liger-kernel-nightly 0.5.8.dev20250416185644__tar.gz → 0.5.8.dev20250422210723__tar.gz

{liger_kernel_nightly-0.5.8.dev20250416185644 → liger_kernel_nightly-0.5.8.dev20250422210723}/.github/workflows/intel-ci.yml

@@ -45,27 +45,40 @@ jobs:
       run: make checkstyle
 
   tests:
-    runs-on: linux-max1550-gpu-8
+    runs-on: linux-max1550-pvc-8
     needs: [checkstyle]
-
+    if: success()
+    container:
+      image: intel/oneapi-basekit:2025.0.1-0-devel-ubuntu24.04
+      options: --privileged -v /dev/dri/by-path:/dev/dri/by-path --device=/dev/dri --ipc=host
     steps:
+    - name: Set up python
+      shell: bash
+      run: |
+        apt-get update && \
+        apt-get install -y python3.12-venv python3-pip && \
+        ln -sf /usr/bin/python3 /usr/bin/python && \
+        apt-get clean && rm -rf /var/lib/apt/lists/*
+
     - name: Checkout code
       uses: actions/checkout@v3
-
-    - name: Set up Python
-      uses: actions/setup-python@v3
-      with:
-        python-version: '3.10'
-
+      
     - name: Setup Dependencies
+      shell: bash
       run: |
-        python -m pip install --upgrade pip
+        python -m venv test-env
+        . test-env/bin/activate
         pip install -e .[dev] --extra-index-url https://download.pytorch.org/whl/test/xpu
-    
+
     - name: List Python Environments
-      run: python -m pip list
+      shell: bash
+      run: |
+        . test-env/bin/activate
+        python -m pip list
     
     - name: Run Unit Tests
+      shell: bash
       run: |
+        . test-env/bin/activate
         make test
         make test-convergence

{liger_kernel_nightly-0.5.8.dev20250416185644 → liger_kernel_nightly-0.5.8.dev20250422210723}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.8.dev20250416185644
+Version: 0.5.8.dev20250422210723
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.8.dev20250416185644 → liger_kernel_nightly-0.5.8.dev20250422210723}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.8.dev20250416185644"
+version = "0.5.8.dev20250422210723"
 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-0.5.8.dev20250416185644 → liger_kernel_nightly-0.5.8.dev20250422210723}/src/liger_kernel/chunked_loss/dpo_loss.py

@@ -68,6 +68,7 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
         compute_nll_loss=False,
         compiled=True,
         use_ref_model=True,
+        average_log_prob=False,
         chunk_size=1,
     ):
         """
@@ -85,6 +86,7 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
             compute_nll_loss (bool): Whether to compute the NLL loss
             compiled (bool): Whether to use torch compile
             use_ref_model (bool): Whether to use a reference model
+            average_log_prob (bool): Whether to average the log probability per non-masked token
             chunk_size (int): Size of chunks for processing.
         Returns:
             torch.Tensor: Computed loss
@@ -104,13 +106,14 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
             ref_input=ref_input,
             ref_weight=ref_weight,
             ref_bias=ref_bias,
+            average_log_prob=average_log_prob,
             chunk_size=chunk_size,
         )
 
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None, None, None, None, None, None
+        return *grads, None, None, None, None, None, None, None, None, None, None
 
 
 class LigerFusedLinearDPOLoss(torch.nn.Module):
@@ -125,6 +128,7 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
         compute_nll_loss: bool = False,
         compiled: bool = True,
         use_ref_model: bool = True,
+        average_log_prob: bool = True,
         chunk_size: int = 1,
     ):
         """
@@ -134,6 +138,7 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
             compute_nll_loss (bool): Whether to compute the NLL loss.
             compiled (bool): Whether to use the torch compiled kernel.
             use_ref_model (bool): Whether to use a reference model for the DPO loss.
+            average_log_prob (bool): Whether to average the log probability per non-masked token.
             chunk_size (int): Size of chunks for processing.
         """
         super().__init__()
@@ -142,6 +147,7 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
         self.compute_nll_loss = compute_nll_loss
         self.compiled = compiled
         self.use_ref_model = use_ref_model
+        self.average_log_prob = average_log_prob
         self.chunk_size = chunk_size
 
     def forward(
@@ -167,5 +173,6 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
             self.compute_nll_loss,
             self.compiled,
             self.use_ref_model,
+            self.average_log_prob,
             self.chunk_size,
         )

{liger_kernel_nightly-0.5.8.dev20250416185644 → liger_kernel_nightly-0.5.8.dev20250422210723}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.8.dev20250416185644
+Version: 0.5.8.dev20250422210723
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.8.dev20250416185644 → liger_kernel_nightly-0.5.8.dev20250422210723}/test/convergence/bf16/test_mini_models_multimodal.py

@@ -886,6 +886,7 @@ def run_mini_model_multimodal(
                     not GEMMA3_AVAILABLE,
                     reason="Gemma3 not available in this version of transformers",
                 ),
+                pytest.mark.skipif(device == "xpu", reason="skip for XPU"),
             ],
         ),
     ],

{liger_kernel_nightly-0.5.8.dev20250416185644 → liger_kernel_nightly-0.5.8.dev20250422210723}/test/convergence/fp32/test_mini_models_multimodal.py

@@ -862,10 +862,13 @@ def run_mini_model_multimodal(
             1e-5,
             5e-3,
             1e-5,
-            marks=pytest.mark.skipif(
-                not GEMMA3_AVAILABLE,
-                reason="Gemma3 not available in this version of transformers",
-            ),
+            marks=[
+                pytest.mark.skipif(
+                    not GEMMA3_AVAILABLE,
+                    reason="Gemma3 not available in this version of transformers",
+                ),
+                pytest.mark.skipif(device == "xpu", reason="skip for XPU"),
+            ],
         ),
     ],
 )