remote-code",
            "--disaggregation-mode",
            "decode",
227
228
229
            "--tp",
            "2",
            "--base-gpu-id",
230
231
232
            "1",
            "--disaggregation-ib-device",
            "mlx5_roce1,mlx5_roce2",
233
234
235
236
237
238
239
240
241
242
        ]
        cls.process_decode = popen_launch_pd_server(
            cls.model,
            cls.decode_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=decode_args,
        )

    @classmethod
    def wait_server_ready(cls, url, timeout=60):
243
        start_time = time.perf_counter()
244
245
246
247
248
249
250
251
252
        while True:
            try:
                response = requests.get(url)
                if response.status_code == 200:
                    print(f"Server {url} is ready")
                    return
            except Exception:
                pass

253
            if time.perf_counter() - start_time > timeout:
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
                raise RuntimeError(f"Server {url} failed to start in {timeout}s")
            time.sleep(1)

    @classmethod
    def tearDownClass(cls):
        # Restore JIT DeepGEMM environment variable
        if cls.original_jit_deepgemm is not None:
            os.environ["SGL_ENABLE_JIT_DEEPGEMM"] = cls.original_jit_deepgemm
        else:
            os.environ.pop("SGL_ENABLE_JIT_DEEPGEMM", None)

        for process in [cls.process_lb, cls.process_decode, cls.process_prefill]:
            if process:
                try:
                    kill_process_tree(process.pid)
                except Exception as e:
                    print(f"Error killing process {process.pid}: {e}")
271
272
        # wait for 5 seconds
        time.sleep(5)
273
274
275
276
277
278
279
280

    def test_gsm8k(self):
        args = SimpleNamespace(
            num_shots=5,
            data_path=None,
            num_questions=200,
            max_new_tokens=512,
            parallel=128,
281
282
            host=f"http://{self.base_host}",
            port=int(self.lb_port),
283
284
285
286
287
288
289
290
291
        )
        metrics = run_eval_few_shot_gsm8k(args)
        print(f"Evaluation metrics: {metrics}")

        self.assertGreater(metrics["accuracy"], 0.60)


if __name__ == "__main__":
    unittest.main()