Fix installation script and improve CLI configuration (#63)

* fix installation script

* fix installation script

* fix pre-commit

* fix tests

fix install

* fix docker compose

* misc improvements

* remove options

* up version

---------

Co-authored-by: pxc <panxuchen.pxc@alibaba-inc.com>

Author: ekzhu

.github/workflows/docker/docker-compose.yml

@@ -1,9 +1,22 @@
 services:
   # use 2 nodes to simulate a cluster environment
   tuft-node-1:
-    image: tuft-unittest:20260123
+    image: nvcr.io/nvidia/cuda:12.8.1-cudnn-devel-ubuntu22.04
     pull_policy: never
-    command: bash -c "source /opt/venv/bin/activate && uv pip install -e .[dev,backend,persistence] && ray start --head --dashboard-host 0.0.0.0 --include-dashboard true --block"
+    command: bash -c "
+      chmod 1777 /tmp && apt update && apt install -y --no-install-recommends \
+      build-essential \
+      curl git wget vim tmux net-tools \
+      python3 python3-pip python3-dev python3-packaging python3-venv \
+      libomp-dev infiniband-diags libibverbs-dev librdmacm-dev rdma-core perftest \
+      && rm -rf /var/lib/apt/lists/* \
+      && ln -sf /usr/bin/python3 /usr/bin/python \
+      && ln -sf /usr/bin/pip3 /usr/bin/pip \
+      && bash /workspace/scripts/install.sh --local-source /workspace \
+      && source $HOME/.local/bin/env \
+      && source /root/.tuft/venv/bin/activate \
+      && uv pip install .[dev] \
+      && ray start --head --dashboard-host 0.0.0.0 --include-dashboard true --block"
     environment:
       - HF_ENDPOINT=https://hf-mirror.com
       - RAY_ADDRESS=auto
@@ -12,7 +25,7 @@ services:
       - TUFT_TEST_MODEL_1=/mnt/models/Qwen3-0.6B
       - TUFT_TEST_MODEL_2=/mnt/models/Qwen3-1.7B
       - TEST_REDIS_URL=redis://tuft-redis:6379
-      - VIRTUAL_ENV=/opt/venv
+      - VIRTUAL_ENV=/root/.tuft/venv
     working_dir: /workspace
     networks:
       - tuft-network
@@ -25,36 +38,7 @@ services:
         reservations:
           devices:
           - driver: nvidia
-            device_ids: ['0', '1']
-            capabilities: [gpu]
-
-  tuft-node-2:
-    image: tuft-unittest:20260123
-    pull_policy: never
-    command: bash -c "source /opt/venv/bin/activate && uv pip install -e .[dev,backend,persistence] && ray start --address=tuft-node-1:6379 --block"
-    environment:
-      - HF_ENDPOINT=https://hf-mirror.com
-      - TUFT_CHECKPOINT_DIR=/mnt/checkpoints
-      - TUFT_TEST_MODEL=/mnt/models/Qwen3-0.6B
-      - TUFT_TEST_MODEL_1=/mnt/models/Qwen3-0.6B
-      - TUFT_TEST_MODEL_2=/mnt/models/Qwen3-1.7B
-      - TEST_REDIS_URL=redis://tuft-redis:6379
-      - VIRTUAL_ENV=/opt/venv
-    working_dir: /workspace
-    volumes:
-      - tuft-volume:/mnt
-      - ../../..:/workspace
-    depends_on:
-      - tuft-node-1
-    networks:
-      - tuft-network
-    shm_size: "64G"
-    deploy:
-      resources:
-        reservations:
-          devices:
-          - driver: nvidia
-            device_ids: ['2', '3']
+            device_ids: ['0', '1', '2', '3']
             capabilities: [gpu]
 
   tuft-redis:

.github/workflows/install-script.yml

@@ -59,36 +59,42 @@ jobs:
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
-      - name: Test tuft (dry run - check config error)
+      - name: Test tuft launch --help
         run: |
           export PATH="${TUFT_HOME}/bin:$PATH"
-          # Should fail with config error, not import error
-          tuft 2>&1 | grep -q "\-\-config" || tuft 2>&1 | grep -q "config"
+          tuft launch --help
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
-      - name: Clean up installation
-        run: rm -rf "${TUFT_HOME}"
-        env:
-          TUFT_HOME: ${{ runner.temp }}/tuft
-
-  test-install-default-with-backend:
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-
-      - name: Run install script (default includes backend)
+      - name: Test tuft launch requires config
         run: |
-          bash scripts/install.sh --local-source "$GITHUB_WORKSPACE"
+          export PATH="${TUFT_HOME}/bin:$PATH"
+          # Should fail with config error when no config provided
+          if tuft launch 2>&1; then
+            echo "Expected tuft launch to fail without config"
+            exit 1
+          fi
+          # Verify error message mentions config
+          tuft launch 2>&1 | grep -qi "config"
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
-      - name: Verify backend dependencies installed
+      - name: Test tuft launch with config file
         run: |
-          "${TUFT_HOME}/venv/bin/python" -c "import peft; print('peft imported successfully')"
-          "${TUFT_HOME}/venv/bin/python" -c "import redis; print('redis imported successfully')"
+          export PATH="${TUFT_HOME}/bin:$PATH"
+          # Create a minimal config file
+          cat > "${TUFT_HOME}/configs/tuft_config.yaml" << 'EOF'
+          model_owner: test
+          supported_models:
+            - model_name: test-model
+              model_path: /nonexistent/path
+              max_model_len: 1024
+          authorized_users:
+            test-key: test-user
+          EOF
+          # Launch should fail due to missing model, but get past config validation
+          # We just verify it doesn't fail on config parsing
+          tuft launch 2>&1 | grep -v "Configuration file must be provided" || true
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
@@ -97,25 +103,23 @@ jobs:
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
-  test-install-without-backend:
+  test-backend-dependencies:
     runs-on: ubuntu-latest
 
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Run install script without backend
+      - name: Run install script
         run: |
-          bash scripts/install.sh --local-source "$GITHUB_WORKSPACE" --without-backend
+          bash scripts/install.sh --local-source "$GITHUB_WORKSPACE"
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
-      - name: Verify minimal install (no peft)
+      - name: Verify backend dependencies installed
         run: |
-          # peft should NOT be installed in minimal mode
-          "${TUFT_HOME}/venv/bin/python" -c "import peft" 2>&1 && exit 1 || echo "peft not installed (expected)"
-          # tuft should still be importable
-          "${TUFT_HOME}/venv/bin/python" -c "import tuft; print('tuft imported successfully')"
+          "${TUFT_HOME}/venv/bin/python" -c "import peft; print('peft imported successfully')"
+          "${TUFT_HOME}/venv/bin/python" -c "import redis; print('redis imported successfully')"
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
@@ -150,13 +154,20 @@ jobs:
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
-      - name: Test upgrade command
+      - name: Test upgrade command (from PyPI)
         run: |
           export PATH="${TUFT_HOME}/bin:$PATH"
           tuft upgrade
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
+      - name: Test upgrade command (from local source)
+        run: |
+          export PATH="${TUFT_HOME}/bin:$PATH"
+          tuft upgrade --local-source "$GITHUB_WORKSPACE"
+        env:
+          TUFT_HOME: ${{ runner.temp }}/tuft
+
       - name: Clean up installation
         run: rm -rf "${TUFT_HOME}"
         env:
@@ -171,7 +182,7 @@ jobs:
 
       - name: Initial install
         run: |
-          bash scripts/install.sh --local-source "$GITHUB_WORKSPACE" --without-backend
+          bash scripts/install.sh --local-source "$GITHUB_WORKSPACE"
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
@@ -184,7 +195,7 @@ jobs:
 
       - name: Reinstall with --clean
         run: |
-          bash scripts/install.sh --local-source "$GITHUB_WORKSPACE" --without-backend --clean
+          bash scripts/install.sh --local-source "$GITHUB_WORKSPACE" --clean
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
@@ -202,36 +213,30 @@ jobs:
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
-  test-install-backend-command:
+  test-upgrade-from-source:
     runs-on: ubuntu-latest
 
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Install without backend first
-        run: |
-          bash scripts/install.sh --local-source "$GITHUB_WORKSPACE" --without-backend
-        env:
-          TUFT_HOME: ${{ runner.temp }}/tuft
-
-      - name: Verify peft is NOT installed
-        run: |
-          "${TUFT_HOME}/venv/bin/python" -c "import peft" 2>&1 && exit 1 || echo "peft not installed (expected)"
+      - name: Install tuft
+        run: bash scripts/install.sh --local-source "$GITHUB_WORKSPACE"
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
-      - name: Run install-backend command
+      - name: Test upgrade --from-source
         run: |
           export PATH="${TUFT_HOME}/bin:$PATH"
-          tuft install-backend
+          tuft upgrade --from-source
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 
-      - name: Verify backend dependencies now installed
+      - name: Verify tuft still works after upgrade
         run: |
-          "${TUFT_HOME}/venv/bin/python" -c "import peft; print('peft imported successfully')"
-          "${TUFT_HOME}/venv/bin/python" -c "import redis; print('redis imported successfully')"
+          export PATH="${TUFT_HOME}/bin:$PATH"
+          tuft version
+          tuft launch --help
         env:
           TUFT_HOME: ${{ runner.temp }}/tuft
 

.github/workflows/unittest.yml

@@ -34,8 +34,7 @@ jobs:
         MAX_RETRIES=20
         RETRY_INTERVAL=5
         for i in $(seq 1 $MAX_RETRIES); do
-          if docker compose exec tuft-node-1 bash -c "source /opt/venv/bin/activate && ray status" \
-            && docker compose exec tuft-node-2 bash -c "source /opt/venv/bin/activate && ray status"; then
+          if docker compose exec tuft-node-1 bash -c "source /root/.tuft/venv/bin/activate && ray status"; then
             break
           fi
           echo "Waiting for ray cluster to be ready... ($i/$MAX_RETRIES)"
@@ -51,7 +50,7 @@ jobs:
       # set a github env variable to indicate tests were run, so that subsequent steps can check it
       run: |
         echo "tests_run=true" >> $GITHUB_ENV
-        docker compose exec tuft-node-1 bash -c "source /opt/venv/bin/activate && pytest tests -v -s --gpu --basetemp /mnt/checkpoints --ctrf report.json"
+        docker compose exec tuft-node-1 bash -c "source /root/.tuft/venv/bin/activate && pytest tests -v -s --gpu --basetemp /mnt/checkpoints --ctrf report.json"
 
     - name: Convert report.json time to ms
       working-directory: tuft-${{ github.run_id }}

docker/Dockerfile

@@ -7,7 +7,7 @@
 #
 # Note:
 # This Dockerfile uses 'uv' to create a virtual environment for better package management.
-# The uv virtual environment is created at `/opt/venv`, use `source /opt/venv/bin/activate` to activate it.
+# The uv virtual environment is created at `/root/.tuft/venv`, use `source /root/.tuft/venv/bin/activate` to activate it.
 # Make sure to use `uv pip` to install packages within the virtual environment.
 
 FROM nvcr.io/nvidia/cuda:12.8.1-cudnn-devel-ubuntu22.04
@@ -23,28 +23,26 @@ RUN chmod 1777 /tmp && apt update && apt install -y --no-install-recommends \
     && ln -sf /usr/bin/python3 /usr/bin/python \
     && ln -sf /usr/bin/pip3 /usr/bin/pip
 
-ENV VIRTUAL_ENV=/opt/venv
+ENV VIRTUAL_ENV=/root/.tuft/venv
 
 # copy the TuFT dir into the workspace
 COPY ./pyproject.toml .
 COPY ./LICENSE .
 COPY ./README.md .
 COPY ./src ./src
+COPY ./scripts ./scripts
 
 #  Uncomment the following line if you want to use AliCloud Mirror to speed up pip install
 # ENV UV_DEFAULT_INDEX=http://mirrors.cloud.aliyuncs.com/pypi/simple/
 
 # Uncomment the following line to use a Hugging Face mirror if you have network connection problem with Hugging Face
 # ENV HF_ENDPOINT=https://hf-mirror.com
 
-# Install uv
-RUN pip install uv && uv venv ${VIRTUAL_ENV} --python=python3.12
-
-# Install minimal TuFT
-RUN . ${VIRTUAL_ENV}/bin/activate && uv pip install -e .[dev,backend,persistence]
-
-# Install flash_attn
-RUN . ${VIRTUAL_ENV}/bin/activate && uv pip install flash_attn==2.8.1 --no-build-isolation
+# Install
+RUN bash ./scripts/install.sh --local-source /workspace \
+    && . $HOME/.local/bin/env \
+    && . /root/.tuft/venv/bin/activate \
+    && uv pip install .[dev]
 
 ENTRYPOINT ["/bin/bash", "-c", "source ${VIRTUAL_ENV}/bin/activate && exec \"$@\"", "--"]
 CMD ["bash"]

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "tuft"
-version = "0.1.1"
+version = "0.1.2"
 description = "A multi-tenant fine-tuning platform for LLMs with Tinker-compatible API"
 authors = [
     { name = "TuFT Developers", email = "tuft@list.alibaba-inc.com" }
@@ -24,7 +24,7 @@ dependencies = [
     "opentelemetry-instrumentation-fastapi>=0.41b0",
     "opentelemetry-instrumentation-logging>=0.41b0",
     "psutil>=5.9.0",
-    "pynvml>=11.5.0",
+    "nvidia-ml-py>=13.0.0",
 ]
 
 [project.scripts]