def _evaluate_config(self, args, yaml_content):
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
mock_config.stop()
return config
def _evaluate_analyze_config(self, args, yaml_content):
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = ConfigCommandAnalyze()
cli = CLI()
cli.add_subcommand(
cmd='analyze',
help='Collect and sort profiling results and generate data and '
'summaries.',
config=config)
cli.parse()
mock_config.stop()
return config
def _evaluate_profile_config(self, args, yaml_content):
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = ConfigCommandProfile()
cli = CLI()
cli.add_subcommand(
cmd='profile',
help='Run model inference profiling based on specified CLI or '
'config options.',
config=config)
cli.parse()
mock_config.stop()
return config
def test_validation(self):
args = [
'model-analyzer', '--model-repository', 'cli_repository', '-f',
'path-to-config-file'
]
# end key should not be included in concurrency
yaml_content = """
model_names:
-
vgg_16_graphdef:
parameters:
concurrency:
start: 4
stop: 12
end: 2
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
with self.assertRaises(TritonModelAnalyzerException):
cli.parse()
mock_config.stop()
args = [
'model-analyzer', '--model-repository', 'cli_repository', '-f',
'path-to-config-file'
]
yaml_content = """
model_names:
-
vgg_16_graphdef:
parameters:
concurrency:
start: 13
stop: 12
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
with self.assertRaises(TritonModelAnalyzerException):
cli.parse()
mock_config.stop()
def _evaluate_config(self, args, yaml_content):
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config_analyze = ConfigCommandAnalyze()
config_report = ConfigCommandReport()
cli = CLI()
cli.add_subcommand(
cmd="analyze",
help=
"Collect and sort profiling results and generate data and summaries.",
config=config_analyze)
cli.add_subcommand(
cmd='report',
help='Generate detailed reports for a single config',
config=config_report)
cli.parse()
mock_config.stop()
ret = config_analyze if config_analyze.export_path else config_report
return ret
def test_config(self):
args = [
'model-analyzer', '--model-repository', 'cli_repository', '-f',
'path-to-config-file', '--model-names', 'vgg11'
]
yaml_content = 'model_repository: yaml_repository'
config = self._evaluate_config(args, yaml_content)
# CLI flag has the highest priority
self.assertTrue(
config.get_all_config()['model_repository'] == 'cli_repository')
args = [
'model-analyzer', '-f', 'path-to-config-file', '--model-names',
'vgg11'
]
yaml_content = 'model_repository: yaml_repository'
config = self._evaluate_config(args, yaml_content)
# If CLI flag doesn't exist, YAML config has the highest priority
self.assertTrue(
config.get_all_config()['model_repository'] == 'yaml_repository')
args = ['model-analyzer', '-f', 'path-to-config-file']
yaml_content = 'model_repository: yaml_repository'
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
# When a required field is not specified, parse will lead to an
# exception
with self.assertRaises(TritonModelAnalyzerException):
cli.parse()
mock_config.stop()
def test_constraints(self):
args = [
'model-analyzer', '--model-repository', 'cli_repository', '-f',
'path-to-config-file'
]
yaml_content = """
model_names:
-
vgg_16_graphdef:
parameters:
concurrency:
- 1
- 2
- 3
- 4
objectives:
perf_throughput: 10
gpu_used_memory: 5
constraints:
gpu_used_memory:
max: 80
- vgg_19_graphdef
"""
config = self._evaluate_config(args, yaml_content)
expected_model_objects = [
ConfigModel('vgg_16_graphdef',
parameters={
'batch_sizes': [1],
'concurrency': [1, 2, 3, 4]
},
objectives={
'perf_throughput': 10,
'gpu_used_memory': 5
},
constraints={'gpu_used_memory': {
'max': 80,
}}),
ConfigModel('vgg_19_graphdef',
parameters={
'batch_sizes': [1],
'concurrency': [1]
},
objectives={'perf_throughput': 10})
]
self._assert_equality_of_model_configs(
config.get_all_config()['model_names'], expected_model_objects)
# GPU Memory shouldn't have min
yaml_content = """
model_names:
-
vgg_16_graphdef:
parameters:
concurrency:
- 1
- 2
- 3
- 4
objectives:
- perf_throughput
- gpu_used_memory
constraints:
gpu_memory:
max: 80
min: 45
- vgg_19_graphdef
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
with self.assertRaises(TritonModelAnalyzerException):
cli.parse()
mock_config.stop()
# Test objective key that is not one of the supported metrics
yaml_content = """
model_names:
-
vgg_16_graphdef:
parameters:
concurrency:
- 1
- 2
- 3
- 4
objectives:
- throughput
constraints:
gpu_used_memory:
max: 80
- vgg_19_graphdef
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
with self.assertRaises(TritonModelAnalyzerException):
cli.parse()
mock_config.stop()
def test_range_and_list_values(self):
args = [
'model-analyzer', '--model-repository', 'cli_repository', '-f',
'path-to-config-file'
]
yaml_content = 'model_names: model_1,model_2'
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
self.assertTrue(
config.get_all_config()['model_names'] == ['model_1', 'model_2'])
mock_config.stop()
yaml_content = """
model_names:
- model_1
- model_2
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
self.assertTrue(
config.get_all_config()['model_names'] == ['model_1', 'model_2'])
mock_config.stop()
args = [
'model-analyzer', '--model-repository', 'cli_repository', '-f',
'path-to-config-file', '--model-names', 'model_1,model_2'
]
yaml_content = """
batch_sizes:
- 2
- 3
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
self.assertTrue(config.get_all_config()['batch_sizes'] == [2, 3])
mock_config.stop()
yaml_content = """
batch_sizes: 2
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
self.assertTrue(config.get_all_config()['batch_sizes'] == [2])
mock_config.stop()
yaml_content = """
concurrency: 2
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
self.assertTrue(config.get_all_config()['concurrency'] == [2])
self.assertTrue(config.get_all_config()['batch_sizes'] == [1])
mock_config.stop()
yaml_content = """
batch_sizes:
start: 2
stop: 6
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
self.assertTrue(
config.get_all_config()['batch_sizes'] == [2, 3, 4, 5, 6])
mock_config.stop()
yaml_content = """
batch_sizes:
start: 2
stop: 6
step: 2
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
self.assertTrue(config.get_all_config()['batch_sizes'] == [2, 4, 6])
mock_config.stop()
def test_constraints(self):
args = [
'model-analyzer', '--model-repository', 'cli_repository', '-f',
'path-to-config-file'
]
yaml_content = """
model_names:
-
vgg_16_graphdef:
parameters:
concurrency:
- 1
- 2
- 3
- 4
objectives:
- throughput
- gpu_memory
constraints:
gpu_memory:
max: 80
- vgg_19_graphdef
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
self.assertTrue(config.get_all_config()['model_names'] == [{
'vgg_16_graphdef': {
'parameters': {
'concurrency': [1, 2, 3, 4]
},
'objectives': ['throughput', 'gpu_memory'],
'constraints': {
'gpu_memory': {
'max': 80,
}
}
}
}, 'vgg_19_graphdef'])
mock_config.stop()
# GPU Memory shouldn't have min
yaml_content = """
model_names:
-
vgg_16_graphdef:
parameters:
concurrency:
- 1
- 2
- 3
- 4
objectives:
- throughput
- gpu_memory
constraints:
gpu_memory:
max: 80
min: 45
- vgg_19_graphdef
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
with self.assertRaises(TritonModelAnalyzerException):
cli.parse()
mock_config.stop()
def test_object(self):
args = [
'model-analyzer', '--model-repository', 'cli_repository', '-f',
'path-to-config-file'
]
yaml_content = """
model_names:
-
vgg_16_graphdef:
parameters:
concurrency:
- 1
- 2
- 3
- 4
- vgg_19_graphdef
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
self.assertTrue(config.get_all_config()['model_names'] == [{
'vgg_16_graphdef': {
'parameters': {
'concurrency': [1, 2, 3, 4]
}
}
}, 'vgg_19_graphdef'])
mock_config.stop()
yaml_content = """
model_names:
vgg_16_graphdef:
parameters:
concurrency:
- 1
- 2
- 3
- 4
vgg_19_graphdef:
parameters:
concurrency:
- 1
- 2
- 3
- 4
batch_sizes:
start: 2
stop: 6
step: 2
"""
mock_config = MockConfig(args, yaml_content)
mock_config.start()
config = AnalyzerConfig()
cli = CLI(config)
cli.parse()
self.assertTrue(
config.get_all_config()['model_names'] == {
'vgg_16_graphdef': {
'parameters': {
'concurrency': [1, 2, 3, 4]
}
},
'vgg_19_graphdef': {
'parameters': {
'concurrency': [1, 2, 3, 4],
'batch_sizes': [2, 4, 6]
}
}
})
mock_config.stop()