liger-kernel-nightly 0.5.5.dev20250402185606__tar.gz → 0.5.5.dev20250402212634__tar.gz

{liger_kernel_nightly-0.5.5.dev20250402185606 → liger_kernel_nightly-0.5.5.dev20250402212634}/.github/workflows/amd-ci.yml

@@ -49,7 +49,7 @@ jobs:
     needs: [checkstyle]
     strategy:
       matrix:
-        rocm_version: ['6.2', '6.3']
+        rocm_version: ['6.3']
 
     steps:
     - name: Checkout code
@@ -62,6 +62,7 @@ jobs:
 
     - name: Setup Dependencies
       run: |
+        rocm-smi
         python -m pip install --upgrade pip
         pip install -e .[dev] --extra-index-url https://download.pytorch.org/whl/nightly/rocm${{ matrix.rocm_version }}
     

{liger_kernel_nightly-0.5.5.dev20250402185606 → liger_kernel_nightly-0.5.5.dev20250402212634}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.5.dev20250402185606
+Version: 0.5.5.dev20250402212634
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.5.dev20250402185606 → liger_kernel_nightly-0.5.5.dev20250402212634}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.5.dev20250402185606"
+version = "0.5.5.dev20250402212634"
 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-0.5.5.dev20250402185606 → liger_kernel_nightly-0.5.5.dev20250402212634}/src/liger_kernel/chunked_loss/fused_linear_ppo.py

@@ -37,6 +37,7 @@ class LigerFusedLinearPPOBase(torch.autograd.Function):
         use_ref_model=False,
         chunk_size=1,
     ):
+        # TODO: check torch compile matmul
         """Chunked forward pass for PPO loss computation.
 
         Args:

{liger_kernel_nightly-0.5.5.dev20250402185606 → liger_kernel_nightly-0.5.5.dev20250402212634}/src/liger_kernel/ops/fused_linear_jsd.py

@@ -8,11 +8,12 @@ from liger_kernel.ops.utils import amp_custom_bwd
 from liger_kernel.ops.utils import amp_custom_fwd
 from liger_kernel.ops.utils import element_mul_kernel
 from liger_kernel.ops.utils import is_hip
+from liger_kernel.utils import infer_device
 
 # The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576 https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
 # However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
 # The optimal maximum block size depends on your hardware, your kernel, and your dtype
-MAX_FUSED_SIZE = 65536 // 2
+MAX_FUSED_SIZE = 4096 if infer_device() == "xpu" else 65536 // 2
 
 
 def fused_linear_jsd_forward(

{liger_kernel_nightly-0.5.5.dev20250402185606 → liger_kernel_nightly-0.5.5.dev20250402212634}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.5.dev20250402185606
+Version: 0.5.5.dev20250402212634
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.5.dev20250402185606 → liger_kernel_nightly-0.5.5.dev20250402212634}/test/convergence/fp32/test_mini_models_with_logits.py

@@ -724,7 +724,7 @@ def run_mini_model(
             1e-4,
             torch.float32,
             1e-8,
-            1e-5,
+            2e-5,
             5e-3,
             1e-5,
             5e-3,
@@ -740,7 +740,7 @@ def run_mini_model(
             1e-4,
             torch.float32,
             1e-8,
-            1e-5,
+            2e-5,
             5e-3,
             1e-5,
             5e-3,

{liger_kernel_nightly-0.5.5.dev20250402185606 → liger_kernel_nightly-0.5.5.dev20250402212634}/test/transformers/test_jsd.py

@@ -78,11 +78,11 @@ _DTYPE_PARAMS = (
     [
         pytest.param(
             torch.bfloat16,
-            1e-8,
+            1e-7,
             5e-2,
             marks=pytest.mark.skipif(not supports_bfloat16(), reason="bfloat16 not supported on this GPU"),
         ),
-        (torch.float32, 1e-8, 1e-6),
+        (torch.float32, 1e-7, 1e-6),
         (torch.float16, 1e-3, 1e-3),
     ],
 )