File size: 13,504 Bytes
a402b9b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 | """
Unit tests for ModelOpt export functionality in SGLang.
These tests verify the integration of ModelOpt export API with SGLang's model loading
and quantization workflow.
"""
import json
import os
import tempfile
import unittest
from unittest.mock import Mock, patch
import torch
from sglang.srt.configs.device_config import DeviceConfig
from sglang.srt.configs.load_config import LoadConfig
from sglang.srt.configs.model_config import ModelConfig
from sglang.srt.model_loader.loader import ModelOptModelLoader
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
register_cuda_ci(est_time=9, suite="stage-b-test-large-1-gpu")
register_amd_ci(est_time=9, suite="stage-b-test-small-1-gpu-amd")
# Note: PYTHONPATH=python should be set when running tests
# Check if modelopt is available
try:
import modelopt # noqa: F401
MODELOPT_AVAILABLE = True
except ImportError:
MODELOPT_AVAILABLE = False
class TestModelOptExport(unittest.TestCase):
"""Test suite for ModelOpt export functionality."""
def setUp(self):
"""Set up test fixtures."""
# Mock distributed functionality to avoid initialization errors
self.mock_tp_rank = patch(
"sglang.srt.distributed.parallel_state.get_tensor_model_parallel_rank",
return_value=0,
)
self.mock_tp_rank.start()
self.mock_rank0_log = patch("sglang.srt.model_loader.loader.rank0_log")
self.mock_rank0_log.start()
# Mock logger to avoid issues
self.mock_logger = patch("sglang.srt.model_loader.loader.logger")
self.mock_logger.start()
# Mock all distributed functions that might be called
self.mock_get_tp_group = patch(
"sglang.srt.distributed.parallel_state.get_tp_group"
)
self.mock_get_tp_group.start()
# Mock model parallel initialization check
self.mock_mp_is_initialized = patch(
"sglang.srt.distributed.parallel_state.model_parallel_is_initialized",
return_value=True,
)
self.mock_mp_is_initialized.start()
self.temp_dir = tempfile.mkdtemp()
self.export_dir = os.path.join(self.temp_dir, "exported_model")
self.checkpoint_dir = os.path.join(self.temp_dir, "checkpoint")
# Mock model
self.mock_model = Mock(spec=torch.nn.Module)
self.mock_model.device = torch.device("cuda:0")
# Mock tokenizer
self.mock_tokenizer = Mock()
# Mock quantization config
self.mock_quant_cfg = Mock()
# Create ModelOptModelLoader instance
self.load_config = LoadConfig()
self.model_loader = ModelOptModelLoader(self.load_config)
def tearDown(self):
"""Clean up test fixtures."""
import shutil
shutil.rmtree(self.temp_dir, ignore_errors=True)
# Stop mocks
self.mock_tp_rank.stop()
self.mock_rank0_log.stop()
self.mock_logger.stop()
self.mock_get_tp_group.stop()
self.mock_mp_is_initialized.stop()
def _create_mock_export_files(self, export_dir: str):
"""Create mock export files for testing validation."""
os.makedirs(export_dir, exist_ok=True)
# Create config.json
config = {
"model_type": "test_model",
"architectures": ["TestModel"],
"quantization_config": {
"quant_method": "modelopt",
"bits": 8,
},
}
with open(os.path.join(export_dir, "config.json"), "w") as f:
json.dump(config, f)
# Create tokenizer_config.json
tokenizer_config = {"tokenizer_class": "TestTokenizer"}
with open(os.path.join(export_dir, "tokenizer_config.json"), "w") as f:
json.dump(tokenizer_config, f)
# Create model file
with open(os.path.join(export_dir, "model.safetensors"), "w") as f:
f.write("mock_model_data")
@unittest.skipIf(not MODELOPT_AVAILABLE, "nvidia-modelopt not available")
@patch("sglang.srt.model_loader.loader.os.makedirs")
@patch("modelopt.torch.export.export_hf_checkpoint")
def test_export_modelopt_checkpoint_success(self, mock_export, mock_makedirs):
"""Test successful model export."""
# Arrange
mock_export.return_value = None
mock_makedirs.return_value = None
# Act
self.model_loader._export_modelopt_checkpoint(self.mock_model, self.export_dir)
# Assert
mock_makedirs.assert_called_once_with(self.export_dir, exist_ok=True)
mock_export.assert_called_once_with(self.mock_model, export_dir=self.export_dir)
@unittest.skipIf(not MODELOPT_AVAILABLE, "nvidia-modelopt not available")
@patch("modelopt.torch.opt.restore")
@patch("modelopt.torch.quantization.utils.is_quantized")
def test_setup_quantization_with_export_from_checkpoint(
self, mock_is_quantized, mock_restore
):
"""Test export functionality when restoring from checkpoint."""
# Arrange
mock_is_quantized.return_value = False
mock_restore.return_value = None
with patch.object(
self.model_loader, "_export_modelopt_checkpoint"
) as mock_export:
# Act
self.model_loader._setup_modelopt_quantization(
self.mock_model,
self.mock_tokenizer,
self.mock_quant_cfg,
quantized_ckpt_restore_path=self.checkpoint_dir,
export_path=self.export_dir,
)
# Assert
mock_restore.assert_called_once_with(self.mock_model, self.checkpoint_dir)
mock_export.assert_called_once_with(self.mock_model, self.export_dir, None)
@unittest.skipIf(not MODELOPT_AVAILABLE, "nvidia-modelopt not available")
@patch("modelopt.torch.quantization.quantize")
@patch("modelopt.torch.quantization.print_quant_summary")
@patch("modelopt.torch.quantization.utils.is_quantized")
@patch("modelopt.torch.utils.dataset_utils.get_dataset_dataloader")
@patch("modelopt.torch.utils.dataset_utils.create_forward_loop")
def test_setup_quantization_with_export_after_calibration(
self,
mock_create_loop,
mock_get_dataloader,
mock_is_quantized,
mock_print_summary,
mock_quantize,
):
"""Test export functionality after calibration-based quantization."""
# Arrange
mock_is_quantized.return_value = False
mock_dataloader = Mock()
mock_get_dataloader.return_value = mock_dataloader
mock_calibrate_loop = Mock()
mock_create_loop.return_value = mock_calibrate_loop
mock_quantize.return_value = None
mock_print_summary.return_value = None
with patch.object(
self.model_loader, "_export_modelopt_checkpoint"
) as mock_export:
# Act
self.model_loader._setup_modelopt_quantization(
self.mock_model,
self.mock_tokenizer,
self.mock_quant_cfg,
export_path=self.export_dir,
)
# Assert
mock_quantize.assert_called_once_with(
self.mock_model, self.mock_quant_cfg, forward_loop=mock_calibrate_loop
)
mock_export.assert_called_once_with(self.mock_model, self.export_dir, None)
@unittest.skipIf(not MODELOPT_AVAILABLE, "nvidia-modelopt not available")
def test_setup_quantization_without_export(self):
"""Test quantization setup without export path specified."""
with patch("modelopt.torch.quantization.utils.is_quantized", return_value=True):
# Act
with patch.object(
self.model_loader, "_export_modelopt_checkpoint"
) as mock_export:
self.model_loader._setup_modelopt_quantization(
self.mock_model,
self.mock_tokenizer,
self.mock_quant_cfg,
export_path=None, # No export path
)
# Assert
mock_export.assert_not_called()
def test_quantize_and_serve_config_validation(self):
"""Test that quantize_and_serve is properly disabled."""
# Test that quantize-and-serve mode raises NotImplementedError
with self.assertRaises(NotImplementedError) as context:
ModelConfig(
model_path="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
quantization="modelopt_fp8",
quantize_and_serve=True,
)
# Verify the error message contains helpful instructions
error_msg = str(context.exception)
self.assertIn("disabled due to compatibility issues", error_msg)
self.assertIn("separate quantize-then-deploy workflow", error_msg)
# Test invalid configuration - no quantization
with self.assertRaises(ValueError) as context:
ModelConfig(
model_path="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
quantize_and_serve=True,
)
self.assertIn("requires ModelOpt quantization", str(context.exception))
@unittest.skipIf(not MODELOPT_AVAILABLE, "nvidia-modelopt not available")
def test_standard_workflow_selection(self):
"""Test that standard workflow is selected by default."""
with patch(
"modelopt.torch.quantization.utils.is_quantized", return_value=False
):
with patch.object(
self.model_loader, "_standard_quantization_workflow"
) as mock_standard:
with patch.object(self.model_loader, "_load_modelopt_base_model"):
mock_standard.return_value = Mock()
# Create model config without quantize_and_serve
model_config = ModelConfig(
model_path="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
quantization="modelopt_fp8",
quantize_and_serve=False,
)
device_config = DeviceConfig()
# Act
self.model_loader.load_model(
model_config=model_config,
device_config=device_config,
)
# Assert
mock_standard.assert_called_once_with(model_config, device_config)
def _get_export_info(self, export_dir: str) -> dict:
"""Get information about an exported model."""
if not self._validate_export(export_dir):
return None
try:
config_path = os.path.join(export_dir, "config.json")
with open(config_path, "r") as f:
config = json.load(f)
return {
"model_type": config.get("model_type", "unknown"),
"architectures": config.get("architectures", []),
"quantization_config": config.get("quantization_config", {}),
"export_dir": export_dir,
}
except Exception:
return None
@unittest.skipIf(not MODELOPT_AVAILABLE, "nvidia-modelopt not available")
class TestModelOptExportIntegration(unittest.TestCase):
"""Integration tests for ModelOpt export with full model loading workflow."""
def setUp(self):
"""Set up integration test fixtures."""
self.temp_dir = tempfile.mkdtemp()
self.export_dir = os.path.join(self.temp_dir, "exported_model")
def tearDown(self):
"""Clean up integration test fixtures."""
import shutil
shutil.rmtree(self.temp_dir, ignore_errors=True)
@patch("sglang.srt.model_loader.loader.get_model_architecture")
@patch("transformers.AutoTokenizer.from_pretrained")
@patch("transformers.AutoModelForCausalLM.from_pretrained")
def test_full_workflow_with_export(self, mock_model, mock_tokenizer, mock_arch):
"""Test the complete workflow from model config to export."""
# Arrange
mock_arch.return_value = ("TestModel", "TestConfig")
mock_tokenizer.return_value = Mock()
mock_model.return_value = Mock(spec=torch.nn.Module)
model_config = ModelConfig(
model_path="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
modelopt_quant="fp8",
modelopt_export_path=self.export_dir,
)
load_config = LoadConfig()
device_config = DeviceConfig()
# Mock the quantization and export process
with patch.object(
ModelOptModelLoader, "_setup_modelopt_quantization"
) as mock_setup:
with patch.object(
ModelOptModelLoader, "_load_modelopt_base_model"
) as mock_load_base:
mock_load_base.return_value = mock_model.return_value
# Act
model_loader = ModelOptModelLoader(load_config)
result = model_loader.load_model(
model_config=model_config,
device_config=device_config,
)
# Assert
self.assertIsNotNone(result)
mock_setup.assert_called_once()
# Verify export_path was passed to setup
args, kwargs = mock_setup.call_args
self.assertEqual(kwargs.get("export_path"), self.export_dir)
if __name__ == "__main__":
unittest.main()
|