liger-kernel-nightly 0.5.10.dev20250611191801__tar.gz → 0.5.10.dev20250613192702__tar.gz

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.10.dev20250611191801
+Version: 0.5.10.dev20250613192702
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -45,6 +45,7 @@ Requires-Dist: datasets>=2.19.2; extra == "dev"
 Requires-Dist: seaborn; extra == "dev"
 Requires-Dist: mkdocs; extra == "dev"
 Requires-Dist: mkdocs-material; extra == "dev"
+Requires-Dist: torchvision>=0.20; extra == "dev"
 
 <a name="readme-top"></a>
 

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_cpo_loss.py

@@ -147,7 +147,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_fused_linear_cpo_loss,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_cross_entropy.py

@@ -109,7 +109,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_cross_entropy,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_distill_jsd_loss.py

@@ -248,7 +248,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_jsd_loss,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_dpo_loss.py

@@ -166,7 +166,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_dpo_loss,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_embedding.py

@@ -112,7 +112,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_embedding,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_fused_linear_cross_entropy.py

@@ -156,7 +156,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_fused_linear_cross_entropy,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_fused_linear_jsd.py

@@ -246,7 +246,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_fused_linear_jsd,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_jsd.py

@@ -143,7 +143,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_jsd,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_args,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_kl_div.py

@@ -110,7 +110,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_kldiv,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_args,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_kto_loss.py

@@ -299,7 +299,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_kto_loss,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_layer_norm.py

@@ -111,7 +111,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_layer_norm,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "backward", "full"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_orpo_loss.py

@@ -149,7 +149,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_fused_linear_orpo_loss,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "full", "backward"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_simpo_loss.py

@@ -147,7 +147,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_fused_linear_simpo_loss,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "full", "backward"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_softmax.py

@@ -124,7 +124,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_softmax,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "full", "backward"],
         metric_name="speed",
         metric_unit="ms",
         overwrite=args.overwrite,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_swiglu.py

@@ -161,7 +161,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_swiglu,
-        kernel_operation_modes=["forward"],
+        kernel_operation_modes=["forward", "full", "backward"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/benchmark_tvd.py

@@ -126,7 +126,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_tvd,
-        kernel_operation_modes=["forward", "full"],
+        kernel_operation_modes=["forward", "full", "backward"],
         metric_name="speed",
         metric_unit="ms",
         **common_args,

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/benchmark/scripts/utils.py

@@ -235,7 +235,7 @@ def update_benchmark_data_csv(
                     pass
             else:
                 existing_data_dict[row_key] = row_dict
-
+    os.makedirs(os.path.dirname(filename_abs_path), exist_ok=True)
     with open(filename_abs_path, mode="w", newline="") as file:
         writer = csv.DictWriter(file, fieldnames=fieldnames)
         writer.writeheader()

liger_kernel_nightly-0.5.10.dev20250613192702/dev/modal/benchmarks.py

@@ -0,0 +1,73 @@
+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_benchmarks", 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=60 * 45)
+def liger_benchmarks():
+    import subprocess
+    import os
+
+    subprocess.run(
+        ["uv pip install -e '.[dev]' --system"],
+        check=True,
+        shell=True,
+        cwd=REMOTE_ROOT_PATH,
+    )
+    subprocess.run(["make run-benchmarks"], check=True, shell=True, cwd=REMOTE_ROOT_PATH)
+
+    file_path = Path(REMOTE_ROOT_PATH) / "benchmark" / "data" / "all_benchmark_data.csv"
+    print(f"Checking if file exists at: {file_path}")
+    print(f"File exists: {os.path.exists(file_path)}")
+    
+    if not os.path.exists(file_path):
+        print("Listing directory contents:")
+        data_dir = file_path.parent
+        if os.path.exists(data_dir):
+            print(f"Contents of {data_dir}:")
+            print(os.listdir(data_dir))
+        else:
+            print(f"Data directory {data_dir} does not exist")
+        raise FileNotFoundError(f"Benchmark data file not found at {file_path}")
+
+    with open(file_path, "rb") as f:
+        data = f.read()
+        print(f"Successfully read {len(data)} bytes of data")
+        return data
+
+
+@app.local_entrypoint()
+def main():
+    try:
+        # Run the benchmarks and get the data
+        print("Starting benchmark run...")
+        benchmark_data = liger_benchmarks.remote()
+        
+        if not benchmark_data:
+            raise ValueError("No data received from remote function")
+            
+        # Save the data locally
+        local_data_path = ROOT_PATH / "benchmark" / "data" / "all_benchmark_data.csv"
+        print(f"Attempting to save data to: {local_data_path}")
+        
+        local_data_path.parent.mkdir(parents=True, exist_ok=True)
+        
+        with open(local_data_path, "wb") as f:
+            f.write(benchmark_data)
+        
+        print(f"Successfully saved {len(benchmark_data)} bytes to: {local_data_path}")
+        
+    except Exception as e:
+        print(f"Error occurred: {str(e)}")
+        raise

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.10.dev20250611191801"
+version = "0.5.10.dev20250613192702"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/setup.py

@@ -43,6 +43,7 @@ def get_optional_dependencies():
             "seaborn",
             "mkdocs",
             "mkdocs-material",
+            "torchvision>=0.20",
         ]
     }
 

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.10.dev20250611191801
+Version: 0.5.10.dev20250613192702
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -45,6 +45,7 @@ Requires-Dist: datasets>=2.19.2; extra == "dev"
 Requires-Dist: seaborn; extra == "dev"
 Requires-Dist: mkdocs; extra == "dev"
 Requires-Dist: mkdocs-material; extra == "dev"
+Requires-Dist: torchvision>=0.20; extra == "dev"
 
 <a name="readme-top"></a>
 

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/src/liger_kernel_nightly.egg-info/requires.txt

@@ -14,3 +14,4 @@ datasets>=2.19.2
 seaborn
 mkdocs
 mkdocs-material
+torchvision>=0.20

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/test/convergence/bf16/test_mini_models_multimodal.py

@@ -20,6 +20,7 @@ from test.utils import FAKE_CONFIGS_PATH
 from test.utils import UNTOKENIZED_DATASET_PATH
 from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
+from test.utils import is_torchvision_available
 from test.utils import load_image_processing_config
 from test.utils import load_processor_config
 from test.utils import load_tokenizer_config
@@ -788,6 +789,7 @@ def run_mini_model_multimodal(
                     not QWEN2_VL_AVAILABLE,
                     reason="Qwen2-VL not available in this version of transformers",
                 ),
+                pytest.mark.skipif(not is_torchvision_available(), reason="Qwen2VLVideoProcessor requires torchvision"),
             ],
         ),
         pytest.param(
@@ -826,6 +828,7 @@ def run_mini_model_multimodal(
                     not QWEN2_5_VL_AVAILABLE,
                     reason="Qwen2.5-VL not available in this version of transformers",
                 ),
+                pytest.mark.skipif(not is_torchvision_available(), reason="Qwen2VLVideoProcessor requires torchvision"),
             ],
         ),
         pytest.param(

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/test/convergence/fp32/test_mini_models_multimodal.py

@@ -20,6 +20,7 @@ from test.utils import FAKE_CONFIGS_PATH
 from test.utils import UNTOKENIZED_DATASET_PATH
 from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
+from test.utils import is_torchvision_available
 from test.utils import load_image_processing_config
 from test.utils import load_processor_config
 from test.utils import load_tokenizer_config
@@ -783,6 +784,7 @@ def run_mini_model_multimodal(
                     not QWEN2_VL_AVAILABLE,
                     reason="Qwen2-VL not available in this version of transformers",
                 ),
+                pytest.mark.skipif(not is_torchvision_available(), reason="Qwen2VLVideoProcessor requires torchvision"),
             ],
         ),
         pytest.param(
@@ -817,6 +819,7 @@ def run_mini_model_multimodal(
                     not QWEN2_5_VL_AVAILABLE,
                     reason="Qwen2.5-VL not available in this version of transformers",
                 ),
+                pytest.mark.skipif(not is_torchvision_available(), reason="Qwen2VLVideoProcessor requires torchvision"),
             ],
         ),
         pytest.param(

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/test/utils.py

@@ -228,6 +228,13 @@ def supports_bfloat16():
         return False
 
 
+def is_torchvision_available():
+    if importlib.util.find_spec("torchvision") is not None:
+        return True
+    else:
+        return False
+
+
 def revert_liger_kernel_to_granite(model_config: MiniModelConfig):
     """
     Revert all Liger kernel patches applied to Granite.

liger_kernel_nightly-0.5.10.dev20250611191801/dev/modal/benchmarks.py

@@ -1,28 +0,0 @@
-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_benchmarks", 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=60 * 45)
-def liger_benchmarks():
-    import subprocess
-
-    subprocess.run(
-        ["uv pip install -e '.[dev]' --system"],
-        check=True,
-        shell=True,
-        cwd=REMOTE_ROOT_PATH,
-    )
-    subprocess.run(["python benchmark/scripts/benchmark_kto_loss.py"], check=True, shell=True, cwd=REMOTE_ROOT_PATH)
-    subprocess.run(["python benchmark/scripts/benchmark_cpo_loss.py"], check=True, shell=True, cwd=REMOTE_ROOT_PATH)

{liger_kernel_nightly-0.5.10.dev20250611191801 → liger_kernel_nightly-0.5.10.dev20250613192702}/.github/workflows/benchmark.yml

@@ -29,6 +29,7 @@ jobs:
       OUTPUT_FILENAME: benchmark.csv
       GENERATED_CSV: benchmark/data/all_benchmark_data.csv
 
+
     steps:
       - name: Checkout code
         uses: actions/checkout@v3
@@ -71,7 +72,6 @@ jobs:
         run: |
           mkdir -p gh-pages/${OUTPUT_DIR}/${{ steps.get_hash.outputs.hash }}
           cp ${GENERATED_CSV} gh-pages/${OUTPUT_DIR}/${{ steps.get_hash.outputs.hash }}/${OUTPUT_FILENAME}
-
       # Step 7: Append commit hash to commits.txt if not already present
       - name: Update commits.txt
         run: |
@@ -84,7 +84,6 @@ jobs:
           if ! grep -q "${{ steps.get_hash.outputs.hash }}" ${OUTPUT_DIR}/commits.txt; then
             echo "${{ steps.get_hash.outputs.hash }}" >> ${OUTPUT_DIR}/commits.txt
           fi
-
       # Step 7: Commit and push
       - name: Commit and push to gh-pages
         run: |
@@ -94,3 +93,4 @@ jobs:
           git add .
           git commit -m "Add benchmark for commit ${{ steps.get_hash.outputs.hash }}" || echo "No changes to commit"
           git push origin gh-pages
+