Merge pull request #53 from mindvalley/chore/optimize-model-build-workflow

remove some models to save memory during build

Author: onimsha

.github/workflows/gar-build-push-model-server-container-on-tag.yml

@@ -45,7 +45,7 @@ jobs:
         id: metadata
         with:
           images: |
-            us-docker.pkg.dev/${{ env.GarProjectID }}/${{ env.GarRepo }}/${{ env.GarImageName }}:${{ github.ref_name }}
+            us-docker.pkg.dev/${{ env.GarProjectID }}/${{ env.GarRepo }}/${{ env.GarImageName }}:${{ github.sha }}
 
       - name: Model Server Image Docker Build and Push
         uses: int128/kaniko-action@v1
@@ -60,7 +60,7 @@ jobs:
             --snapshot-mode=redo
             --cache-ttl=730h
           tags: |
-            us-docker.pkg.dev/${{ env.GarProjectID }}/${{ env.GarRepo }}/${{ env.GarImageName }}:${{ github.ref_name }}
+            us-docker.pkg.dev/${{ env.GarProjectID }}/${{ env.GarRepo }}/${{ env.GarImageName }}:${{ github.sha }}
             us-docker.pkg.dev/${{ env.GarProjectID }}/${{ env.GarRepo }}/${{ env.GarImageName }}:latest
           build-args: |
-            DANSWER_VERSION=${{ github.ref_name }}
+            DANSWER_VERSION=${{ github.sha }}

backend/Dockerfile.model_server

@@ -21,19 +21,16 @@ RUN apt-get remove -y --allow-remove-essential perl-base && \
 # Download tokenizers, distilbert for the Danswer model
 # Download model weights
 # Run Nomic to pull in the custom architecture and have it cached locally
-RUN python -c "from transformers import AutoTokenizer; \
-AutoTokenizer.from_pretrained('distilbert-base-uncased'); \
-AutoTokenizer.from_pretrained('mixedbread-ai/mxbai-rerank-xsmall-v1'); \
-from huggingface_hub import snapshot_download; \
-snapshot_download(repo_id='danswer/hybrid-intent-token-classifier', revision='v1.0.3'); \
-snapshot_download('nomic-ai/nomic-embed-text-v1'); \
-snapshot_download('mixedbread-ai/mxbai-rerank-xsmall-v1'); \
-from sentence_transformers import SentenceTransformer; \
-SentenceTransformer(model_name_or_path='nomic-ai/nomic-embed-text-v1', trust_remote_code=True);"
+# RUN python -c "from transformers import AutoTokenizer; \
+# from huggingface_hub import snapshot_download; \
+# snapshot_download(repo_id='danswer/hybrid-intent-token-classifier', revision='v1.0.3'); \
+# snapshot_download('nomic-ai/nomic-embed-text-v1'); \
+# from sentence_transformers import SentenceTransformer; \
+# SentenceTransformer(model_name_or_path='nomic-ai/nomic-embed-text-v1', trust_remote_code=True);"
 
 # In case the user has volumes mounted to /root/.cache/huggingface that they've downloaded while
 # running Danswer, don't overwrite it with the built in cache folder
-RUN mv /root/.cache/huggingface /root/.cache/temp_huggingface
+# RUN mv /root/.cache/huggingface /root/.cache/temp_huggingface
 
 WORKDIR /app