fix: update deps for evaluate

Author: terryyz

Docker/Evaluate.Dockerfile

@@ -16,7 +16,7 @@ RUN rm -rf /bigcodebench
 ADD "https://api.github.com/repos/bigcode-project/bigcodebench/commits?per_page=1" latest_commit
 RUN git clone https://github.com/bigcode-project/bigcodebench.git /bigcodebench
 
-RUN cd /bigcodebench && pip install .
+RUN cd /bigcodebench && pip install .[evaluate] --no-deps
 
 # Pre-install the dataset
 RUN python3 -c "from bigcodebench.data import get_bigcodebench; get_bigcodebench()"

Docker/Gradio.Dockerfile

@@ -18,7 +18,7 @@ RUN rm -rf /bigcodebench
 ADD "https://api.github.com/repos/bigcode-project/bigcodebench/commits?per_page=1" latest_commit
 RUN git clone https://github.com/bigcode-project/bigcodebench.git /bigcodebench
 
-RUN cd /bigcodebench && pip install .
+RUN cd /bigcodebench && pip install .[evaluate] --no-deps
 
 # Pre-install the dataset
 RUN python3 -c "from bigcodebench.data import get_bigcodebench; get_bigcodebench()"

release_docker.sh

@@ -25,11 +25,6 @@ docker tag bigcodebench/bigcodebench-evaluate:$version bigcodebench/bigcodebench
 docker push bigcodebench/bigcodebench-evaluate:$version
 docker push bigcodebench/bigcodebench-evaluate:latest
 
-docker build -f Docker/Generate.Dockerfile . -t bigcodebench/bigcodebench-generate:$version
-docker tag bigcodebench/bigcodebench-generate:$version bigcodebench/bigcodebench-generate:latest
-docker push bigcodebench/bigcodebench-generate:$version
-docker push bigcodebench/bigcodebench-generate:latest
-
 docker build -f Docker/Gradio.Dockerfile . -t bigcodebench/bigcodebench-gradio:$version
 docker tag bigcodebench/bigcodebench-gradio:$version bigcodebench/bigcodebench-gradio:latest
 docker push bigcodebench/bigcodebench-gradio:$version

setup.cfg

@@ -38,6 +38,24 @@ install_requires =
     mistralai>=0.2.0,<1.0.0
     openai>=1.11.1
 
+[options.extras_require]
+# Define optional dependencies (grouped by extras)
+evaluate =
+    appdirs>=1.4.4
+    fire>=0.6.0
+    multipledispatch>=0.6.0
+    pqdm>=0.2.0
+    tempdir>=0.7.1
+    termcolor>=2.0.0
+    tqdm>=4.56.0
+    tree_sitter_languages>=1.10.2
+    tree-sitter==0.21.3
+    wget>=3.2
+    datasets
+    gradio-client
+    numpy
+    rich
+
 [options.entry_points]
 console_scripts =
     bigcodebench.evaluate = bigcodebench.evaluate:main