run PR e2e docker CI tests in Modal (#1217) [skip ci]

* wip modal for ci * handle falcon layernorms better * update * rebuild the template each time with the pseudo-ARGS * fix ref * update tests to use modal * cleanup ci script * make sure to install jinja2 also * kickoff the gh action on gh hosted runners and specify num gpus
2024-01-26 16:13:27 -05:00
parent af29d81f80
commit 36d053f6f0
6 changed files with 141 additions and 42 deletions
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -58,10 +58,15 @@ jobs:
  docker-e2e-tests:
    if: github.repository_owner == 'OpenAccess-AI-Collective'
    # this job needs to be run on self-hosted GPU runners...
-    runs-on: [self-hosted, gpu, docker]
+    runs-on: ubuntu-latest
    timeout-minutes: 30
    needs: [pre-commit, pytest]

+    env:
+      MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
+      MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
+      MODAL_ENVIRONMENT: axolotl-ci-cd
+
    strategy:
      fail-fast: false
      matrix:
@@ -70,43 +75,29 @@ jobs:
            cuda_version: 11.8.0
            python_version: "3.10"
            pytorch: 2.0.1
+            num_gpus: 1
          - cuda: 121
            cuda_version: 12.1.0
            python_version: "3.10"
            pytorch: 2.1.2
+            num_gpus: 1
    steps:
      - name: Checkout
        uses: actions/checkout@v4
-      - name: Docker metadata
-        id: metadata
-        uses: docker/metadata-action@v5
+      - name: Install Python
+        uses: actions/setup-python@v5
        with:
-          images: winglian/axolotl-tests
-      - name: Build Docker image
+          python-version: "3.10"
+      - name: Install Modal
        run: |
-          # Set up build arguments
-          BASE_TAG="main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}"
-          CUDA="${{ matrix.cuda }}"
-          PYTORCH_VERSION="${{ matrix.pytorch }}"
-          # Build the Docker image
-          docker build . \
-            --file ./docker/Dockerfile-tests \
-            --build-arg BASE_TAG=$BASE_TAG \
-            --build-arg CUDA=$CUDA \
-            --build-arg GITHUB_REF=$GITHUB_REF \
-            --build-arg PYTORCH_VERSION=$PYTORCH_VERSION \
-            --tag ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }} \
-            --no-cache
-      - name: Unit Tests w docker image
+          python -m pip install --upgrade pip
+          pip install modal jinja2
+      - name: Update env vars
        run: |
-          docker run --rm ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }} pytest --ignore=tests/e2e/ /workspace/axolotl/tests/
-      - name: GPU Unit Tests w docker image
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
+          echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
+          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+      - name: Run training job on Modal
        run: |
-          docker run --privileged --gpus "all" --env WANDB_DISABLED=true --rm ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }} pytest --ignore=tests/e2e/patched/ /workspace/axolotl/tests/e2e/
-      - name: GPU Unit Tests monkeypatched w docker image
-        run: |
-          docker run --privileged --gpus "all" --env WANDB_DISABLED=true --rm ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }} pytest /workspace/axolotl/tests/e2e/patched/
-      - name: Prune image from docker
-        if: github.ref != 'refs/heads/main'
-        run: |
-          docker rmi -f ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+          modal run cicd.tests
--- a/cicd/Dockerfile.jinja
+++ b/cicd/Dockerfile.jinja
@@ -0,0 +1,38 @@
+FROM winglian/axolotl-base:{{ BASE_TAG }}
+
+ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
+ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
+ENV CUDA="{{ CUDA }}"
+ENV BNB_CUDA_VERSION="{{ CUDA }}"
+ENV PYTORCH_VERSION="{{ PYTORCH_VERSION }}"
+ENV GITHUB_REF="{{ GITHUB_REF }}"
+ENV GITHUB_SHA="{{ GITHUB_SHA }}"
+
+RUN apt-get update && \
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
+
+WORKDIR /workspace
+
+RUN git clone --depth=1 https://github.com/OpenAccess-AI-Collective/axolotl.git
+
+WORKDIR /workspace/axolotl
+
+RUN git fetch origin +$GITHUB_REF && \
+    git checkout FETCH_HEAD
+
+# If AXOLOTL_EXTRAS is set, append it in brackets
+RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS]; \
+    else \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm]; \
+    fi
+
+# So we can test the Docker image
+RUN pip install pytest
+
+# fix so that git fetch/pull from remote works
+RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \
+    git config --get remote.origin.fetch
+
+# helper for huggingface-login cli
+RUN git config --global credential.helper store
--- a/cicd/tests.py
+++ b/cicd/tests.py
@@ -0,0 +1,69 @@
+"""
+modal application to run axolotl gpu tests in Modal
+"""
+import os
+import pathlib
+import tempfile
+
+import jinja2
+import modal
+from jinja2 import select_autoescape
+from modal import Image, Stub
+
+cicd_path = pathlib.Path(__file__).parent.resolve()
+
+template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
+template_env = jinja2.Environment(
+    loader=template_loader, autoescape=select_autoescape()
+)
+df_template = template_env.get_template("Dockerfile.jinja")
+
+df_args = {
+    "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.0.1"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.10-cu118-2.0.1"),
+    "CUDA": os.environ.get("CUDA", "118"),
+    "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
+    "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
+}
+
+dockerfile_contents = df_template.render(**df_args)
+
+temp_dir = tempfile.mkdtemp()
+with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
+    f.write(dockerfile_contents)
+
+cicd_image = Image.from_dockerfile(
+    pathlib.Path(temp_dir) / "Dockerfile",
+    force_build=True,
+    gpu="A10G",
+).env(df_args)
+
+stub = Stub("Axolotl CI/CD", secrets=[])
+
+
+N_GPUS = int(os.environ.get("N_GPUS", 1))
+GPU_CONFIG = modal.gpu.A10G(count=N_GPUS)
+
+
+def run_cmd(cmd: str, run_folder: str):
+    import subprocess  # nosec
+
+    # Propagate errors from subprocess.
+    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
+        exit(exit_code)  # pylint: disable=consider-using-sys-exit
+
+
+@stub.function(
+    image=cicd_image,
+    gpu=GPU_CONFIG,
+    timeout=60 * 30,
+)
+def cicd_pytest():
+    cmd = "pytest /workspace/axolotl/tests/e2e/patched/"
+    run_cmd(cmd, "/workspace/axolotl")
+
+
+@stub.local_entrypoint()
+def main():
+    cicd_pytest.remote()
--- a/docker/Dockerfile-modal
+++ b/docker/Dockerfile-modal
@@ -1,14 +1,11 @@
-ARG BASE_TAG=main-base
-FROM winglian/axolotl-base:$BASE_TAG
+FROM winglian/axolotl-base:main-base

-ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
-ARG AXOLOTL_EXTRAS=""
-ARG CUDA="118"
-ENV BNB_CUDA_VERSION=$CUDA
-ARG PYTORCH_VERSION="2.0.1"
-ARG GITHUB_REF="main"
-
-ENV PYTORCH_VERSION=$PYTORCH_VERSION
+ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
+ENV AXOLOTL_EXTRAS=""
+ENV CUDA="118"
+ENV BNB_CUDA_VERSION="118"
+ENV PYTORCH_VERSION="2.0.1"
+ENV GITHUB_REF="main"

 RUN apt-get update && \
    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,4 +1,5 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
+jinja2
 packaging==23.2
 peft==0.7.1
 transformers==4.37.0
--- a/src/axolotl/utils/models.py
+++ b/src/axolotl/utils/models.py
@@ -645,7 +645,10 @@ def load_model(
    if not cfg.fsdp:
        # FSDP doesn't like mixed Float and BFloat16
        for name, module in model.named_modules():
-            if any(m in name for m in ["norm", "gate"]):
+            if (
+                any(m in name for m in ["norm", "gate"])
+                or "LayerNorm" in module.__class__.__name__
+            ):
                module.to(torch.float32)
            if model_config.model_type == "btlm":
                # don't upcast lm_head for btlm
@@ -684,7 +687,7 @@ def load_model(
    if needs_fa2_dtype or cfg.flash_attention:
        LOG.info("converting modules to %s for flash attention", cfg.torch_dtype)
        for name, module in model.named_modules():
-            if "norm" in name:
+            if "norm" in name or "LayerNorm" in module.__class__.__name__:
                module.to(cfg.torch_dtype)
            if any(m in name for m in embedding_modules):
                if hasattr(module, "weight"):