liger-kernel 0.5.10__py3-none-any.whl → 0.6.1__py3-none-any.whl

liger_kernel/transformers/multi_token_attention.py

@@ -0,0 +1,64 @@
+import math
+
+import torch
+import torch.nn as nn
+
+from torch.nn.modules.utils import _pair
+
+from liger_kernel.ops.multi_token_attention import LigerMultiTokenAttentionFunction
+
+
+class LigerMultiTokenAttention(nn.Module):
+    """
+    Multi-Token Attention:
+        out = mask_{0}(conv2d(softmax(mask_{-\inf}(scores))))
+
+    Reference: https://arxiv.org/pdf/2504.00927
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int,
+        stride: int = 1,
+        padding: int = 0,
+        dilation: int = 1,
+        groups: int = 1,
+        bias: bool = True,
+        sparse: bool = False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = _pair(kernel_size)
+        self.stride = _pair(stride)
+        self.padding = _pair(padding)
+        self.dilation = _pair(dilation)
+        self.groups = groups
+        self.sparse = sparse
+
+        self.weight = nn.Parameter(torch.empty(out_channels, in_channels // groups, *self.kernel_size))
+        if bias:
+            self.bias = nn.Parameter(torch.empty(out_channels))
+        else:
+            self.register_parameter("bias", None)
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+        if self.bias is not None:
+            nn.init.zeros_(self.bias)
+
+    def forward(self, scores: torch.Tensor) -> torch.Tensor:
+        return LigerMultiTokenAttentionFunction.apply(
+            scores,
+            self.weight,
+            self.bias,
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+            self.sparse,
+        )

liger_kernel/transformers/rms_norm.py

@@ -13,6 +13,7 @@ class LigerRMSNorm(nn.Module):
         casting_mode="llama",
         init_fn="ones",
         in_place=True,
+        row_mode=None,
     ):
         super().__init__()
         assert init_fn in [
@@ -20,11 +21,12 @@ class LigerRMSNorm(nn.Module):
             "zeros",
         ], f"init_fn must be either 'ones' or 'zeros', got {init_fn}"
         self.weight = nn.Parameter(torch.ones(hidden_size) if init_fn == "ones" else torch.zeros(hidden_size))
-        self.variance_epsilon, self.offset, self.casting_mode, self.in_place = (
+        self.variance_epsilon, self.offset, self.casting_mode, self.in_place, self.row_mode = (
             eps,
             offset,
             casting_mode,
             in_place,
+            row_mode,
         )
 
     def forward(self, hidden_states):
@@ -35,9 +37,43 @@ class LigerRMSNorm(nn.Module):
             self.offset,
             self.casting_mode,
             self.in_place,
+            self.row_mode,
         )
 
     def extra_repr(self):
-        return (
-            f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}, offset={self.offset}, in_place={self.in_place}"
-        )
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}, offset={self.offset}, in_place={self.in_place}, row_mode={self.row_mode}"
+
+
+class LigerRMSNormForGemma(LigerRMSNorm):
+    def __init__(
+        self, hidden_size, eps=1e-6, offset=1.0, casting_mode="gemma", init_fn="zeros", in_place=True, row_mode=None
+    ):
+        super().__init__(hidden_size, eps, offset, casting_mode, init_fn, in_place, row_mode)
+
+
+class LigerRMSNormForGemma2(LigerRMSNorm):
+    def __init__(
+        self, hidden_size, eps=1e-6, offset=1.0, casting_mode="gemma", init_fn="zeros", in_place=False, row_mode=None
+    ):
+        super().__init__(hidden_size, eps, offset, casting_mode, init_fn, in_place, row_mode)
+
+
+class LigerRMSNormForGemma3(LigerRMSNorm):
+    """Gemma3RMSNorm has a dim argument not hidden_size used in q_norm and k_norm."""
+
+    def __init__(self, dim, eps=0.000001, offset=1.0, casting_mode="gemma", init_fn="zeros", in_place=False):
+        super().__init__(dim, eps, offset, casting_mode, init_fn, in_place)
+
+
+class LigerRMSNormForOlmo2(LigerRMSNorm):
+    def __init__(
+        self, hidden_size, eps=1e-6, offset=0.0, casting_mode="llama", init_fn="ones", in_place=False, row_mode=None
+    ):
+        super().__init__(hidden_size, eps, offset, casting_mode, init_fn, in_place, row_mode)
+
+
+class LigerRMSNormForGlm4(LigerRMSNorm):
+    def __init__(
+        self, hidden_size, eps=1e-6, offset=0.0, casting_mode="llama", init_fn="ones", in_place=False, row_mode=None
+    ):
+        super().__init__(hidden_size, eps, offset, casting_mode, init_fn, in_place, row_mode)

liger_kernel/transformers/softmax.py

@@ -0,0 +1,12 @@
+import torch
+import torch.nn as nn
+
+from liger_kernel.ops.softmax import LigerSoftmaxFunction
+
+
+class LigerSoftmax(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: torch.Tensor):
+        return LigerSoftmaxFunction.apply(x)

{liger_kernel-0.5.10.dist-info → liger_kernel-0.6.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: liger_kernel
-Version: 0.5.10
+Version: 0.6.1
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -33,7 +33,7 @@ License-File: NOTICE
 Requires-Dist: torch>=2.1.2
 Requires-Dist: triton>=2.3.1
 Provides-Extra: dev
-Requires-Dist: transformers>=4.44.2; extra == "dev"
+Requires-Dist: transformers>=4.49.0; extra == "dev"
 Requires-Dist: matplotlib>=3.7.2; extra == "dev"
 Requires-Dist: flake8>=4.0.1.1; extra == "dev"
 Requires-Dist: black>=24.4.2; extra == "dev"
@@ -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"
 Dynamic: license-file
 Dynamic: provides-extra
 Dynamic: requires-dist
@@ -83,7 +84,7 @@ Dynamic: requires-dist
         </td>
         <td style="padding: 10px;">
             <a href="https://discord.gg/gpumode">
-                <img src="https://dcbadge.vercel.app/api/server/gpumode?style=flat" alt="Join Our Discord">
+                <img src="https://dcbadge.limes.pink/api/server/gpumode?style=flat" alt="Join Our Discord">
             </a>
         </td>
     </tr>
@@ -114,6 +115,8 @@ Dynamic: requires-dist
 
 We've also added optimized Post-Training kernels that deliver **up to 80% memory savings** for alignment and distillation tasks. We support losses like DPO, CPO, ORPO, SimPO, KTO, JSD, and many more. Check out [how we optimize the memory](https://x.com/hsu_byron/status/1866577403918917655).
 
+You can view the documentation site for additional installation, usage examples, and API references:https://linkedin.github.io/Liger-Kernel/
+
 ## Supercharge Your Model with Liger Kernel
 
 ![Banner](https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/banner.GIF)
@@ -290,6 +293,7 @@ loss.backward()
 
 | **Model**   | **API**                                                      | **Supported Operations**                                                |
 |-------------|--------------------------------------------------------------|-------------------------------------------------------------------------|
+| Llama4 (Text) & (Multimodal)      | `liger_kernel.transformers.apply_liger_kernel_to_llama4`   | RMSNorm, LayerNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | LLaMA 2 & 3 | `liger_kernel.transformers.apply_liger_kernel_to_llama`   | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | LLaMA 3.2-Vision | `liger_kernel.transformers.apply_liger_kernel_to_mllama`   | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Mistral     | `liger_kernel.transformers.apply_liger_kernel_to_mistral`  | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
@@ -303,7 +307,7 @@ loss.backward()
 | Qwen2-VL, & QVQ       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen2.5-VL       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_5_vl`    | RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen3   | `liger_kernel.transformers.apply_liger_kernel_to_qwen3`    |  RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
-| Qwen3 MoE | `liger_kernel_transformers.apply_liger_kernel_to_qwen3_moe` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
+| Qwen3 MoE | `liger_kernel.transformers.apply_liger_kernel_to_qwen3_moe` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
 | OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
@@ -326,6 +330,8 @@ 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`                  |
 
 
@@ -408,21 +414,19 @@ loss.backward()
 
 - For issues, create a Github ticket in this repository
 - For open discussion, join [our discord channel on GPUMode](https://discord.com/channels/1189498204333543425/1275130785933951039)
-- For formal collaboration, send an email to yannchen@linkedin.com and hning@linkedin.com
+- For formal collaboration, send an email to Yanning Chen(yannchen@linkedin.com) and Zhipeng Wang(zhipwang@linkedin.com)
 
 ## Cite this work
 
 Biblatex entry:
 ```bib
-@article{hsu2024ligerkernelefficienttriton,
-      title={Liger Kernel: Efficient Triton Kernels for LLM Training},
-      author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},
-      year={2024},
-      eprint={2410.10989},
-      archivePrefix={arXiv},
-      primaryClass={cs.LG},
-      url={https://arxiv.org/abs/2410.10989},
-      journal={arXiv preprint arXiv:2410.10989},
+@inproceedings{
+hsu2025ligerkernel,
+title={Liger-Kernel: Efficient Triton Kernels for {LLM} Training},
+author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen and Zhipeng Wang},
+booktitle={Championing Open-source DEvelopment in ML Workshop @ ICML25},
+year={2025},
+url={https://openreview.net/forum?id=36SjAIT42G}
 }
 ```
 

{liger_kernel-0.5.10.dist-info → liger_kernel-0.6.1.dist-info}/RECORD

@@ -2,10 +2,11 @@ liger_kernel/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 liger_kernel/env_report.py,sha256=uhdEC8OydxoZlb7B6YYcAaBF3crGFdIck-4cxaW4NJY,1728
 liger_kernel/utils.py,sha256=BQleeZWHSZPNuPcYcoZTOp1kcNEZONZilPP5-AmjgWI,2024
 liger_kernel/chunked_loss/README.md,sha256=0FmkFC3hKBqyoDT5uTlIYmrvRkF-EOCR1y-EBU1LpWU,2248
-liger_kernel/chunked_loss/__init__.py,sha256=ATu-xX5Fc49Cr6yBOGBRNTo593ZrU5ZCsIuvoIbJWw4,603
+liger_kernel/chunked_loss/__init__.py,sha256=J5_jNnzZ4gZmA38W5f_4oab7xMoNk1Xy-yh3X_Xlf-s,714
+liger_kernel/chunked_loss/cosine_similarity_loss.py,sha256=pZ07OQ6RI-c8uk96tDRlUXdt31-da7yWhfwircZlKRw,4198
 liger_kernel/chunked_loss/cpo_loss.py,sha256=Gzz1eU4kgcbdubFVRy55e8A1Cr-r45UgNicXwZIjmBU,5454
 liger_kernel/chunked_loss/dpo_loss.py,sha256=tapMiNdI8_ufW55iG0Ud4dmiW39gu1DzlvtoOCHrdGg,6259
-liger_kernel/chunked_loss/functional.py,sha256=9G3nKm-Bi7uoZRFkL8wwGMl6juDl4bSzDvTa5GHZPzg,955
+liger_kernel/chunked_loss/functional.py,sha256=-XPDbLml9dHmvoSU2VNTUrBDFehuzvuAGPikVetBMtI,1132
 liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=ooR-qnZCyWJN935oHCSWLaKKKyaYERyhNczRGi1VOiw,11935
 liger_kernel/chunked_loss/fused_linear_ppo.py,sha256=AA19cpv6D8mo5RbSK5GRCcZoOSnpxV_Z1eJlAsC5eic,13434
 liger_kernel/chunked_loss/fused_linear_preference.py,sha256=FIH85uUXAOgYx5Ax8MjFhJHVu-2pKtY7wSegd0zSyyY,18336
@@ -17,74 +18,83 @@ liger_kernel/chunked_loss/orpo_loss.py,sha256=nu9UYG16dcMw93lvHi4_hYs3Q0FK1KnlmM
 liger_kernel/chunked_loss/simpo_loss.py,sha256=fy2w8KbhMrBv7b1jdIeH3bBFxY52bPQPZb3KwBvmurM,5385
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 liger_kernel/ops/cross_entropy.py,sha256=e8THGnhOcy_0SbOLABx67HEM7-B8a8pG7nDKbCRpQKM,19123
-liger_kernel/ops/dyt.py,sha256=Y180EIvtUc2z83mhyub0EVOCQHJmWX3JnscqkOJqswk,5467
+liger_kernel/ops/dyt.py,sha256=gCLz4S8aul8SY9nvIGaoK67aGb7U9MJRQdo3ONqmQYs,5417
+liger_kernel/ops/fused_add_rms_norm.py,sha256=UBqmlqFCmhSAIpkNKd8rrfXatX7Z4J9bp2dX9A0lrJQ,14017
 liger_kernel/ops/fused_linear_cross_entropy.py,sha256=5fbGhN85n3zf0uIdJ7PYHWIRzTf0VTFiS0ARtOmqIP0,11020
 liger_kernel/ops/fused_linear_jsd.py,sha256=CSoprxb-YcJy-YUKiTcYkxN8sb9h2kdk_iHuncvSV5c,9683
-liger_kernel/ops/geglu.py,sha256=axGvCIvlBzuluoAIrWTsp2iZM4BFKNInkPov8YVvH9E,4126
+liger_kernel/ops/fused_neighborhood_attention.py,sha256=vPi5xbnh6wxyZehaqo6Tuilqo2fN5SGDiONjnNmIKqs,35556
+liger_kernel/ops/geglu.py,sha256=r0WSq9E93zzynL44Wh8femzOWK07_SseBM_pJUyxT3s,4144
 liger_kernel/ops/group_norm.py,sha256=qD4D4lSjSgVtO52EBNLC2iTseALRgPgqXE50U2woggk,10837
 liger_kernel/ops/grpo_loss.py,sha256=anRnv7k1-AV3pCC6_TqP0GMg78YYUfRAJrbpx6PVhl0,9448
 liger_kernel/ops/jsd.py,sha256=onHp5T3MbvJaVz5Vup7Ww6EQp_HTaZeayTjJk6FgQMY,7042
 liger_kernel/ops/kl_div.py,sha256=ZjGdDLKWksHT9dZ0xF_TDgAkj5cuMTwwT5tr9E-_24o,8734
-liger_kernel/ops/layer_norm.py,sha256=vWCyOm-F2GMAilB-ozJcFeUQQLCJoTE_uiXq-_0uYuI,8356
+liger_kernel/ops/layer_norm.py,sha256=BHPDuaogMTfIJkBJdqLZbOQouNWTf3fJVyOQOD7blCE,9901
+liger_kernel/ops/multi_token_attention.py,sha256=Oz_RXDp-OSS_R_HuGmaETHdAJ7Toda_70OfE7TXMUlY,7645
 liger_kernel/ops/qwen2vl_mrope.py,sha256=3GExhYpLgB4VUtyZyjRk8XjEur3W4EWF6HQ67ML5vBU,8481
-liger_kernel/ops/rms_norm.py,sha256=PP27OIBmV9By63i13jot9ylDowW0nuxY_JFIkaPLgL4,12078
-liger_kernel/ops/rope.py,sha256=ofmBOkUpZZO-Q8Z5B_LOFYYLD-YT-8WnJ4vGOrDYouI,8943
-liger_kernel/ops/sparsemax.py,sha256=t7JWIyzq1piikXUufayFzsfkzVaCYU-hXPuMs7839pk,4850
-liger_kernel/ops/swiglu.py,sha256=KmgMjaJQnbLLgZn2nEpbwHU_xpnYRweCyrLQSVvM1vA,3015
+liger_kernel/ops/rms_norm.py,sha256=DtvsWN5YktFAoc0JYSAwVeoZfryBFJlX-ipU7ooP01A,18891
+liger_kernel/ops/rope.py,sha256=v-7JHRrv-5ImoROkpKfl30WwWI4qTa2tAl7zQeB4ml4,8956
+liger_kernel/ops/softmax.py,sha256=tgORx6MK1IDDtZKqGarj0IPIVjqAIEUXXYPiinhRdtI,5864
+liger_kernel/ops/sparsemax.py,sha256=AeWe1xgkHJFEKWTj2vu_0hj7LztGvjqXAps-QTpCY0U,5087
+liger_kernel/ops/swiglu.py,sha256=D7nd4u_LInwsIRNCDdY77lqnTz8-W5dJrpEAt8zEO_A,3033
 liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
 liger_kernel/ops/utils.py,sha256=uoFKQqo-34N2TWQNvXMFywqGiOMMXNEVBxVojzlUAa0,3836
 liger_kernel/ops/experimental/embedding.py,sha256=tolj3tItkzpSb30zWqDN2_yX4ectflaQ8HMyKyFIQc8,4172
 liger_kernel/ops/experimental/mm_int8int2.py,sha256=TrS9lpwekrik_w5qE7AhMJD1bcq-OidjtbsW80oZ6IM,13314
-liger_kernel/transformers/__init__.py,sha256=0KX0rxyy0E_uNWVE0PSTzEVzKqc5KdFHtvdHhJm23Kk,7077
+liger_kernel/transformers/__init__.py,sha256=VoHQp5emsAJAouql37RuvtGFeZCoMIHgoIxfsyYMTc8,7564
 liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
 liger_kernel/transformers/cross_entropy.py,sha256=z3KTWQnFxr_IZaVjtYt0ZNEWQdDdYThN35xWkHlDGH0,1683
 liger_kernel/transformers/dyt.py,sha256=i-4GPaMrl-jab9TVI5qN0-H9qycn_mCbV82ozU4nbmU,723
 liger_kernel/transformers/fsdp.py,sha256=CUiyjTmjkjY7pLXQv8ly9rnzgXw6529csd9pvtJNMYc,3096
-liger_kernel/transformers/functional.py,sha256=2YBfvtdU1GRZuRpJhHgJXeGYa1RvmO6-qQvrKQrLJK4,5259
+liger_kernel/transformers/functional.py,sha256=PXnACWD7kzgge50RdOUuvtmOTS7DVkkrL7mm0cX5bOc,7734
+liger_kernel/transformers/fused_add_rms_norm.py,sha256=7_Bzg-x6lLe6W1qG2DtjDALhEpNZlC6N5GppEs9cTYY,1199
 liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=O8Sg5BT81nTaY9fSGoOY9dOD9ekibwwiuXhdUHaxntQ,1742
 liger_kernel/transformers/fused_linear_jsd.py,sha256=bZ4otCvWBuOnA5XdQL-FzZVItJlDt-ht9e_pG7PG93E,3999
+liger_kernel/transformers/fused_neighborhood_attention.py,sha256=TxYDUAt9B6WSP14aJP66C_2Mbds2sSIPGnamhUSTrC8,7957
 liger_kernel/transformers/geglu.py,sha256=mrgqzIUVd6lN7fkDKLkw5YaESDxDtFgbot430WwPVOQ,1107
-liger_kernel/transformers/gema3_rms.py,sha256=LTmZOXe6WEnv6ZroW-kU1TE2B36-z5v8OLmKr3XEVFo,353
 liger_kernel/transformers/group_norm.py,sha256=6qMAWOprr4SzP0YhNVNGQIBpM5aUHplUD2VuGJrMBz0,2173
 liger_kernel/transformers/grpo_loss.py,sha256=uAkUNKSnUGEOqa82L9w2e6AI1kcmG8K45-QxyaT8zhM,3897
 liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
 liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
 liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
-liger_kernel/transformers/monkey_patch.py,sha256=DKv5-4KyXLiVhAJ9WVFv1I1i1DzjaudTrhqx6EVYViU,74505
+liger_kernel/transformers/monkey_patch.py,sha256=tXKo4EKVp3szpdqPh051oLZFrlg_hCbWRv0RpSX_kfY,89238
+liger_kernel/transformers/multi_token_attention.py,sha256=l9VDICK0dfmifUDW668hGscP8AHq2rYcM2oGUa3baRQ,1751
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
-liger_kernel/transformers/rms_norm.py,sha256=GqCEJuGt0YdqqlMcToE0Wp4A8YFquDa4UUSyH2uFW2A,1191
+liger_kernel/transformers/rms_norm.py,sha256=vkekcvTeWY8vL4H6hg3t0XeY0Ew_3OFMPHuzqlxPPVw,2719
 liger_kernel/transformers/rope.py,sha256=ZTrTORSAyfcFIKjk6XEeYmk4ROH7xXED9L4g2NFntlE,999
+liger_kernel/transformers/softmax.py,sha256=yadlAgE4V2JByMwrDDa2s5SUBp8Jgd57xwnVvAWoBaI,264
 liger_kernel/transformers/sparsemax.py,sha256=0lQA0UEOs4mu8CMruZ3VLhImxQVXJWhPsAKUsYA7vj8,403
 liger_kernel/transformers/swiglu.py,sha256=LZ8YeLIdv2k46JleZMjzubGk98smt6t780kSgcVLsQk,3454
 liger_kernel/transformers/trainer_integration.py,sha256=W3ON51O5GkyzNJsItz0y5rKx-uy2f2cFfveZpqbUdhw,123
 liger_kernel/transformers/tvd.py,sha256=XrRfyJIqN6HFxXk8MYyFVZM1OLz3mtSbRZvWfZ_JerQ,450
 liger_kernel/transformers/experimental/embedding.py,sha256=2P0QYdlFyFrG5OqTzTa1wcRgDSyjBMv5i1a7BrDPDQw,881
 liger_kernel/transformers/model/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/transformers/model/gemma.py,sha256=gi5fVeFPryoYy0_T3rzU2wm7v_xiJnLCnTkQYR86_nk,9504
-liger_kernel/transformers/model/gemma2.py,sha256=61uH9JSZM6cPDoGHr2kNUVq2O4A3XIy2Qea36XhkkPQ,10761
-liger_kernel/transformers/model/gemma3.py,sha256=e-o7rcOJAJMZDJBB-blkLz5ildWjuDneSkakqwrADBc,15630
-liger_kernel/transformers/model/glm4.py,sha256=yYbQEcSrSTMleNTpwJosMhBf4VC9-79EyC__utmOSFg,5031
-liger_kernel/transformers/model/llama.py,sha256=ALVgzpD_YRYE7-6npb0KkjSBwrhCsgk_4lbaymOyRVw,12226
-liger_kernel/transformers/model/llava.py,sha256=RjLVnpHtOClc1jJkkPSqke7fcgWC3Jjh1rrGyvh5kb8,17008
+liger_kernel/transformers/model/gemma.py,sha256=mNX-mIwV6jI4zfbrUHp0C468pOmjzsL7mjXipGt-eS0,10007
+liger_kernel/transformers/model/gemma2.py,sha256=R_JFPyWTk7RyA7D05ZiIaNO5pX8gWcvfWf-6rdCRMxs,11296
+liger_kernel/transformers/model/gemma3.py,sha256=FKO4j3t4W_5uECRA1lhVnXC-It2GhirHm4tpCf9ApAc,12785
+liger_kernel/transformers/model/glm4.py,sha256=GlnEhdGJuDIqp2R9qC54biY3HwV1tWmfpJm6ijoAsrM,5257
+liger_kernel/transformers/model/llama.py,sha256=i8jJgyZsMKWQ-zKloETLugtwFpUOdaWxLDceciFXKd4,12832
+liger_kernel/transformers/model/llama4.py,sha256=IgbB8sTh3dlETQnaNNy1bZLuXy-Nt7qmeAjF27ydGpg,4210
+liger_kernel/transformers/model/llava.py,sha256=bLCioday_SOm69ogMDBhy_4UsVkH2-BSl93-EXY6-7I,15076
 liger_kernel/transformers/model/loss_utils.py,sha256=WWAMdiONPaXpIvxyOim_0igLrYh0yyOok5Q9_L9xvZw,1787
-liger_kernel/transformers/model/mistral.py,sha256=0lt1Jq37zWjxLZF-Vuj9jUyIEnWlMuT7PB5xB42KXBs,5313
-liger_kernel/transformers/model/mixtral.py,sha256=KpxDHtj7OCrZj_KrUWByRKM3A_x9o1S26rU3XGd1Ro8,11170
-liger_kernel/transformers/model/mllama.py,sha256=eElsJpBjdLfWhAZsYcfWnp_1tAf6t8jvliszu-v7sVg,11054
-liger_kernel/transformers/model/olmo2.py,sha256=FH_BY6pTiLgcjqsO1rprl9vcL_iZgBHBszelXgVj47Y,5033
-liger_kernel/transformers/model/paligemma.py,sha256=zXVV7FkhBnuHrbMg-CTOK21B90but6NqFd0DCeEefQE,18562
-liger_kernel/transformers/model/phi3.py,sha256=jYFqWcfP9wT9WUZeOC0SWjX_ZtWzQSDHDWH40m91TGE,10150
-liger_kernel/transformers/model/qwen2.py,sha256=b0fF5HX009VRrAGu9O2pG73YDDR05x_oy7JV9dvHuak,9432
-liger_kernel/transformers/model/qwen2_5_vl.py,sha256=F3lnFpKxTyij7ToEWc0hmXXyrdSsnbEfPSNCh9tAF0Y,8946
-liger_kernel/transformers/model/qwen2_vl.py,sha256=q3AMpxFfwHjaMu9Q3jpwpMPRzrE-eLqppg_8Z0ixjaQ,9357
-liger_kernel/transformers/model/qwen3.py,sha256=u_0cCRwr1jcwMkSknbBVb9my1OepCGU718uxKhNUOVM,4657
-liger_kernel/transformers/model/qwen3_moe.py,sha256=lIWGunVtNP-d7VfRvEGY820howzecb10g6ZeWRgsfl8,5463
+liger_kernel/transformers/model/mistral.py,sha256=syYNL8dLThX2-4uC13Lu0krEZ5zw3InviDUR3AJmc-I,5500
+liger_kernel/transformers/model/mixtral.py,sha256=VY-y73IyjcCyWyI7ahxXLw0fJrhgjYfr1xwRYtsHX0o,11396
+liger_kernel/transformers/model/mllama.py,sha256=my29NXk-p6ckQaP8qDIN8e318yI_9mQZHt38MV3SqLY,11280
+liger_kernel/transformers/model/olmo2.py,sha256=6L_bo-ZUgO1lYppdJneOtYxNIylQKS6BiGp13g7Uq9E,5259
+liger_kernel/transformers/model/paligemma.py,sha256=xuIx3oOwTgftU3jqLfWOxUxgCLBNJh0yNC21an9qDjo,18773
+liger_kernel/transformers/model/phi3.py,sha256=zAzBVNOA16B16yy2HWsEgOMHhLoYkpWOWPgBT4z95WI,10655
+liger_kernel/transformers/model/qwen2.py,sha256=3fpOTEOkniQmkCfN1KUa3KhseHJVzhj2Ht9FdYPUy-E,9962
+liger_kernel/transformers/model/qwen2_5_vl.py,sha256=zEVVwotCXnAm3RRc8-1Nc8uitSWrwW4B9dYY2uOZDwg,6331
+liger_kernel/transformers/model/qwen2_vl.py,sha256=5vK-vtCDpKZ2w33xYp2BS8kQYWUbKMqaiKvQcI27Mss,5884
+liger_kernel/transformers/model/qwen3.py,sha256=w2jBHuK9kK9EmOr5dnEIXNQXUgUSV_sJUkXSEwxLPHs,4885
+liger_kernel/transformers/model/qwen3_moe.py,sha256=BkpfFH3fOH0yRfA7LF-AoHTLut2GV0Y4MOlkiIYewfU,5511
+liger_kernel/transformers/model/smollm3.py,sha256=mqayvpwpMbp2yd_Ue7IPzy-dA4KHSDi_ROZW5vHCHfQ,7596
 liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7HHWHwku25A-GYL0WU,193
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
 liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
 liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
-liger_kernel-0.5.10.dist-info/licenses/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel-0.5.10.dist-info/licenses/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel-0.5.10.dist-info/METADATA,sha256=RZkuxdPfcMh52OXxN59XhLMmlDUSjGLd4uzf4SQmzIw,24014
-liger_kernel-0.5.10.dist-info/WHEEL,sha256=zaaOINJESkSfm_4HQVc5ssNzHCPXhJm0kEUakpsEHaU,91
-liger_kernel-0.5.10.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel-0.5.10.dist-info/RECORD,,
+liger_kernel-0.6.1.dist-info/licenses/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel-0.6.1.dist-info/licenses/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel-0.6.1.dist-info/METADATA,sha256=_of0e7dKufrp2upc26bnv4VLBZvAbcdDA8Fssm3mIfk,24545
+liger_kernel-0.6.1.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+liger_kernel-0.6.1.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel-0.6.1.dist-info/RECORD,,

{liger_kernel-0.5.10.dist-info → liger_kernel-0.6.1.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (80.8.0)
+Generator: setuptools (80.9.0)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

liger_kernel/transformers/gema3_rms.py

@@ -1,8 +0,0 @@
-from .rms_norm import LigerRMSNorm
-
-
-class LigerRMSNormForGemma3(LigerRMSNorm):
-    """Gemma3RMSNorm has a dim argument not hidden_size used in q_norm and k_norm."""
-
-    def __init__(self, dim, eps=0.000001, offset=1.0, casting_mode="gemma", init_fn="zeros", in_place=False):
-        super().__init__(dim, eps, offset, casting_mode, init_fn, in_place)