Skip to content

Commit 3eda55d

Browse files
vertex-sdk-botcopybara-github
vertex-sdk-bot
authored and
copybara-github
committed
feat: tune_model method for class TextEmbeddingModel.
PiperOrigin-RevId: 627266791
1 parent 415912e commit 3eda55d

File tree

2 files changed

+492
-78
lines changed

2 files changed

+492
-78
lines changed

‎tests/unit/aiplatform/test_language_models.py

+201-1
Original file line numberDiff line numberDiff line change
@@ -101,7 +101,18 @@
101101
"prediction_schema_uri": "gs://google-cloud-aiplatform/schema/predict/prediction/text_generation_1.0.0.yaml",
102102
},
103103
}
104-
104+
_TEXT_GECKO_PUBLISHER_MODEL_DICT = {
105+
"name": "publishers/google/models/textembedding-gecko",
106+
"version_id": "003",
107+
"open_source_category": "PROPRIETARY",
108+
"launch_stage": gca_publisher_model.PublisherModel.LaunchStage.GA,
109+
"publisher_model_template": "projects/{user-project}/locations/{location}/publishers/google/models/textembedding-gecko@003",
110+
"predict_schemata": {
111+
"instance_schema_uri": "gs://google-cloud-aiplatform/schema/predict/instance/text_embedding_1.0.0.yaml",
112+
"parameters_schema_uri": "gs://google-cloud-aiplatfrom/schema/predict/params/text_embedding_1.0.0.yaml",
113+
"prediction_schema_uri": "gs://google-cloud-aiplatform/schema/predict/prediction/text_embedding_1.0.0.yaml",
114+
},
115+
}
105116
_CHAT_BISON_PUBLISHER_MODEL_DICT = {
106117
"name": "publishers/google/models/chat-bison",
107118
"version_id": "001",
@@ -528,6 +539,105 @@ def reverse_string_2(s):""",
528539
},
529540
)
530541

542+
_EMBEDING_MODEL_TUNING_PIPELINE_SPEC = {
543+
"components": {},
544+
"deploymentSpec": {},
545+
"pipelineInfo": {
546+
"description": "Pipeline definition for v1.1.x embedding tuning pipelines.",
547+
"name": "tune-text-embedding-model",
548+
},
549+
"root": {
550+
"dag": {"tasks": {}},
551+
"inputDefinitions": {
552+
"parameters": {
553+
"accelerator_count": {
554+
"defaultValue": 4,
555+
"description": "how many accelerators to use when running the\ncontainer.",
556+
"isOptional": True,
557+
"parameterType": "NUMBER_INTEGER",
558+
},
559+
"accelerator_type": {
560+
"defaultValue": "NVIDIA_TESLA_V100",
561+
"description": "the accelerator type for running the trainer component.",
562+
"isOptional": True,
563+
"parameterType": "STRING",
564+
},
565+
"base_model_version_id": {
566+
"defaultValue": "textembedding-gecko@001",
567+
"description": "which base model to tune. This may be any stable\nnumbered version, for example `textembedding-gecko@001`.",
568+
"isOptional": True,
569+
"parameterType": "STRING",
570+
},
571+
"batch_size": {
572+
"defaultValue": 128,
573+
"description": "training batch size.",
574+
"isOptional": True,
575+
"parameterType": "NUMBER_INTEGER",
576+
},
577+
"corpus_path": {
578+
"description": "the GCS path to the corpus data location.",
579+
"parameterType": "STRING",
580+
},
581+
"iterations": {
582+
"defaultValue": 1000,
583+
"description": "the number of steps to perform fine-tuning.",
584+
"isOptional": True,
585+
"parameterType": "NUMBER_INTEGER",
586+
},
587+
"location": {
588+
"defaultValue": "us-central1",
589+
"description": "GCP region to run the pipeline.",
590+
"isOptional": True,
591+
"parameterType": "STRING",
592+
},
593+
"machine_type": {
594+
"defaultValue": "n1-standard-16",
595+
"description": "the type of the machine to run the trainer component. For\nmore details about this input config, see:\nhttps://cloud.google.com/vertex-ai/docs/training/configure-compute.",
596+
"isOptional": True,
597+
"parameterType": "STRING",
598+
},
599+
"model_display_name": {
600+
"defaultValue": "tuned-text-embedding-model",
601+
"description": "output model display name.",
602+
"isOptional": True,
603+
"parameterType": "STRING",
604+
},
605+
"project": {
606+
"description": "user's project id.",
607+
"parameterType": "STRING",
608+
},
609+
"queries_path": {
610+
"description": "the GCS path to the queries location.",
611+
"parameterType": "STRING",
612+
},
613+
"task_type": {
614+
"defaultValue": "DEFAULT",
615+
"description": "the task type expected to be used during inference. Valid\nvalues are `DEFAULT`, `RETRIEVAL_QUERY`, `RETRIEVAL_DOCUMENT`,\n`SEMANTIC_SIMILARITY`, `CLASSIFICATION`, and `CLUSTERING`.",
616+
"isOptional": True,
617+
"parameterType": "STRING",
618+
},
619+
"test_label_path": {
620+
"defaultValue": "",
621+
"description": "the GCS path to the test label data location.",
622+
"isOptional": True,
623+
"parameterType": "STRING",
624+
},
625+
"train_label_path": {
626+
"description": "the GCS path to the train label data location.",
627+
"parameterType": "STRING",
628+
},
629+
"validation_label_path": {
630+
"defaultValue": "",
631+
"description": "The GCS path to the validation label data location.",
632+
"isOptional": True,
633+
"parameterType": "STRING",
634+
},
635+
}
636+
},
637+
},
638+
"schemaVersion": "2.1.0",
639+
"sdkVersion": "kfp-2.6.0",
640+
}
531641
_TEST_PIPELINE_SPEC = {
532642
"components": {},
533643
"pipelineInfo": {"name": "evaluation-llm-text-generation-pipeline"},
@@ -641,6 +751,9 @@ def reverse_string_2(s):""",
641751
}
642752

643753

754+
_EMBEDING_MODEL_TUNING_PIPELINE_SPEC_JSON = json.dumps(
755+
_EMBEDING_MODEL_TUNING_PIPELINE_SPEC,
756+
)
644757
_TEST_PIPELINE_SPEC_JSON = json.dumps(
645758
_TEST_PIPELINE_SPEC,
646759
)
@@ -1460,6 +1573,18 @@ def mock_request_urlopen(request: str) -> Tuple[str, mock.MagicMock]:
14601573
yield request.param, mock_urlopen
14611574

14621575

1576+
@pytest.fixture
1577+
def mock_request_urlopen_gecko(request: str) -> Tuple[str, mock.MagicMock]:
1578+
data = _EMBEDING_MODEL_TUNING_PIPELINE_SPEC
1579+
with mock.patch.object(urllib_request, "urlopen") as mock_urlopen:
1580+
mock_read_response = mock.MagicMock()
1581+
mock_decode_response = mock.MagicMock()
1582+
mock_decode_response.return_value = json.dumps(data)
1583+
mock_read_response.return_value.decode = mock_decode_response
1584+
mock_urlopen.return_value.read = mock_read_response
1585+
yield request.param, mock_urlopen
1586+
1587+
14631588
@pytest.fixture
14641589
def mock_request_urlopen_rlhf(request: str) -> Tuple[str, mock.MagicMock]:
14651590
data = _TEST_RLHF_PIPELINE_SPEC
@@ -1528,6 +1653,21 @@ def get_endpoint_mock():
15281653
yield get_endpoint_mock
15291654

15301655

1656+
@pytest.fixture
1657+
def mock_get_tuned_embedding_model(get_endpoint_mock):
1658+
with mock.patch.object(
1659+
_language_models._TunableEmbeddingModelMixin, "get_tuned_model"
1660+
) as mock_text_generation_model:
1661+
mock_text_generation_model.return_value._model_id = (
1662+
test_constants.ModelConstants._TEST_MODEL_RESOURCE_NAME
1663+
)
1664+
mock_text_generation_model.return_value._endpoint_name = (
1665+
test_constants.EndpointConstants._TEST_ENDPOINT_NAME
1666+
)
1667+
mock_text_generation_model.return_value._endpoint = get_endpoint_mock
1668+
yield mock_text_generation_model
1669+
1670+
15311671
@pytest.fixture
15321672
def mock_get_tuned_model(get_endpoint_mock):
15331673
with mock.patch.object(
@@ -2134,6 +2274,66 @@ def test_text_generation_response_repr(self):
21342274
assert "blocked" in response_repr
21352275
assert "Violent" in response_repr
21362276

2277+
@pytest.mark.parametrize(
2278+
"job_spec",
2279+
[_EMBEDING_MODEL_TUNING_PIPELINE_SPEC_JSON],
2280+
)
2281+
@pytest.mark.parametrize(
2282+
"mock_request_urlopen_gecko",
2283+
["https://us-central1-kfp.pkg.dev/proj/repo/pack/latest"],
2284+
indirect=True,
2285+
)
2286+
def test_tune_text_embedding_model(
2287+
self,
2288+
mock_pipeline_service_create,
2289+
mock_pipeline_job_get,
2290+
mock_pipeline_bucket_exists,
2291+
job_spec,
2292+
mock_load_yaml_and_json,
2293+
mock_gcs_from_string,
2294+
mock_gcs_upload,
2295+
mock_request_urlopen_gecko,
2296+
mock_get_tuned_embedding_model,
2297+
):
2298+
"""Tests tuning the text embedding model."""
2299+
aiplatform.init(
2300+
project=_TEST_PROJECT,
2301+
location=_TEST_LOCATION,
2302+
encryption_spec_key_name=_TEST_ENCRYPTION_KEY_NAME,
2303+
)
2304+
with mock.patch.object(
2305+
target=model_garden_service_client.ModelGardenServiceClient,
2306+
attribute="get_publisher_model",
2307+
return_value=gca_publisher_model.PublisherModel(
2308+
_TEXT_GECKO_PUBLISHER_MODEL_DICT
2309+
),
2310+
):
2311+
model = language_models.TextEmbeddingModel.from_pretrained(
2312+
"textembedding-gecko@003"
2313+
)
2314+
tuning_job = model.tune_model(
2315+
training_data="gs://bucket/training.tsv",
2316+
corpus_data="gs://bucket/corpus.jsonl",
2317+
queries_data="gs://bucket/queries.jsonl",
2318+
test_data="gs://bucket/test.tsv",
2319+
tuned_model_location="us-central1",
2320+
train_steps=10,
2321+
accelerator="NVIDIA_TESLA_A100",
2322+
)
2323+
call_kwargs = mock_pipeline_service_create.call_args[1]
2324+
pipeline_arguments = call_kwargs[
2325+
"pipeline_job"
2326+
].runtime_config.parameter_values
2327+
assert pipeline_arguments["iterations"] == 10
2328+
assert pipeline_arguments["accelerator_type"] == "NVIDIA_TESLA_A100"
2329+
2330+
# Testing the tuned model
2331+
tuned_model = tuning_job.get_tuned_model()
2332+
assert (
2333+
tuned_model._endpoint_name
2334+
== test_constants.EndpointConstants._TEST_ENDPOINT_NAME
2335+
)
2336+
21372337
@pytest.mark.parametrize(
21382338
"job_spec",
21392339
[_TEST_PIPELINE_SPEC_JSON, _TEST_PIPELINE_JOB],

0 commit comments

Comments
 (0)