def get_embedding(self, face_img):
if not isinstance(face_img, list):
face_img = [face_img]
face_img = np.stack(face_img)
input_size = tuple(face_img[0].shape[0:2][::-1])
blob = cv2.dnn.blobFromImages(
face_img,
1.0 / self.input_std,
input_size, (self.input_mean, self.input_mean, self.input_mean),
swapRB=True)
blob = blob.astype(triton_to_np_dtype(self.dtype))
inputs = []
inputs.append(
grpcclient.InferInput(self.input_name,
[blob.shape[0], self.c, self.h, self.w],
"FP32"))
# inputs[0].set_data_from_numpy(face_img)
cudashm.set_shared_memory_region(self.in_handle, [blob])
input_bytesize = 12 * blob.shape[0] * self.w * self.h
inputs[-1].set_shared_memory(self.in_handle_name, input_bytesize)
outputs = []
out_bytesize = 12 * 512 * self.max_batch_size
outputs.append(grpcclient.InferRequestedOutput(self.output_name[0]))
outputs[-1].set_shared_memory(self.out_handle_name, out_bytesize)
out = self.triton_client.infer(self.model_name,
inputs,
model_version=self.model_version,
outputs=outputs)
out = [
cudashm.get_contents_as_numpy(self.out_handle,
triton_to_np_dtype(self.dtype),
[blob.shape[0], 512])
]
# out = [out.as_numpy(e) for e in self.output_name]
return out[0]
def infer_exact(tester,
pf,
tensor_shape,
batch_size,
input_dtype,
output0_dtype,
output1_dtype,
output0_raw=True,
output1_raw=True,
model_version=None,
swap=False,
outputs=("OUTPUT0", "OUTPUT1"),
use_http=True,
use_grpc=True,
use_http_json_tensors=True,
skip_request_id_check=False,
use_streaming=True,
correlation_id=0,
shm_region_names=None,
precreated_shm_regions=None,
use_system_shared_memory=False,
use_cuda_shared_memory=False,
priority=0,
timeout_us=0):
tester.assertTrue(use_http or use_grpc or use_streaming)
# configs [ url, protocol, async stream, binary data ]
configs = []
if use_http:
configs.append(("localhost:8000", "http", False, True))
if output0_raw == output1_raw:
# Float16 not supported for Input and Output via JSON
if use_http_json_tensors and (input_dtype != np.float16) and \
(output0_dtype != np.float16) and (output1_dtype != np.float16):
configs.append(("localhost:8000", "http", False, False))
if use_grpc:
configs.append(("localhost:8001", "grpc", False, False))
if use_streaming:
configs.append(("localhost:8001", "grpc", True, False))
# outputs are sum and difference of inputs so set max input
# values so that they will not overflow the output. This
# allows us to do an exact match. For float types use 8, 16,
# 32 int range for fp 16, 32, 64 respectively. When getting
# class outputs the result value/probability is returned as a
# float so must use fp32 range in that case.
rinput_dtype = _range_repr_dtype(input_dtype)
routput0_dtype = _range_repr_dtype(
output0_dtype if output0_raw else np.float32)
routput1_dtype = _range_repr_dtype(
output1_dtype if output1_raw else np.float32)
val_min = max(
np.iinfo(rinput_dtype).min,
np.iinfo(routput0_dtype).min,
np.iinfo(routput1_dtype).min) / 2
val_max = min(
np.iinfo(rinput_dtype).max,
np.iinfo(routput0_dtype).max,
np.iinfo(routput1_dtype).max) / 2
num_classes = 3
input0_array = np.random.randint(low=val_min,
high=val_max,
size=tensor_shape,
dtype=rinput_dtype)
input1_array = np.random.randint(low=val_min,
high=val_max,
size=tensor_shape,
dtype=rinput_dtype)
if input_dtype != np.object:
input0_array = input0_array.astype(input_dtype)
input1_array = input1_array.astype(input_dtype)
if not swap:
output0_array = input0_array + input1_array
output1_array = input0_array - input1_array
else:
output0_array = input0_array - input1_array
output1_array = input0_array + input1_array
if output0_dtype == np.object:
output0_array = np.array([
unicode(str(x), encoding='utf-8')
for x in (output0_array.flatten())
],
dtype=object).reshape(output0_array.shape)
else:
output0_array = output0_array.astype(output0_dtype)
if output1_dtype == np.object:
output1_array = np.array([
unicode(str(x), encoding='utf-8')
for x in (output1_array.flatten())
],
dtype=object).reshape(output1_array.shape)
else:
output1_array = output1_array.astype(output1_dtype)
if input_dtype == np.object:
in0n = np.array(
[str(x) for x in input0_array.reshape(input0_array.size)],
dtype=object)
input0_array = in0n.reshape(input0_array.shape)
in1n = np.array(
[str(x) for x in input1_array.reshape(input1_array.size)],
dtype=object)
input1_array = in1n.reshape(input1_array.shape)
# prepend size of string to output string data
if output0_dtype == np.object:
if batch_size == 1:
output0_array_tmp = serialize_byte_tensor_list([output0_array])
else:
output0_array_tmp = serialize_byte_tensor_list(output0_array)
else:
output0_array_tmp = output0_array
if output1_dtype == np.object:
if batch_size == 1:
output1_array_tmp = serialize_byte_tensor_list([output1_array])
else:
output1_array_tmp = serialize_byte_tensor_list(output1_array)
else:
output1_array_tmp = output1_array
# Get model platform
model_name = tu.get_model_name(pf, input_dtype, output0_dtype,
output1_dtype)
if configs[0][1] == "http":
metadata_client = httpclient.InferenceServerClient(configs[0][0],
verbose=True)
metadata = metadata_client.get_model_metadata(model_name)
platform = metadata["platform"]
else:
metadata_client = grpcclient.InferenceServerClient(configs[0][0],
verbose=True)
metadata = metadata_client.get_model_metadata(model_name)
platform = metadata.platform
if platform == "pytorch_libtorch":
OUTPUT0 = "OUTPUT__0"
OUTPUT1 = "OUTPUT__1"
INPUT0 = "INPUT__0"
INPUT1 = "INPUT__1"
else:
OUTPUT0 = "OUTPUT0"
OUTPUT1 = "OUTPUT1"
INPUT0 = "INPUT0"
INPUT1 = "INPUT1"
output0_byte_size = sum([o0.nbytes for o0 in output0_array_tmp])
output1_byte_size = sum([o1.nbytes for o1 in output1_array_tmp])
if batch_size == 1:
input0_list = [input0_array]
input1_list = [input1_array]
else:
input0_list = [x for x in input0_array]
input1_list = [x for x in input1_array]
# Serialization of string tensors in the case of shared memory must be done manually
if input_dtype == np.object:
input0_list_tmp = serialize_byte_tensor_list(input0_list)
input1_list_tmp = serialize_byte_tensor_list(input1_list)
else:
input0_list_tmp = input0_list
input1_list_tmp = input1_list
input0_byte_size = sum([i0.nbytes for i0 in input0_list_tmp])
input1_byte_size = sum([i1.nbytes for i1 in input1_list_tmp])
# Create system/cuda shared memory regions if needed
shm_regions, shm_handles = su.create_set_shm_regions(
input0_list_tmp, input1_list_tmp, output0_byte_size, output1_byte_size,
outputs, shm_region_names, precreated_shm_regions,
use_system_shared_memory, use_cuda_shared_memory)
if model_version is not None:
model_version = str(model_version)
else:
model_version = ""
# Run inference and check results for each config
for config in configs:
model_name = tu.get_model_name(pf, input_dtype, output0_dtype,
output1_dtype)
if config[1] == "http":
triton_client = httpclient.InferenceServerClient(config[0],
verbose=True)
else:
triton_client = grpcclient.InferenceServerClient(config[0],
verbose=True)
inputs = []
if config[1] == "http":
inputs.append(
httpclient.InferInput(INPUT0, tensor_shape,
np_to_triton_dtype(input_dtype)))
inputs.append(
httpclient.InferInput(INPUT1, tensor_shape,
np_to_triton_dtype(input_dtype)))
else:
inputs.append(
grpcclient.InferInput(INPUT0, tensor_shape,
np_to_triton_dtype(input_dtype)))
inputs.append(
grpcclient.InferInput(INPUT1, tensor_shape,
np_to_triton_dtype(input_dtype)))
if not (use_cuda_shared_memory or use_system_shared_memory):
if config[1] == "http":
inputs[0].set_data_from_numpy(input0_array,
binary_data=config[3])
inputs[1].set_data_from_numpy(input1_array,
binary_data=config[3])
else:
inputs[0].set_data_from_numpy(input0_array)
inputs[1].set_data_from_numpy(input1_array)
else:
# Register necessary shared memory regions/handles
su.register_add_shm_regions(inputs, outputs, shm_regions,
precreated_shm_regions, shm_handles,
input0_byte_size, input1_byte_size,
output0_byte_size, output1_byte_size,
use_system_shared_memory,
use_cuda_shared_memory, triton_client)
if batch_size == 1:
expected0_sort_idx = [
np.flip(np.argsort(x.flatten()), 0)
for x in output0_array.reshape((1, ) + tensor_shape)
]
expected1_sort_idx = [
np.flip(np.argsort(x.flatten()), 0)
for x in output1_array.reshape((1, ) + tensor_shape)
]
else:
expected0_sort_idx = [
np.flip(np.argsort(x.flatten()), 0)
for x in output0_array.reshape(tensor_shape)
]
expected1_sort_idx = [
np.flip(np.argsort(x.flatten()), 0)
for x in output1_array.reshape(tensor_shape)
]
# Force binary_data = False for shared memory and class
output_req = []
i = 0
if "OUTPUT0" in outputs:
if len(shm_regions) != 0:
if config[1] == "http":
output_req.append(
httpclient.InferRequestedOutput(OUTPUT0,
binary_data=config[3]))
else:
output_req.append(grpcclient.InferRequestedOutput(OUTPUT0))
output_req[-1].set_shared_memory(shm_regions[2] + '_data',
output0_byte_size)
else:
if output0_raw:
if config[1] == "http":
output_req.append(
httpclient.InferRequestedOutput(
OUTPUT0, binary_data=config[3]))
else:
output_req.append(
grpcclient.InferRequestedOutput(OUTPUT0))
else:
if config[1] == "http":
output_req.append(
httpclient.InferRequestedOutput(
OUTPUT0,
binary_data=config[3],
class_count=num_classes))
else:
output_req.append(
grpcclient.InferRequestedOutput(
OUTPUT0, class_count=num_classes))
i += 1
if "OUTPUT1" in outputs:
if len(shm_regions) != 0:
if config[1] == "http":
output_req.append(
httpclient.InferRequestedOutput(OUTPUT1,
binary_data=config[3]))
else:
output_req.append(grpcclient.InferRequestedOutput(OUTPUT1))
output_req[-1].set_shared_memory(shm_regions[2 + i] + '_data',
output1_byte_size)
else:
if output1_raw:
if config[1] == "http":
output_req.append(
httpclient.InferRequestedOutput(
OUTPUT1, binary_data=config[3]))
else:
output_req.append(
grpcclient.InferRequestedOutput(OUTPUT1))
else:
if config[1] == "http":
output_req.append(
httpclient.InferRequestedOutput(
OUTPUT1,
binary_data=config[3],
class_count=num_classes))
else:
output_req.append(
grpcclient.InferRequestedOutput(
OUTPUT1, class_count=num_classes))
if config[2]:
user_data = UserData()
triton_client.start_stream(partial(completion_callback, user_data))
try:
results = triton_client.async_stream_infer(
model_name,
inputs,
model_version=model_version,
outputs=output_req,
request_id=str(_unique_request_id()))
except Exception as e:
triton_client.stop_stream()
raise e
triton_client.stop_stream()
(results, error) = user_data._completed_requests.get()
if error is not None:
raise error
else:
results = triton_client.infer(model_name,
inputs,
model_version=model_version,
outputs=output_req,
request_id=str(_unique_request_id()))
last_response = results.get_response()
if not skip_request_id_check:
global _seen_request_ids
if config[1] == "http":
request_id = int(last_response["id"])
else:
request_id = int(last_response.id)
tester.assertFalse(request_id in _seen_request_ids,
"request_id: {}".format(request_id))
_seen_request_ids.add(request_id)
if config[1] == "http":
response_model_name = last_response["model_name"]
if model_version != "":
response_model_version = last_response["model_version"]
response_outputs = last_response["outputs"]
else:
response_model_name = last_response.model_name
if model_version != "":
response_model_version = last_response.model_version
response_outputs = last_response.outputs
tester.assertEqual(response_model_name, model_name)
if model_version != "":
tester.assertEqual(str(response_model_version), model_version)
tester.assertEqual(len(response_outputs), len(outputs))
for result in response_outputs:
if config[1] == "http":
result_name = result["name"]
else:
result_name = result.name
if ((result_name == OUTPUT0 and output0_raw)
or (result_name == OUTPUT1 and output1_raw)):
if use_system_shared_memory or use_cuda_shared_memory:
if result_name == OUTPUT0:
shm_handle = shm_handles[2]
else:
shm_handle = shm_handles[3]
output = results.get_output(result_name)
if config[1] == "http":
output_datatype = output['datatype']
output_shape = output['shape']
else:
output_datatype = output.datatype
output_shape = output.shape
output_dtype = triton_to_np_dtype(output_datatype)
if use_system_shared_memory:
output_data = shm.get_contents_as_numpy(
shm_handle, output_dtype, output_shape)
elif use_cuda_shared_memory:
output_data = cudashm.get_contents_as_numpy(
shm_handle, output_dtype, output_shape)
else:
output_data = results.as_numpy(result_name)
if (output_data.dtype == np.object) and (config[3] == False):
output_data = output_data.astype(np.bytes_)
if result_name == OUTPUT0:
tester.assertTrue(
np.array_equal(output_data, output0_array),
"{}, {} expected: {}, got {}".format(
model_name, OUTPUT0, output0_array, output_data))
elif result_name == OUTPUT1:
tester.assertTrue(
np.array_equal(output_data, output1_array),
"{}, {} expected: {}, got {}".format(
model_name, OUTPUT1, output1_array, output_data))
else:
tester.assertTrue(
False, "unexpected raw result {}".format(result_name))
else:
for b in range(batch_size):
# num_classes values must be returned and must
# match expected top values
if "nobatch" in pf:
class_list = results.as_numpy(result_name)
else:
class_list = results.as_numpy(result_name)[b]
tester.assertEqual(len(class_list), num_classes)
if batch_size == 1:
expected0_flatten = output0_array.flatten()
expected1_flatten = output1_array.flatten()
else:
expected0_flatten = output0_array[b].flatten()
expected1_flatten = output1_array[b].flatten()
for idx, class_label in enumerate(class_list):
# can't compare indices since could have different
# indices with the same value/prob, so check that
# the value of each index equals the expected value.
# Only compare labels when the indices are equal.
if type(class_label) == str:
ctuple = class_label.split(':')
else:
ctuple = "".join(chr(x)
for x in class_label).split(':')
cval = float(ctuple[0])
cidx = int(ctuple[1])
if result_name == OUTPUT0:
tester.assertEqual(cval, expected0_flatten[cidx])
tester.assertEqual(
cval,
expected0_flatten[expected0_sort_idx[b][idx]])
if cidx == expected0_sort_idx[b][idx]:
tester.assertEqual(
ctuple[2], 'label{}'.format(
expected0_sort_idx[b][idx]))
elif result_name == OUTPUT1:
tester.assertEqual(cval, expected1_flatten[cidx])
tester.assertEqual(
cval,
expected1_flatten[expected1_sort_idx[b][idx]])
else:
tester.assertTrue(
False, "unexpected class result {}".format(
result_name))
# Unregister system/cuda shared memory regions if they exist
su.unregister_cleanup_shm_regions(shm_regions, shm_handles,
precreated_shm_regions, outputs,
use_system_shared_memory,
use_cuda_shared_memory)
return results
def infer_zero(tester,
pf,
batch_size,
tensor_dtype,
input_shapes,
output_shapes,
model_version=None,
use_http=True,
use_grpc=True,
use_http_json_tensors=True,
use_streaming=True,
shm_region_name_prefix=None,
use_system_shared_memory=False,
use_cuda_shared_memory=False,
priority=0,
timeout_us=0):
tester.assertTrue(use_http or use_grpc or use_streaming)
configs = []
if use_http:
configs.append(("localhost:8000", "http", False, True))
if use_http_json_tensors and (tensor_dtype != np.float16):
configs.append(("localhost:8000", "http", False, False))
if use_grpc:
configs.append(("localhost:8001", "grpc", False, False))
if use_streaming:
configs.append(("localhost:8001", "grpc", True, False))
tester.assertEqual(len(input_shapes), len(output_shapes))
io_cnt = len(input_shapes)
if shm_region_name_prefix is None:
shm_region_name_prefix = ["input", "output"]
input_dict = {}
expected_dict = {}
shm_ip_handles = list()
shm_op_handles = list()
# Get model platform
model_name = tu.get_zero_model_name(pf, io_cnt, tensor_dtype)
if configs[0][1] == "http":
metadata_client = httpclient.InferenceServerClient(configs[0][0],
verbose=True)
metadata = metadata_client.get_model_metadata(model_name)
platform = metadata["platform"]
else:
metadata_client = grpcclient.InferenceServerClient(configs[0][0],
verbose=True)
metadata = metadata_client.get_model_metadata(model_name)
platform = metadata.platform
for io_num in range(io_cnt):
if platform == "pytorch_libtorch":
input_name = "INPUT__{}".format(io_num)
output_name = "OUTPUT__{}".format(io_num)
else:
input_name = "INPUT{}".format(io_num)
output_name = "OUTPUT{}".format(io_num)
input_shape = input_shapes[io_num]
output_shape = output_shapes[io_num]
rtensor_dtype = _range_repr_dtype(tensor_dtype)
if (rtensor_dtype != np.bool):
input_array = np.random.randint(low=np.iinfo(rtensor_dtype).min,
high=np.iinfo(rtensor_dtype).max,
size=input_shape,
dtype=rtensor_dtype)
else:
input_array = np.random.choice(a=[False, True], size=input_shape)
if tensor_dtype != np.object:
input_array = input_array.astype(tensor_dtype)
expected_array = np.ndarray.copy(input_array)
else:
expected_array = np.array([
unicode(str(x), encoding='utf-8')
for x in input_array.flatten()
],
dtype=object)
input_array = np.array([str(x) for x in input_array.flatten()],
dtype=object).reshape(input_array.shape)
expected_array = expected_array.reshape(output_shape)
expected_dict[output_name] = expected_array
output_byte_size = expected_array.nbytes
if batch_size == 1:
input_list = [input_array]
else:
input_list = [x for x in input_array]
# Serialization of string tensors in the case of shared memory must be done manually
if tensor_dtype == np.object:
input_list_tmp = serialize_byte_tensor_list(input_list)
else:
input_list_tmp = input_list
input_byte_size = sum([ip.nbytes for ip in input_list_tmp])
# create and register shared memory region for inputs and outputs
shm_io_handles = su.create_set_either_shm_region(
[
shm_region_name_prefix[0] + str(io_num),
shm_region_name_prefix[1] + str(io_num)
], input_list_tmp, input_byte_size, output_byte_size,
use_system_shared_memory, use_cuda_shared_memory)
if len(shm_io_handles) != 0:
shm_ip_handles.append(shm_io_handles[0])
shm_op_handles.append(shm_io_handles[1])
input_dict[input_name] = input_array
if model_version is not None:
model_version = str(model_version)
else:
model_version = ""
# Run inference and check results for each config
for config in configs:
model_name = tu.get_zero_model_name(pf, io_cnt, tensor_dtype)
if config[1] == "http":
triton_client = httpclient.InferenceServerClient(config[0],
verbose=True)
else:
triton_client = grpcclient.InferenceServerClient(config[0],
verbose=True)
inputs = []
output_req = []
for io_num, (input_name, output_name) in enumerate(
zip(input_dict.keys(), expected_dict.keys())):
input_data = input_dict[input_name]
input_byte_size = input_data.nbytes
output_byte_size = expected_dict[output_name].nbytes
if config[1] == "http":
inputs.append(
httpclient.InferInput(input_name, input_data.shape,
np_to_triton_dtype(tensor_dtype)))
output_req.append(
httpclient.InferRequestedOutput(output_name,
binary_data=config[3]))
else:
inputs.append(
grpcclient.InferInput(input_name, input_data.shape,
np_to_triton_dtype(tensor_dtype)))
output_req.append(grpcclient.InferRequestedOutput(output_name))
if not (use_cuda_shared_memory or use_system_shared_memory):
if config[1] == "http":
inputs[-1].set_data_from_numpy(input_data,
binary_data=config[3])
else:
inputs[-1].set_data_from_numpy(input_data)
else:
# Register necessary shared memory regions/handles
su.register_add_either_shm_regions(
inputs, output_req, shm_region_name_prefix,
(shm_ip_handles, shm_op_handles), io_num, input_byte_size,
output_byte_size, use_system_shared_memory,
use_cuda_shared_memory, triton_client)
if config[2]:
user_data = UserData()
triton_client.start_stream(partial(completion_callback, user_data))
try:
results = triton_client.async_stream_infer(
model_name,
inputs,
model_version=model_version,
outputs=output_req,
request_id=str(_unique_request_id()),
priority=priority,
timeout=timeout_us)
except Exception as e:
triton_client.stop_stream()
raise e
triton_client.stop_stream()
(results, error) = user_data._completed_requests.get()
if error is not None:
raise error
else:
results = triton_client.infer(model_name,
inputs,
model_version=model_version,
outputs=output_req,
request_id=str(_unique_request_id()),
priority=priority,
timeout=timeout_us)
last_response = results.get_response()
if config[1] == "http":
response_model_name = last_response["model_name"]
if model_version != "":
response_model_version = last_response["model_version"]
response_outputs = last_response["outputs"]
else:
response_model_name = last_response.model_name
if model_version != "":
response_model_version = last_response.model_version
response_outputs = last_response.outputs
tester.assertEqual(response_model_name, model_name)
if model_version != "":
tester.assertEqual(response_model_version, model_version)
tester.assertEqual(len(response_outputs), io_cnt)
for result in response_outputs:
if config[1] == "http":
result_name = result["name"]
else:
result_name = result.name
tester.assertTrue(result_name in expected_dict)
if use_system_shared_memory or use_cuda_shared_memory:
if platform == "pytorch_libtorch":
io_num = int(result_name.split("OUTPUT__")[1])
else:
io_num = int(result_name.split("OUTPUT")[1])
shm_handle = shm_op_handles[io_num]
output = results.get_output(result_name)
if config[1] == "http":
output_datatype = output['datatype']
output_shape = output['shape']
else:
output_datatype = output.datatype
output_shape = output.shape
output_dtype = triton_to_np_dtype(output_datatype)
if use_system_shared_memory:
output_data = shm.get_contents_as_numpy(
shm_handle, output_dtype, output_shape)
elif use_cuda_shared_memory:
output_data = cudashm.get_contents_as_numpy(
shm_handle, output_dtype, output_shape)
else:
output_data = results.as_numpy(result_name)
if (output_data.dtype == np.object) and (config[3] == False):
output_data = output_data.astype(np.bytes_)
expected = expected_dict[result_name]
tester.assertEqual(output_data.shape, expected.shape)
tester.assertTrue(
np.array_equal(output_data, expected),
"{}, {}, expected: {}, got {}".format(model_name, result_name,
expected, output_data))
if len(shm_ip_handles) != 0:
for io_num in range(io_cnt):
if use_cuda_shared_memory:
triton_client.unregister_cuda_shared_memory(
shm_region_name_prefix[0] + str(io_num) + '_data')
triton_client.unregister_cuda_shared_memory(
shm_region_name_prefix[0] + str(io_num) + '_data')
cudashm.destroy_shared_memory_region(shm_ip_handles[io_num])
cudashm.destroy_shared_memory_region(shm_op_handles[io_num])
else:
triton_client.unregister_system_shared_memory(
shm_region_name_prefix[1] + str(io_num) + '_data')
triton_client.unregister_system_shared_memory(
shm_region_name_prefix[1] + str(io_num) + '_data')
shm.destroy_shared_memory_region(shm_ip_handles[io_num])
shm.destroy_shared_memory_region(shm_op_handles[io_num])
return results
def check_sequence_async(self,
trial,
model_name,
input_dtype,
correlation_id,
sequence_thresholds,
values,
expected_result,
shm_region_handles,
batch_size=1,
sequence_name="<unknown>",
tensor_shape=(1,)):
"""Perform sequence of inferences using stream async run.
The 'values' holds a list of tuples, one for each inference with format:
(flag_str, value, pre_delay_ms)
"""
if (("savedmodel" not in trial) and ("graphdef" not in trial) and
("custom" not in trial) and ("onnx" not in trial) and
("libtorch" not in trial) and ("plan" not in trial)):
self.assertFalse(True, "unknown trial type: " + trial)
self.assertFalse(
_test_system_shared_memory and _test_cuda_shared_memory,
"Cannot set both System and CUDA shared memory flags to 1")
full_shape = tensor_shape if "nobatch" in trial else (
batch_size,) + tensor_shape
client_utils = grpcclient
triton_client = client_utils.InferenceServerClient(f"{_tritonserver_ipaddr}:8001",
verbose=True)
user_data = UserData()
triton_client.start_stream(partial(completion_callback, user_data))
# Execute the sequence of inference...
try:
seq_start_ms = int(round(time.time() * 1000))
INPUT = "INPUT__0" if trial.startswith("libtorch") else "INPUT"
OUTPUT = "OUTPUT__0" if trial.startswith("libtorch") else "OUTPUT"
sent_count = 0
for flag_str, value, pre_delay_ms in values:
seq_start = False
seq_end = False
if flag_str is not None:
seq_start = ("start" in flag_str)
seq_end = ("end" in flag_str)
# Construct request IOs
inputs = []
outputs = []
inputs.append(
client_utils.InferInput(INPUT, full_shape,
np_to_triton_dtype(input_dtype)))
outputs.append(client_utils.InferRequestedOutput(OUTPUT))
if not (_test_system_shared_memory or _test_cuda_shared_memory):
if input_dtype == np.object_:
in0 = np.full(full_shape, value, dtype=np.int32)
in0n = np.array([str(x) for x in in0.reshape(in0.size)],
dtype=object)
in0 = in0n.reshape(full_shape)
else:
in0 = np.full(full_shape, value, dtype=input_dtype)
inputs[0].set_data_from_numpy(in0)
else:
offset = 2 * sent_count
inputs[0].set_shared_memory(shm_region_handles[offset][0],
shm_region_handles[offset][1])
outputs[0].set_shared_memory(
shm_region_handles[offset + 1][0],
shm_region_handles[offset + 1][1])
if pre_delay_ms is not None:
time.sleep(pre_delay_ms / 1000.0)
triton_client.async_stream_infer(model_name,
inputs,
outputs=outputs,
sequence_id=correlation_id,
sequence_start=seq_start,
sequence_end=seq_end)
sent_count += 1
# Wait for the results in the order sent
result = None
processed_count = 0
while processed_count < sent_count:
(results, error) = user_data._completed_requests.get()
if error is not None:
raise error
# Get value of "OUTPUT", for shared memory, need to get it via
# shared memory utils
if (not _test_system_shared_memory) and (
not _test_cuda_shared_memory):
out = results.as_numpy(OUTPUT)
else:
output = results.get_output(OUTPUT)
offset = 2 * processed_count + 1
output_shape = output.shape
output_type = input_dtype
if _test_system_shared_memory:
out = shm.get_contents_as_numpy(
shm_region_handles[offset][2], output_type,
output_shape)
else:
out = cudashm.get_contents_as_numpy(
shm_region_handles[offset][2], output_type,
output_shape)
result = out[0] if "nobatch" in trial else out[0][0]
print("{}: {}".format(sequence_name, result))
processed_count += 1
seq_end_ms = int(round(time.time() * 1000))
if input_dtype == np.object_:
self.assertEqual(int(result), expected_result)
else:
self.assertEqual(result, expected_result)
if sequence_thresholds is not None:
lt_ms = sequence_thresholds[0]
gt_ms = sequence_thresholds[1]
if lt_ms is not None:
if _test_jetson:
lt_ms *= _jetson_slowdown_factor
self.assertTrue((seq_end_ms - seq_start_ms) < lt_ms,
"sequence expected less than " +
str(lt_ms) + "ms response time, got " +
str(seq_end_ms - seq_start_ms) + " ms")
if gt_ms is not None:
self.assertTrue((seq_end_ms - seq_start_ms) > gt_ms,
"sequence expected greater than " +
str(gt_ms) + "ms response time, got " +
str(seq_end_ms - seq_start_ms) + " ms")
except Exception as ex:
self.add_deferred_exception(ex)
triton_client.stop_stream()
def check_sequence(self,
trial,
model_name,
input_dtype,
correlation_id,
sequence_thresholds,
values,
expected_result,
protocol,
batch_size=1,
sequence_name="<unknown>",
tensor_shape=(1,)):
"""Perform sequence of inferences. The 'values' holds a list of
tuples, one for each inference with format:
(flag_str, value, (ls_ms, gt_ms), (pre_delay_ms, post_delay_ms)
"""
if (("savedmodel" not in trial) and ("graphdef" not in trial) and
("custom" not in trial) and ("onnx" not in trial) and
("libtorch" not in trial) and ("plan" not in trial)):
self.assertFalse(True, "unknown trial type: " + trial)
# Can only send the request exactly once since it is a
# sequence model with state, so can have only a single config.
configs = []
if protocol == "http":
configs.append((f"{_tritonserver_ipaddr}:8000", "http", False))
if protocol == "grpc":
configs.append((f"{_tritonserver_ipaddr}:8001", "grpc", False))
if protocol == "streaming":
configs.append((f"{_tritonserver_ipaddr}:8001", "grpc", True))
self.assertFalse(
_test_system_shared_memory and _test_cuda_shared_memory,
"Cannot set both System and CUDA shared memory flags to 1")
self.assertEqual(len(configs), 1)
full_shape = tensor_shape if "nobatch" in trial else (
batch_size,) + tensor_shape
# create and register shared memory output region in advance,
# knowing that this function will not be called concurrently.
if _test_system_shared_memory or _test_cuda_shared_memory:
self.triton_client_.unregister_system_shared_memory()
self.triton_client_.unregister_cuda_shared_memory()
output_byte_size = 512
if _test_system_shared_memory:
shm_op_handle = shm.create_shared_memory_region(
"output_data", "/output", output_byte_size)
self.triton_client_.register_system_shared_memory(
"output_data", "/output", output_byte_size)
elif _test_cuda_shared_memory:
shm_op_handle = cudashm.create_shared_memory_region(
"output_data", output_byte_size, 0)
self.triton_client_.register_cuda_shared_memory(
"output_data", cudashm.get_raw_handle(shm_op_handle), 0,
output_byte_size)
shm_ip_handles = []
for config in configs:
client_utils = grpcclient if config[1] == "grpc" else httpclient
triton_client = client_utils.InferenceServerClient(config[0],
verbose=True)
if config[2]:
user_data = UserData()
triton_client.start_stream(
partial(completion_callback, user_data))
# Execute the sequence of inference...
try:
seq_start_ms = int(round(time.time() * 1000))
INPUT = "INPUT__0" if trial.startswith("libtorch") else "INPUT"
OUTPUT = "OUTPUT__0" if trial.startswith(
"libtorch") else "OUTPUT"
for flag_str, value, thresholds, delay_ms in values:
if _test_valgrind or _test_jetson:
if delay_ms is not None:
delay_ms[0] = max(_valgrind_delay_ms, delay_ms[0])
delay_ms[1] = max(_valgrind_delay_ms, delay_ms[1])
else:
delay_ms = (_valgrind_delay_ms, _valgrind_delay_ms)
if delay_ms is not None:
time.sleep(delay_ms[0] / 1000.0)
seq_start = False
seq_end = False
if flag_str is not None:
seq_start = ("start" in flag_str)
seq_end = ("end" in flag_str)
# Construct request IOs
inputs = []
outputs = []
inputs.append(
client_utils.InferInput(
INPUT, full_shape, np_to_triton_dtype(input_dtype)))
outputs.append(client_utils.InferRequestedOutput(OUTPUT))
if input_dtype == np.object_:
in0 = np.full(full_shape, value, dtype=np.int32)
in0n = np.array([str(x) for x in in0.reshape(in0.size)],
dtype=object)
in0 = in0n.reshape(full_shape)
else:
in0 = np.full(full_shape, value, dtype=input_dtype)
# create input shared memory and copy input data values into it
if _test_system_shared_memory or _test_cuda_shared_memory:
if input_dtype == np.object_:
input_list_tmp = iu.serialize_byte_tensor_list(
[in0])
input_byte_size = sum([
serialized_byte_size(i0)
for i0 in input_list_tmp
])
else:
input_list_tmp = [in0]
input_byte_size = sum(
[i0.nbytes for i0 in input_list_tmp])
ip_name = "ip{}".format(len(shm_ip_handles))
if _test_system_shared_memory:
shm_ip_handles.append(
shm.create_shared_memory_region(
ip_name, "/" + ip_name, input_byte_size))
shm.set_shared_memory_region(
shm_ip_handles[-1], input_list_tmp)
triton_client.register_system_shared_memory(
ip_name, "/" + ip_name, input_byte_size)
elif _test_cuda_shared_memory:
shm_ip_handles.append(
cudashm.create_shared_memory_region(
ip_name, input_byte_size, 0))
cudashm.set_shared_memory_region(
shm_ip_handles[-1], input_list_tmp)
triton_client.register_cuda_shared_memory(
ip_name,
cudashm.get_raw_handle(shm_ip_handles[-1]), 0,
input_byte_size)
inputs[0].set_shared_memory(ip_name, input_byte_size)
outputs[0].set_shared_memory("output_data",
output_byte_size)
else:
inputs[0].set_data_from_numpy(in0)
start_ms = int(round(time.time() * 1000))
if config[2]:
triton_client.async_stream_infer(
model_name,
inputs,
outputs=outputs,
sequence_id=correlation_id,
sequence_start=seq_start,
sequence_end=seq_end)
(results, error) = user_data._completed_requests.get()
if error is not None:
raise error
else:
results = triton_client.infer(
model_name,
inputs,
outputs=outputs,
sequence_id=correlation_id,
sequence_start=seq_start,
sequence_end=seq_end)
end_ms = int(round(time.time() * 1000))
# Get value of "OUTPUT", for shared memory, need to get it via
# shared memory utils
if (not _test_system_shared_memory) and (
not _test_cuda_shared_memory):
out = results.as_numpy(OUTPUT)
else:
output = results.get_output(OUTPUT)
if config[1] == "http":
output_shape = output["shape"]
else:
output_shape = output.shape
output_type = input_dtype
if _test_system_shared_memory:
out = shm.get_contents_as_numpy(
shm_op_handle, output_type, output_shape)
else:
out = cudashm.get_contents_as_numpy(
shm_op_handle, output_type, output_shape)
result = out[0] if "nobatch" in trial else out[0][0]
print("{}: {}".format(sequence_name, result))
if thresholds is not None:
lt_ms = thresholds[0]
gt_ms = thresholds[1]
if lt_ms is not None:
self.assertTrue((end_ms - start_ms) < lt_ms,
"expected less than " + str(lt_ms) +
"ms response time, got " +
str(end_ms - start_ms) + " ms")
if gt_ms is not None:
self.assertTrue(
(end_ms - start_ms) > gt_ms,
"expected greater than " + str(gt_ms) +
"ms response time, got " +
str(end_ms - start_ms) + " ms")
if delay_ms is not None:
time.sleep(delay_ms[1] / 1000.0)
seq_end_ms = int(round(time.time() * 1000))
if input_dtype == np.object_:
self.assertEqual(int(result), expected_result)
else:
self.assertEqual(result, expected_result)
if sequence_thresholds is not None:
lt_ms = sequence_thresholds[0]
gt_ms = sequence_thresholds[1]
if lt_ms is not None:
if _test_jetson:
lt_ms *= _jetson_slowdown_factor
self.assertTrue((seq_end_ms - seq_start_ms) < lt_ms,
"sequence expected less than " +
str(lt_ms) + "ms response time, got " +
str(seq_end_ms - seq_start_ms) + " ms")
if gt_ms is not None:
self.assertTrue((seq_end_ms - seq_start_ms) > gt_ms,
"sequence expected greater than " +
str(gt_ms) + "ms response time, got " +
str(seq_end_ms - seq_start_ms) + " ms")
except Exception as ex:
self.add_deferred_exception(ex)
if config[2]:
triton_client.stop_stream()
if _test_system_shared_memory or _test_cuda_shared_memory:
self.triton_client_.unregister_system_shared_memory()
self.triton_client_.unregister_cuda_shared_memory()
destroy_func = shm.destroy_shared_memory_region if _test_system_shared_memory else cudashm.destroy_shared_memory_region
destroy_func(shm_op_handle)
for shm_ip_handle in shm_ip_handles:
destroy_func(shm_ip_handle)
binary_data=True))
outputs[-1].set_shared_memory("output0_data", output_byte_size)
outputs.append(httpclient.InferRequestedOutput('OUTPUT1',
binary_data=True))
outputs[-1].set_shared_memory("output1_data", output_byte_size)
results = triton_client.infer(model_name=model_name,
inputs=inputs,
outputs=outputs)
# Read results from the shared memory.
output0 = results.get_output("OUTPUT0")
if output0 is not None:
output0_data = cudashm.get_contents_as_numpy(
shm_op0_handle, utils.triton_to_np_dtype(output0['datatype']),
output0['shape'])
else:
print("OUTPUT0 is missing in the response.")
sys.exit(1)
output1 = results.get_output("OUTPUT1")
if output1 is not None:
output1_data = cudashm.get_contents_as_numpy(
shm_op1_handle, utils.triton_to_np_dtype(output1['datatype']),
output1['shape'])
else:
print("OUTPUT1 is missing in the response.")
sys.exit(1)
for i in range(16):
def test_buffer_attributes(self):
model_name = 'bls'
# Infer
clients = [
httpclient.InferenceServerClient(url='localhost:8000'),
grpcclient.InferenceServerClient(url='localhost:8001')
]
triton_clients = [httpclient, grpcclient]
for i, client in enumerate(clients):
# To make sure no shared memory regions are registered with the
# server.
client.unregister_system_shared_memory()
client.unregister_cuda_shared_memory()
triton_client = triton_clients[i]
inputs = []
outputs = []
inputs.append(
triton_client.InferInput('INPUT0', [1, 1000], "INT32"))
input0_data = np.arange(start=0, stop=1000, dtype=np.int32)
input0_data = np.expand_dims(input0_data, axis=0)
input_byte_size = input0_data.size * input0_data.itemsize
output_byte_size = input_byte_size
shm_ip0_handle = cudashm.create_shared_memory_region(
"input0_data", input_byte_size, 0)
shm_op0_handle = cudashm.create_shared_memory_region(
"output0_data", output_byte_size, 0)
client.register_cuda_shared_memory(
"input0_data", cudashm.get_raw_handle(shm_ip0_handle), 0,
input_byte_size)
client.register_cuda_shared_memory(
"output0_data", cudashm.get_raw_handle(shm_op0_handle), 0,
input_byte_size)
cudashm.set_shared_memory_region(shm_ip0_handle, [input0_data])
inputs[0].set_shared_memory("input0_data", input_byte_size)
if triton_client is grpcclient:
outputs.append(triton_client.InferRequestedOutput('OUTPUT0'))
outputs[0].set_shared_memory("output0_data", output_byte_size)
else:
outputs.append(
triton_client.InferRequestedOutput('OUTPUT0',
binary_data=True))
outputs[0].set_shared_memory("output0_data", output_byte_size)
results = client.infer(model_name=model_name,
inputs=inputs,
outputs=outputs)
output0 = results.get_output("OUTPUT0")
self.assertIsNotNone(output0)
if triton_client is grpcclient:
output0_data = cudashm.get_contents_as_numpy(
shm_op0_handle, triton_to_np_dtype(output0.datatype),
output0.shape)
else:
output0_data = cudashm.get_contents_as_numpy(
shm_op0_handle, triton_to_np_dtype(output0['datatype']),
output0['shape'])
self.assertTrue(np.all(output0_data == input0_data))
