liger-kernel-nightly 0.5.10.dev20250528223524__tar.gz → 0.5.10.dev20250601024230__tar.gz

{liger_kernel_nightly-0.5.10.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.10.dev20250528223524
+Version: 0.5.10.dev20250601024230
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -325,6 +325,8 @@ loss.backward()
 | GeGLU                           | `liger_kernel.transformers.LigerGEGLUMLP`                   |
 | CrossEntropy                    | `liger_kernel.transformers.LigerCrossEntropyLoss`           |
 | Fused Linear CrossEntropy       | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+| Multi Token Attention           | `liger_kernel.transformers.LigerMultiTokenAttention`        |
+| Softmax                         | `liger_kernel.transformers.LigerSoftmax`                    |
 | Sparsemax                       | `liger_kernel.transformers.LigerSparsemax`                  |
 
 

{liger_kernel_nightly-0.5.10.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/README.md

@@ -277,6 +277,8 @@ loss.backward()
 | GeGLU                           | `liger_kernel.transformers.LigerGEGLUMLP`                   |
 | CrossEntropy                    | `liger_kernel.transformers.LigerCrossEntropyLoss`           |
 | Fused Linear CrossEntropy       | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+| Multi Token Attention           | `liger_kernel.transformers.LigerMultiTokenAttention`        |
+| Softmax                         | `liger_kernel.transformers.LigerSoftmax`                    |
 | Sparsemax                       | `liger_kernel.transformers.LigerSparsemax`                  |
 
 

{liger_kernel_nightly-0.5.10.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/benchmark/scripts/benchmark_softmax.py

@@ -8,7 +8,7 @@ from utils import _test_memory
 from utils import parse_benchmark_script_args
 from utils import run_benchmarks
 
-from liger_kernel.transformers.softmax import LigerKernelSoftmax
+from liger_kernel.transformers.softmax import LigerSoftmax
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -23,7 +23,7 @@ def bench_speed_softmax(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOut
     dtype = extra_benchmark_config["dtype"]
 
     x_shape = (M, N)
-    liger_softmax = LigerKernelSoftmax().to(device).to(dtype)
+    liger_softmax = LigerSoftmax().to(device).to(dtype)
     torch_softmax = torch.nn.Softmax(dim=-1).to(device).to(dtype)
 
     x = torch.randn(x_shape, dtype=dtype, device=device)
@@ -72,7 +72,7 @@ def bench_memory_softmax(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOu
     dtype = extra_benchmark_config.get("dtype", torch.float32)
 
     torch_softmax = torch.nn.Softmax(dim=-1)
-    liger_softmax = LigerKernelSoftmax().to(device).to(dtype)
+    liger_softmax = LigerSoftmax().to(device).to(dtype)
 
     x = torch.randn(shape, device=device, dtype=dtype, requires_grad=True)
 

{liger_kernel_nightly-0.5.10.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/docs/Low-Level-APIs.md

@@ -9,6 +9,8 @@
 | GeGLU                           | `liger_kernel.transformers.LigerGEGLUMLP`                   |
 | CrossEntropy                    | `liger_kernel.transformers.LigerCrossEntropyLoss`           |
 | Fused Linear CrossEntropy       | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+| Multi Token Attention           | `liger_kernel.transformers.LigerMultiTokenAttention`        |
+| Softmax                         | `liger_kernel.transformers.LigerSoftmax`                    |
 | Sparsemax                       | `liger_kernel.transformers.LigerSparsemax`                  |
 
 
@@ -51,6 +53,19 @@ This kernel combines linear transformations with cross-entropy loss calculations
 !!! Example "Try it out"
     You can experiment as shown in this example [here](https://colab.research.google.com/drive/1Z2QtvaIiLm5MWOs7X6ZPS1MN3hcIJFbj?usp=sharing)
 
+### Multi Token Attention
+
+The Multi Token Attention kernel implementation provides and optimized fused implementation of multi-token attention over the implemented Pytorch model baseline. This is a new attention mechanism that can operate on multiple Q and K inputs introduced by Meta Research.
+
+Paper: https://arxiv.org/abs/2504.00927
+
+### Softmax
+
+The Softmax kernel implementation provides an optimized implementation of the softmax operation, which is a fundamental component in neural networks for converting raw scores into probability distributions.
+
+The implementation shows notable speedups compared to the Softmax PyTorch implementation
+
+
 ### Sparsemax
 
 Sparsemax is a sparse alternative to softmax that produces sparse probability distributions. This kernel implements an efficient version of the sparsemax operation that can be used as a drop-in replacement for softmax in attention mechanisms or classification tasks.

{liger_kernel_nightly-0.5.10.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.10.dev20250528223524"
+version = "0.5.10.dev20250601024230"
 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.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/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.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/src/liger_kernel/transformers/monkey_patch.py

@@ -776,7 +776,8 @@ 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.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/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.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/src/liger_kernel/transformers/softmax.py

@@ -4,7 +4,7 @@ import torch.nn as nn
 from liger_kernel.ops.softmax import LigerSoftmaxFunction
 
 
-class LigerKernelSoftmax(nn.Module):
+class LigerSoftmax(nn.Module):
     def __init__(self):
         super().__init__()
 

{liger_kernel_nightly-0.5.10.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.10.dev20250528223524
+Version: 0.5.10.dev20250601024230
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -325,6 +325,8 @@ loss.backward()
 | GeGLU                           | `liger_kernel.transformers.LigerGEGLUMLP`                   |
 | CrossEntropy                    | `liger_kernel.transformers.LigerCrossEntropyLoss`           |
 | Fused Linear CrossEntropy       | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+| Multi Token Attention           | `liger_kernel.transformers.LigerMultiTokenAttention`        |
+| Softmax                         | `liger_kernel.transformers.LigerSoftmax`                    |
 | Sparsemax                       | `liger_kernel.transformers.LigerSparsemax`                  |
 
 

{liger_kernel_nightly-0.5.10.dev20250528223524 → liger_kernel_nightly-0.5.10.dev20250601024230}/test/transformers/test_softmax.py

@@ -6,7 +6,7 @@ from test.utils import set_seed
 from test.utils import supports_bfloat16
 
 from liger_kernel.transformers.functional import liger_softmax
-from liger_kernel.transformers.softmax import LigerKernelSoftmax
+from liger_kernel.transformers.softmax import LigerSoftmax
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -47,7 +47,7 @@ def test_liger_softmax(shape, dtype, atol, rtol):
 
     torch_softmax = torch.nn.Softmax(dim=-1)
     ref_out = torch_softmax(x1)
-    liger_softmax = LigerKernelSoftmax().to(device).to(dtype)
+    liger_softmax = LigerSoftmax().to(device).to(dtype)
     liger_out = liger_softmax(x2)
 
     assert_verbose_allclose(ref_out, liger_out, atol=atol, rtol=rtol)