feat: inference API cleanup

.github/workflows/tests.yml

@@ -38,6 +38,10 @@ jobs:
         run: |
           make unit-test
 
+      - name: Run inference api unit tests
+        run: |
+          make inference-api-e2e
+
       - name: Upload Codecov report
         uses: codecov/codecov-action@v4
         with:

Makefile

@@ -83,6 +83,10 @@ unit-test: ## Run unit tests.
 	go test -v $(shell go list ./pkg/... ./api/... | grep -v /vendor) -race -coverprofile=coverage.txt -covermode=atomic
 	go tool cover -func=coverage.txt
 
+inference-api-e2e: 
+	pip install -r presets/inference/text-generation/requirements.txt
+	pytest -o log_cli=true -o log_cli_level=INFO .
+
 $(E2E_TEST):
 	(cd test/e2e && go test -c . -o $(E2E_TEST))
 

docker/presets/tfs-onnx/Dockerfile

@@ -17,7 +17,7 @@ RUN echo $VERSION > /workspace/tfs/version.txt
 COPY kaito/presets/inference/${MODEL_TYPE}/requirements.txt /workspace/tfs/requirements.txt
 RUN pip install --no-cache-dir -r requirements.txt
 
-COPY kaito/presets/inference/${MODEL_TYPE}/inference-api.py /workspace/tfs/inference-api.py
+COPY kaito/presets/inference/${MODEL_TYPE}/inference_api.py /workspace/tfs/inference_api.py
 
 # Convert to ONNX Runtime
 # RUN python convert_to_onnx.py ${MODEL_NAME} 

docker/presets/tfs/Dockerfile

@@ -1,4 +1,4 @@
-FROM python:3.8-slim
+FROM python:3.10-slim
 
 ARG WEIGHTS_PATH
 ARG MODEL_TYPE
@@ -16,7 +16,7 @@ RUN echo $VERSION > /workspace/tfs/version.txt
 COPY kaito/presets/inference/${MODEL_TYPE}/requirements.txt /workspace/tfs/requirements.txt
 RUN pip install --no-cache-dir -r requirements.txt
 
-COPY kaito/presets/inference/${MODEL_TYPE}/inference-api.py /workspace/tfs/inference-api.py
+COPY kaito/presets/inference/${MODEL_TYPE}/inference_api.py /workspace/tfs/inference_api.py
 
 # Copy the entire model weights to the weights directory
 COPY ${WEIGHTS_PATH} /workspace/tfs/weights