kubeflow
/
pipelines
mirror of https://github.com/kubeflow/pipelines.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix(components): Pass tuned model checkpoint to inference pipeline after RLHF tuning 

PiperOrigin-RevId: 610918020
This commit is contained in:
Googler 2024-02-27 16:53:03 -08:00 committed by Google Cloud Pipeline Components maintainers
parent 144761c948
commit 755c1f9898
2 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
components/google-cloud/RELEASE.md
View File

@ -1,5 +1,6 @@
## Upcoming release
* Add `v1.automl.forecasting.learn_to_learn_forecasting_pipeline`, `v1.automl.forecasting.sequence_to_sequence_forecasting_pipeline`, `v1.automl.forecasting.temporal_fusion_transformer_forecasting_pipeline`, `v1.automl.forecasting.time_series_dense_encoder_forecasting_pipeline` as Forecasting on Pipelines moves to GA.
* Fix bug in `preview.llm.rlhf_pipeline` that caused wrong output artifact to be used for inference after training.
## Release 2.10.0
* Fix the missing output of pipeline remote runner. `AutoMLImageTrainingJobRunOp` now passes the model artifacts correctly to downstream components.

4
components/google-cloud/google_cloud_pipeline_components/preview/llm/rlhf/component.py
View File

@ -152,7 +152,7 @@ def rlhf_pipeline(
name='Perform Inference',
):
has_model_checkpoint = function_based.value_exists(
value=rl_model_pipeline.outputs['output_adapter_path']
value=rl_model_pipeline.outputs['output_model_path']
).set_display_name('Resolve Model Checkpoint')
with kfp.dsl.Condition(
has_model_checkpoint.output == True, # pylint: disable=singleton-comparison
@ -162,7 +162,7 @@ def rlhf_pipeline(
project=project,
location=location,
large_model_reference=large_model_reference,
model_checkpoint=rl_model_pipeline.outputs['output_adapter_path'],
model_checkpoint=rl_model_pipeline.outputs['output_model_path'],
prompt_dataset=eval_dataset,
prompt_sequence_length=prompt_sequence_length,
target_sequence_length=target_sequence_length,