Spaces:
Paused
Paused
File size: 12,501 Bytes
2f5127c |
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 |
# Copyright 2020-2025 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import signal
import subprocess
import unittest
import psutil
import pytest
from transformers import AutoModelForCausalLM
from transformers.testing_utils import require_torch_multi_accelerator, torch_device
from trl.extras.vllm_client import VLLMClient
from trl.scripts.vllm_serve import chunk_list
from .testing_utils import require_3_accelerators
class TestChunkList(unittest.TestCase):
def test_even_split(self):
self.assertEqual(chunk_list([1, 2, 3, 4, 5, 6], 2), [[1, 2, 3], [4, 5, 6]])
def test_uneven_split(self):
self.assertEqual(chunk_list([1, 2, 3, 4, 5, 6], 4), [[1, 2], [3, 4], [5], [6]])
def test_more_chunks_than_elements(self):
self.assertEqual(chunk_list([1, 2, 3, 4, 5, 6], 8), [[1], [2], [3], [4], [5], [6], [], []])
def test_n_equals_len(self):
self.assertEqual(chunk_list([1, 2, 3], 3), [[1], [2], [3]])
def test_n_is_1(self):
self.assertEqual(chunk_list([1, 2, 3], 1), [[1, 2, 3]])
def test_single_element_list(self):
self.assertEqual(chunk_list([42], 2), [[42], []])
def test_any_dtype(self):
self.assertEqual(
chunk_list([1, "two", 3.0, {"four": 4}, ["f", "i", "v", "e"]], 2),
[[1, "two", 3.0], [{"four": 4}, ["f", "i", "v", "e"]]],
)
@pytest.mark.slow
@require_torch_multi_accelerator
class TestVLLMClientServer(unittest.TestCase):
model_id = "Qwen/Qwen2.5-1.5B"
@classmethod
def setUpClass(cls):
# We want the server to run on accelerator 1, so we set VISIBLE_DEVICES to "1"
env = os.environ.copy()
VISIBLE_DEVICES = "ZE_AFFINITY_MASK" if torch_device == "xpu" else "CUDA_VISIBLE_DEVICES"
env[VISIBLE_DEVICES] = "1" # Restrict to accelerator 1
# Start the server process
cls.server_process = subprocess.Popen(
["trl", "vllm-serve", "--model", cls.model_id], stdout=subprocess.PIPE, stderr=subprocess.PIPE, env=env
)
# Initialize the client
cls.client = VLLMClient(connection_timeout=240)
cls.client.init_communicator()
def test_generate(self):
prompts = ["Hello, AI!", "Tell me a joke"]
outputs = self.client.generate(prompts)
# Check that the output is a list
self.assertIsInstance(outputs, list)
# Check that the number of generated sequences is equal to the number of prompts
self.assertEqual(len(outputs), len(prompts))
# Check that the generated sequences are lists of integers
for seq in outputs:
self.assertTrue(all(isinstance(tok, int) for tok in seq))
def test_generate_with_params(self):
prompts = ["Hello, AI!", "Tell me a joke"]
outputs = self.client.generate(prompts, n=2, repetition_penalty=0.9, temperature=0.8, max_tokens=32)
# Check that the output is a list
self.assertIsInstance(outputs, list)
# Check that the number of generated sequences is 2 times the number of prompts
self.assertEqual(len(outputs), 2 * len(prompts))
# Check that the generated sequences are lists of integers
for seq in outputs:
self.assertTrue(all(isinstance(tok, int) for tok in seq))
# Check that the length of the generated sequences is less than or equal to 32
for seq in outputs:
self.assertLessEqual(len(seq), 32)
def test_update_model_params(self):
model = AutoModelForCausalLM.from_pretrained(self.model_id, device_map=torch_device)
self.client.update_model_params(model)
def test_reset_prefix_cache(self):
# Test resetting the prefix cache
self.client.reset_prefix_cache()
@classmethod
def tearDownClass(cls):
super().tearDownClass()
# Close the client
cls.client.close_communicator()
# vLLM x pytest (or Popen) seems not to handle process termination well. To avoid zombie processes, we need to
# kill the server process and its children explicitly.
parent = psutil.Process(cls.server_process.pid)
children = parent.children(recursive=True)
for child in children:
child.send_signal(signal.SIGTERM)
cls.server_process.terminate()
cls.server_process.wait()
# Same as above but using base_url to instantiate the client.
@pytest.mark.slow
@require_torch_multi_accelerator
class TestVLLMClientServerBaseURL(unittest.TestCase):
model_id = "Qwen/Qwen2.5-1.5B"
@classmethod
def setUpClass(cls):
# We want the server to run on accelerator 1, so we set VISIBLE_DEVICES to "1"
env = os.environ.copy()
VISIBLE_DEVICES = "ZE_AFFINITY_MASK" if torch_device == "xpu" else "CUDA_VISIBLE_DEVICES"
env[VISIBLE_DEVICES] = "1" # Restrict to accelerator 1
# Start the server process
cls.server_process = subprocess.Popen(
["trl", "vllm-serve", "--model", cls.model_id], stdout=subprocess.PIPE, stderr=subprocess.PIPE, env=env
)
# Initialize the client
cls.client = VLLMClient(base_url="http://localhost:8000", connection_timeout=240)
cls.client.init_communicator()
def test_generate(self):
prompts = ["Hello, AI!", "Tell me a joke"]
outputs = self.client.generate(prompts)
# Check that the output is a list
self.assertIsInstance(outputs, list)
# Check that the number of generated sequences is equal to the number of prompts
self.assertEqual(len(outputs), len(prompts))
# Check that the generated sequences are lists of integers
for seq in outputs:
self.assertTrue(all(isinstance(tok, int) for tok in seq))
def test_generate_with_params(self):
prompts = ["Hello, AI!", "Tell me a joke"]
outputs = self.client.generate(prompts, n=2, repetition_penalty=0.9, temperature=0.8, max_tokens=32)
# Check that the output is a list
self.assertIsInstance(outputs, list)
# Check that the number of generated sequences is 2 times the number of prompts
self.assertEqual(len(outputs), 2 * len(prompts))
# Check that the generated sequences are lists of integers
for seq in outputs:
self.assertTrue(all(isinstance(tok, int) for tok in seq))
# Check that the length of the generated sequences is less than or equal to 32
for seq in outputs:
self.assertLessEqual(len(seq), 32)
def test_update_model_params(self):
model = AutoModelForCausalLM.from_pretrained(self.model_id, device_map=torch_device)
self.client.update_model_params(model)
def test_reset_prefix_cache(self):
# Test resetting the prefix cache
self.client.reset_prefix_cache()
@classmethod
def tearDownClass(cls):
super().tearDownClass()
# Close the client
cls.client.close_communicator()
# vLLM x pytest (or Popen) seems not to handle process termination well. To avoid zombie processes, we need to
# kill the server process and its children explicitly.
parent = psutil.Process(cls.server_process.pid)
children = parent.children(recursive=True)
for child in children:
child.send_signal(signal.SIGTERM)
cls.server_process.terminate()
cls.server_process.wait()
@pytest.mark.slow
@require_3_accelerators
class TestVLLMClientServerTP(unittest.TestCase):
model_id = "Qwen/Qwen2.5-1.5B"
@classmethod
def setUpClass(cls):
# We want the server to run on accelerator 1 and 2, so we set VISIBLE_DEVICES to "1,2"
env = os.environ.copy()
VISIBLE_DEVICES = "ZE_AFFINITY_MASK" if torch_device == "xpu" else "CUDA_VISIBLE_DEVICES"
env[VISIBLE_DEVICES] = "1,2" # Restrict to accelerator 1 and 2
# Start the server process
cls.server_process = subprocess.Popen(
["trl", "vllm-serve", "--model", cls.model_id, "--tensor_parallel_size", "2"],
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
env=env,
)
# Initialize the client
cls.client = VLLMClient(connection_timeout=240)
cls.client.init_communicator()
def test_generate(self):
prompts = ["Hello, AI!", "Tell me a joke"]
outputs = self.client.generate(prompts)
# Check that the output is a list
self.assertIsInstance(outputs, list)
# Check that the number of generated sequences is equal to the number of prompts
self.assertEqual(len(outputs), len(prompts))
# Check that the generated sequences are lists of integers
for seq in outputs:
self.assertTrue(all(isinstance(tok, int) for tok in seq))
def test_update_model_params(self):
model = AutoModelForCausalLM.from_pretrained(self.model_id, device_map=torch_device)
self.client.update_model_params(model)
def test_reset_prefix_cache(self):
# Test resetting the prefix cache
self.client.reset_prefix_cache()
@classmethod
def tearDownClass(cls):
super().tearDownClass()
# Close the client
cls.client.close_communicator()
# vLLM x pytest (or Popen) seems not to handle process termination well. To avoid zombie processes, we need to
# kill the server process and its children explicitly.
parent = psutil.Process(cls.server_process.pid)
children = parent.children(recursive=True)
for child in children:
child.send_signal(signal.SIGTERM)
cls.server_process.terminate()
cls.server_process.wait()
@pytest.mark.slow
@require_3_accelerators
class TestVLLMClientServerDP(unittest.TestCase):
model_id = "Qwen/Qwen2.5-1.5B"
@classmethod
def setUpClass(cls):
# We want the server to run on accelerator 1 and 2, so we set VISIBLE_DEVICES to "1,2"
env = os.environ.copy()
VISIBLE_DEVICES = "ZE_AFFINITY_MASK" if torch_device == "xpu" else "CUDA_VISIBLE_DEVICES"
env[VISIBLE_DEVICES] = "1,2" # Restrict to accelerator 1 and 2
# Start the server process
cls.server_process = subprocess.Popen(
["trl", "vllm-serve", "--model", cls.model_id, "--data_parallel_size", "2"],
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
env=env,
)
# Initialize the client
cls.client = VLLMClient(connection_timeout=240)
def test_generate(self):
prompts = ["Hello, AI!", "Tell me a joke"]
outputs = self.client.generate(prompts)
# Check that the output is a list
self.assertIsInstance(outputs, list)
# Check that the number of generated sequences is equal to the number of prompts
self.assertEqual(len(outputs), len(prompts))
# Check that the generated sequences are lists of integers
for seq in outputs:
self.assertTrue(all(isinstance(tok, int) for tok in seq))
def test_update_model_params(self):
model = AutoModelForCausalLM.from_pretrained(self.model_id, device_map=torch_device)
self.client.update_model_params(model)
def test_reset_prefix_cache(self):
# Test resetting the prefix cache
self.client.reset_prefix_cache()
@classmethod
def tearDownClass(cls):
super().tearDownClass()
# Close the client
cls.client.close_communicator()
# vLLM x pytest (or Popen) seems not to handle process termination well. To avoid zombie processes, we need to
# kill the server process and its children explicitly.
parent = psutil.Process(cls.server_process.pid)
children = parent.children(recursive=True)
for child in children:
child.send_signal(signal.SIGTERM)
cls.server_process.terminate()
cls.server_process.wait()
|