liger-kernel-nightly 0.5.10.dev20250531184114__tar.gz → 0.5.10.dev20250602014906__tar.gz

{liger_kernel_nightly-0.5.10.dev20250531184114 → liger_kernel_nightly-0.5.10.dev20250602014906}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.10.dev20250531184114
+Version: 0.5.10.dev20250602014906
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.10.dev20250531184114 → liger_kernel_nightly-0.5.10.dev20250602014906}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.10.dev20250531184114"
+version = "0.5.10.dev20250602014906"
 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.dev20250531184114 → liger_kernel_nightly-0.5.10.dev20250602014906}/src/liger_kernel/ops/rms_norm.py

@@ -193,6 +193,7 @@ def _rms_norm_backward_kernel(
 
     tl.store(dW_ptr + row_block_id * dW_row_stride + col_offsets, dW_row, mask=mask)
 
+
 @triton.jit
 def _block_rms_norm_forward_kernel(
     Y_ptr,
@@ -225,8 +226,11 @@ def _block_rms_norm_forward_kernel(
     row_mask = row_idx < n_rows
     col_mask = col_offsets < n_cols
 
-
-    X_row = tl.load(X_ptr + row_idx[:, None] * X_row_stride + col_offsets[None, :], mask=row_mask[:, None] & col_mask[None, :] , other=0)
+    X_row = tl.load(
+        X_ptr + row_idx[:, None] * X_row_stride + col_offsets[None, :],
+        mask=row_mask[:, None] & col_mask[None, :],
+        other=0,
+    )
     X_row_dtype = X_row.dtype
     W_row = tl.load(W_ptr + col_offsets, mask=col_mask, other=0)
 
@@ -262,7 +266,12 @@ def _block_rms_norm_forward_kernel(
     if casting_mode == _CASTING_MODE_GEMMA:
         Y_row = Y_row.to(X_row_dtype)
 
-    tl.store(Y_ptr + row_idx[:, None] * Y_row_stride + col_offsets[None, :], Y_row, mask=row_mask[:, None] & col_mask[None, :])
+    tl.store(
+        Y_ptr + row_idx[:, None] * Y_row_stride + col_offsets[None, :],
+        Y_row,
+        mask=row_mask[:, None] & col_mask[None, :],
+    )
+
 
 @triton.jit
 def _block_rms_norm_backward_kernel(
@@ -306,8 +315,16 @@ def _block_rms_norm_backward_kernel(
     for start in range(pid * BLOCK_ROW, n_rows, NUM_SMS * BLOCK_ROW):
         row_idx = start + tl.arange(0, BLOCK_ROW)
         row_mask = row_idx < n_rows
-        dY_row = tl.load(dY_ptr + row_idx[:, None] * dY_row_stride + col_offsets[None, :], mask=row_mask[:, None] & col_mask[None, :], other=0.0)
-        X_row = tl.load(X_ptr + row_idx[:, None] * X_row_stride + col_offsets[None, :], mask=row_mask[:, None] & col_mask[None, :], other=0.0)
+        dY_row = tl.load(
+            dY_ptr + row_idx[:, None] * dY_row_stride + col_offsets[None, :],
+            mask=row_mask[:, None] & col_mask[None, :],
+            other=0.0,
+        )
+        X_row = tl.load(
+            X_ptr + row_idx[:, None] * X_row_stride + col_offsets[None, :],
+            mask=row_mask[:, None] & col_mask[None, :],
+            other=0.0,
+        )
 
         # Get cached rms
         rstd_row = tl.load(RSTD_ptr + row_idx * RSTD_row_stride, row_mask)
@@ -326,7 +343,9 @@ def _block_rms_norm_backward_kernel(
 
         dX_row = rstd_row[:, None] * m
 
-        dX_row += (rstd_row[:, None]) * (-(1 / n_cols) * (rstd_row * rstd_row * tl.sum(m * X_row, axis=1))[:, None] * X_row)
+        dX_row += (rstd_row[:, None]) * (
+            -(1 / n_cols) * (rstd_row * rstd_row * tl.sum(m * X_row, axis=1))[:, None] * X_row
+        )
 
         # calculate the gradient of W
         if casting_mode == _CASTING_MODE_LLAMA:
@@ -335,8 +354,11 @@ def _block_rms_norm_backward_kernel(
             # here X_row is already in fp32 (see previous if block)
             dW_row += tl.sum(dY_row * (X_row * rstd_row[:, None]), 0)
 
-        tl.store(dX_ptr + row_idx[:, None] * dX_row_stride + col_offsets[None, :], dX_row, mask=row_mask[:, None] & col_mask[None, :])
-
+        tl.store(
+            dX_ptr + row_idx[:, None] * dX_row_stride + col_offsets[None, :],
+            dX_row,
+            mask=row_mask[:, None] & col_mask[None, :],
+        )
 
     tl.store(dW_ptr + pid * dW_row_stride + col_offsets, dW_row, mask=col_mask)
 
@@ -549,15 +571,6 @@ class LigerRMSNormFunction(torch.autograd.Function):
         """
         X, W, RSTD = ctx.saved_tensors
         dX, dW = rms_norm_backward(
-            dY,
-            X,
-            W,
-            RSTD,
-            ctx.offset,
-            ctx.casting_mode,
-            ctx.BLOCK_SIZE,
-            ctx.num_warps,
-            ctx.in_place,
-            ctx.row_mode
+            dY, X, W, RSTD, ctx.offset, ctx.casting_mode, ctx.BLOCK_SIZE, ctx.num_warps, ctx.in_place, ctx.row_mode
         )
         return dX, dW, None, None, None, None, None

{liger_kernel_nightly-0.5.10.dev20250531184114 → liger_kernel_nightly-0.5.10.dev20250602014906}/src/liger_kernel/transformers/monkey_patch.py

@@ -428,13 +428,14 @@ def apply_liger_kernel_to_mllama(
         if isinstance(model, MllamaForConditionalGeneration):
             language_model: MllamaForCausalLM = model.language_model
             vision_model: MllamaVisionModel = model.vision_model
-            text_model: MllamaTextModel = language_model.model
+            text_model: MllamaTextModel = language_model
         elif isinstance(model, MllamaForCausalLM):
             text_model = model.model
             vision_model = None
         elif isinstance(model, MllamaTextModel):
             text_model = model
             vision_model = None
+
         else:
             raise ValueError(f"Unsupported Mllama model type: {type(model)}")
 
@@ -776,7 +777,8 @@ def apply_liger_kernel_to_gemma3_text(
 
     from transformers.models.gemma3 import modeling_gemma3
     from transformers.models.gemma3.modeling_gemma3 import Gemma3DecoderLayer
-    from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM, Gemma3TextModel
+    from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM
+    from transformers.models.gemma3.modeling_gemma3 import Gemma3TextModel
 
     from liger_kernel.transformers.gema3_rms import LigerRMSNormForGemma3
     from liger_kernel.transformers.model.gemma3 import causal_forward

{liger_kernel_nightly-0.5.10.dev20250531184114 → liger_kernel_nightly-0.5.10.dev20250602014906}/src/liger_kernel/transformers/rms_norm.py

@@ -37,7 +37,7 @@ class LigerRMSNorm(nn.Module):
             self.offset,
             self.casting_mode,
             self.in_place,
-            self.row_mode
+            self.row_mode,
         )
 
     def extra_repr(self):

{liger_kernel_nightly-0.5.10.dev20250531184114 → liger_kernel_nightly-0.5.10.dev20250602014906}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.10.dev20250531184114
+Version: 0.5.10.dev20250602014906
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.10.dev20250531184114 → liger_kernel_nightly-0.5.10.dev20250602014906}/test/transformers/test_monkey_patch.py

@@ -348,10 +348,10 @@ def test_apply_liger_kernel_to_instance_for_mllama_for_conditional_generation():
         assert isinstance(dummy_model_instance, MllamaForConditionalGeneration)
 
         # Check that model instance variables are not yet patched with Liger modules
-        assert inspect.getsource(dummy_model_instance.language_model.model.norm.forward) != inspect.getsource(
+        assert inspect.getsource(dummy_model_instance.language_model.norm.forward) != inspect.getsource(
             LigerRMSNorm.forward
         )
-        for layer in dummy_model_instance.language_model.model.layers:
+        for layer in dummy_model_instance.language_model.layers:
             assert inspect.getsource(layer.mlp.forward) != inspect.getsource(LigerSwiGLUMLP.forward)
             assert inspect.getsource(layer.input_layernorm.forward) != inspect.getsource(LigerRMSNorm.forward)
             assert inspect.getsource(layer.post_attention_layernorm.forward) != inspect.getsource(LigerRMSNorm.forward)
@@ -377,10 +377,10 @@ def test_apply_liger_kernel_to_instance_for_mllama_for_conditional_generation():
         _apply_liger_kernel_to_instance(model=dummy_model_instance)
 
         # Check that the model's instance variables were correctly patched with Liger modules
-        assert inspect.getsource(dummy_model_instance.language_model.model.norm.forward) == inspect.getsource(
+        assert inspect.getsource(dummy_model_instance.language_model.norm.forward) == inspect.getsource(
             LigerRMSNorm.forward
         )
-        for layer in dummy_model_instance.language_model.model.layers:
+        for layer in dummy_model_instance.language_model.layers:
             assert inspect.getsource(layer.mlp.forward) == inspect.getsource(LigerSwiGLUMLP.forward)
             assert inspect.getsource(layer.input_layernorm.forward) == inspect.getsource(LigerRMSNorm.forward)
             assert inspect.getsource(layer.post_attention_layernorm.forward) == inspect.getsource(LigerRMSNorm.forward)