fix(components): Upload the tuned adapter to Model Registry instead o…

…f model checkpoint from `preview.llm.rlhf_pipeline` PiperOrigin-RevId: 591393116
kubeflow · Dec 16, 2023 · 2e2ba9e · 2e2ba9e
1 parent 075d58f
commit 2e2ba9e
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 1 deletion.
diff --git a/components/google-cloud/RELEASE.md b/components/google-cloud/RELEASE.md
@@ -1,6 +1,7 @@
 ## Upcoming release
 * Use `large_model_reference` for `model_reference_name` when uploading models from `preview.llm.rlhf_pipeline` instead of hardcoding value as `text-bison@001`.
 * Disable caching when resolving model display names for RLHF-tuned models so a unique name is generated on each `preview.llm.rlhf_pipeline` run.
+* Upload the tuned adapter to Model Registry instead of model checkpoint from `preview.llm.rlhf_pipeline`.
 
 ## Release 2.8.0
 * Release AutoSxS pipeline to preview.

diff --git a/components/google-cloud/google_cloud_pipeline_components/preview/llm/rlhf/component.py b/components/google-cloud/google_cloud_pipeline_components/preview/llm/rlhf/component.py
@@ -128,7 +128,7 @@ def rlhf_pipeline(
     )
 
   llm_model_handler = deployment_graph.pipeline(
-      output_adapter_path=rl_model_pipeline.outputs['output_model_path'],
+      output_adapter_path=rl_model_pipeline.outputs['output_adapter_path'],
       large_model_reference=large_model_reference,
       model_display_name=model_display_name,
       deploy_model=deploy_model,