A client library for caikit-nlp
Install from PyPi
pip install caikit-nlp-clientA few examples follow, see example.py
To use the http protocol
from caikit_nlp_client import HttpClient
host = "localhost"
port = 8080
model_name = "flan-t5-small-caikit"
http_client = HttpClient(f"http://{host}:{port}")
text = http_client.generate_text(model_name, "What is the boiling point of Nitrogen?")To use the gRPC protocol
from caikit_nlp_client import GrpcClient
host, port= "localhost", 8085
model_name = "flan-t5-small-caikit"
grpc_client = GrpcClient(host, port, insecure=True) # plain text mode
text = grpc_client.generate_text(model_name, "What is the boiling point of Nitrogen?")Text generation methods may accept text generation parameters, which can be provided as kwargs
to generate_text and generate_text_stream.
Available values and types be retrieved as a dict for both the grpc and http clients:
for param, default_value in client.get_text_generation_parameters():
print(f"{param=}, {default_value=}")To use a self-signed certificate, assuming we have a certificate authority cert ca.pem
http_client = HttpClient(f"https://{host}:{http_port}", ca_cert_path="ca.pem")
with open("ca.pem", "rb") as fh:
ca_cert = fh.read()
grpc_client = GrpcClient(host, grpc_port, ca_cert=ca_cert)To skip certificate validation:
# http
http_client = HttpClient(f"https://{host}:{http_port}", verify=False)
# grpc
grpc_client = GrpcClient(host, port, verify=False)Assuming we have a client.pem and client-key.pem certificate files, and we require ca.pem to validate the server certificate:
# http
http_client = HttpClient(
f"https://{host}:{http_port}",
ca_cert_path="ca.pem",
client_cert_path="client.pem",
client_key_path="client-key.pem"
)
# grpc
with open("ca.pem", "rb") as fh:
ca_cert = fh.read()
with open("client.pem", "rb") as fh:
client_cert = fh.read()
with open("client-key.pem", "rb") as fh:
client_key = fh.read()
grpc_client = GrpcClient(
host,
port,
ca_cert=ca_cert,
client_key=client_key,
client_cert=client_cert,
)
# alternatively you can pass the paths directly to the client constructor
grpc_client = GrpcClient(
host,
port,
ca_cert="ca.pem",
client_cert="client.pem",
client_key="client-key.pem"
)Set up pre-commit for linting/style/misc fixes:
pip install pre-commit
pre-commit installThis project uses nox to manage test automation:
pip install nox
nox --list # list available sessions
nox --python 3.10 -s tests # run tests session for a specific python versionTests are run against a mocked instance of a TGIS backend by default. To test against a real
caikit+tgis stack, it is sufficient to run using the --real-caikit flag when running pytest:
nox -s tests -- --real-caikit testsNotes:
- The required images (
caikit-tgis-serving,text-generation-inference), so it could take a while for tests to start while compose is pulling the required images, it may seem like the tests are hanging. - This uses a real model (google/flan-t5-small), which will be downloaded
to
tests/fixtures/resources/flan-t5-small-caikitand is around 300MB in size.