Support transfomers v5

.github/workflows/nvi-ci.yml

@@ -40,7 +40,32 @@ jobs:
     - name: Run checkstyle
       run: make checkstyle
 
-  tests:
+  # tests:
+  #   runs-on: ubuntu-latest
+  #   needs: [checkstyle]
+  #   env:
+  #     MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
+  #     MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
+
+  #   steps:
+  #   - name: Checkout code
+  #     uses: actions/checkout@v6
+
+  #   - name: Set up Python
+  #     uses: actions/setup-python@v6
+  #     with:
+  #       python-version: '3.10'
+
+  #   - name: Install dependencies
+  #     run: |
+  #       python -m pip install --upgrade pip
+  #       pip install modal
+
+  #   - name: Run tests
+  #     run: |
+  #       modal run dev.modal.tests
+
+  correctness:
     runs-on: ubuntu-latest
     needs: [checkstyle]
     env:
@@ -63,15 +88,14 @@ jobs:
 
     - name: Run tests
       run: |
-        modal run dev.modal.tests
+        modal run dev.modal.test_correctness
 
-  tests-bwd:
+  convergence:
     runs-on: ubuntu-latest
     needs: [checkstyle]
     env:
       MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
       MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
-      REBUILD_IMAGE: ${{ github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
 
     steps:
     - name: Checkout code
@@ -89,4 +113,4 @@ jobs:
 
     - name: Run tests
       run: |
-        modal run dev.modal.tests_bwd
+        modal run dev.modal.test_convergence

dev/modal/tests_bwd.py → dev/modal/test_convergence.py

@@ -8,14 +8,14 @@
 
 image = modal.Image.debian_slim(python_version=PYTHON_VERSION).pip_install("uv")
 
-app = modal.App("liger_tests_bwd", image=image)
+app = modal.App("liger_tests", image=image)
 
 # mount: add local files to the remote container
 repo = image.add_local_dir(ROOT_PATH, remote_path=REMOTE_ROOT_PATH)
 
 
 @app.function(gpu="H100!", image=repo, timeout=90 * 60)
-def liger_bwd_tests():
+def liger_tests():
     import subprocess
 
     subprocess.run(
@@ -24,12 +24,4 @@ def liger_bwd_tests():
         shell=True,
         cwd=REMOTE_ROOT_PATH,
     )
-    # force install transformers==4.49.0
-    subprocess.run(
-        ["uv pip install transformers==4.49.0 --system"],
-        check=True,
-        shell=True,
-        cwd=REMOTE_ROOT_PATH,
-    )
-    subprocess.run(["make test"], check=True, shell=True, cwd=REMOTE_ROOT_PATH)
     subprocess.run(["make test-convergence"], check=True, shell=True, cwd=REMOTE_ROOT_PATH)

dev/modal/test_correctness.py

@@ -0,0 +1,27 @@
+from pathlib import Path
+
+import modal
+
+ROOT_PATH = Path(__file__).parent.parent.parent
+REMOTE_ROOT_PATH = "/root/liger-kernel"
+PYTHON_VERSION = "3.12"
+
+image = modal.Image.debian_slim(python_version=PYTHON_VERSION).pip_install("uv")
+
+app = modal.App("liger_tests", image=image)
+
+# mount: add local files to the remote container
+repo = image.add_local_dir(ROOT_PATH, remote_path=REMOTE_ROOT_PATH)
+
+
+@app.function(gpu="H100!", image=repo, timeout=90 * 60)
+def liger_tests():
+    import subprocess
+
+    subprocess.run(
+        ["uv pip install -e '.[dev]' --system"],
+        check=True,
+        shell=True,
+        cwd=REMOTE_ROOT_PATH,
+    )
+    subprocess.run(["make test"], check=True, shell=True, cwd=REMOTE_ROOT_PATH)

setup.py

@@ -32,7 +32,7 @@ def get_optional_dependencies():
     """Get optional dependency groups."""
     return {
         "dev": [
-            "transformers>=4.49.0",
+            "transformers==5.0.0rc3",
             "matplotlib>=3.7.2",
             "ruff>=0.12.0",
             "pytest>=7.1.2",