liger-kernel 0.5.8__py3-none-any.whl → 0.5.10__py3-none-any.whl

liger_kernel/transformers/monkey_patch.py

@@ -35,6 +35,13 @@ from liger_kernel.transformers.swiglu import LigerBlockSparseTop2MLP
 from liger_kernel.transformers.swiglu import LigerPhi3SwiGLUMLP
 from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
 
+try:
+    import peft
+
+    PEFT_AVAILABLE = True
+except ImportError:
+    PEFT_AVAILABLE = False
+
 transformer_version = version.parse(transformers.__version__)
 
 logger = logging.getLogger(__name__)
@@ -48,22 +55,68 @@ def _bind_method_to_module(module, method_name: str, new_method: Callable):
 
 
 def _patch_rms_norm_module(module, offset=0.0, eps=1e-6, casting_mode="llama", in_place=True):
-    module.offset = offset
-    module.casting_mode = casting_mode
-    module.variance_epsilon = getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
-    module.in_place = in_place
-    _bind_method_to_module(module, "forward", LigerRMSNorm.forward)
-    _bind_method_to_module(module, "extra_repr", LigerRMSNorm.extra_repr)
-    module.__class__.__name__ = LigerRMSNorm.__name__
+    # Check if the module is a PEFT ModulesToSaveWrapper
+    # If it is, we need to patch the modules_to_save.default and original_modules
+    if PEFT_AVAILABLE and isinstance(module, peft.utils.other.ModulesToSaveWrapper):
+        module.modules_to_save.default.offset = offset
+        module.modules_to_save.default.casting_mode = casting_mode
+        module.modules_to_save.default.variance_epsilon = (
+            getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
+        )
+        module.modules_to_save.default.in_place = in_place
+        module.original_module.offset = offset
+        module.original_module.casting_mode = casting_mode
+        module.original_module.variance_epsilon = (
+            getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
+        )
+        module.original_module.in_place = in_place
+        _bind_method_to_module(module.modules_to_save.default, "forward", LigerRMSNorm.forward)
+        _bind_method_to_module(module.modules_to_save.default, "extra_repr", LigerRMSNorm.extra_repr)
+        _bind_method_to_module(module.original_module, "forward", LigerRMSNorm.forward)
+        _bind_method_to_module(module.original_module, "extra_repr", LigerRMSNorm.extra_repr)
+        module.modules_to_save.default.__class__.__name__ = LigerRMSNorm.__name__
+        module.original_module.__class__.__name__ = LigerRMSNorm.__name__
+    else:
+        module.offset = offset
+        module.casting_mode = casting_mode
+        module.variance_epsilon = getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
+        module.in_place = in_place
+        _bind_method_to_module(module, "forward", LigerRMSNorm.forward)
+        _bind_method_to_module(module, "extra_repr", LigerRMSNorm.extra_repr)
+        module.__class__.__name__ = LigerRMSNorm.__name__
 
 
 def _patch_layer_norm_module(module, eps=1e-6):
-    module.variance_epsilon = getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
-    module.hidden_size = getattr(module, "hidden_size", None) or getattr(module, "normalized_shape", None)
-
-    _bind_method_to_module(module, "forward", LigerLayerNorm.forward)
-    _bind_method_to_module(module, "extra_repr", LigerLayerNorm.extra_repr)
-    module.__class__.__name__ = LigerLayerNorm.__name__
+    # Check if the module is a PEFT ModulesToSaveWrapper
+    # If it is, we need to patch the modules_to_save.default and original_modules
+    if PEFT_AVAILABLE and isinstance(module, peft.utils.other.ModulesToSaveWrapper):
+        module.hidden_size = module.normalized_shape
+        _bind_method_to_module(module, "forward", LigerLayerNorm.forward)
+        _bind_method_to_module(module, "extra_repr", LigerLayerNorm.extra_repr)
+        module.modules_to_save.default.variance_epsilon = (
+            getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
+        )
+        module.original_module.hidden_size = getattr(module, "hidden_size", None) or getattr(
+            module, "normalized_shape", None
+        )
+        module.original_module.variance_epsilon = (
+            getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
+        )
+        module.original_module.hidden_size = getattr(module, "hidden_size", None) or getattr(
+            module, "normalized_shape", None
+        )
+        _bind_method_to_module(module.modules_to_save.default, "forward", LigerRMSNorm.forward)
+        _bind_method_to_module(module.modules_to_save.default, "extra_repr", LigerRMSNorm.extra_repr)
+        _bind_method_to_module(module.original_module, "forward", LigerRMSNorm.forward)
+        _bind_method_to_module(module.original_module, "extra_repr", LigerRMSNorm.extra_repr)
+        module.modules_to_save.default.__class__.__name__ = LigerLayerNorm.__name__
+        module.original_module.__class__.__name__ = LigerLayerNorm.__name__
+    else:
+        module.variance_epsilon = getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
+        module.hidden_size = getattr(module, "hidden_size", None) or getattr(module, "normalized_shape", None)
+        _bind_method_to_module(module, "forward", LigerLayerNorm.forward)
+        _bind_method_to_module(module, "extra_repr", LigerLayerNorm.extra_repr)
+        module.__class__.__name__ = LigerLayerNorm.__name__
 
 
 def _patch_swiglu_module(module, liger_module):
@@ -1048,6 +1101,115 @@ def apply_liger_kernel_to_qwen2(
     print("Applied Liger kernels to Qwen2")
 
 
+def apply_liger_kernel_to_qwen3(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Qwen3 models.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.qwen3 import modeling_qwen3
+    from transformers.models.qwen3.modeling_qwen3 import Qwen3Model
+
+    from liger_kernel.transformers.model.qwen3 import lce_forward as qwen3_lce_forward
+
+    if rope:
+        modeling_qwen3.apply_rotary_pos_emb = liger_rotary_pos_emb
+
+    if rms_norm:
+        modeling_qwen3.Qwen3RMSNorm = LigerRMSNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        modeling_qwen3.Qwen3ForCausalLM.forward = qwen3_lce_forward
+
+    if swiglu:
+        modeling_qwen3.Qwen3MLP = LigerSwiGLUMLP
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        # get the base model from the model instance
+        base_model: Qwen3Model = getattr(model, model.base_model_prefix, model)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.input_layernorm)
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+
+def apply_liger_kernel_to_qwen3_moe(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Qwen3 models.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.qwen3_moe import modeling_qwen3_moe
+    from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeModel
+
+    from liger_kernel.transformers.model.qwen3_moe import lce_forward as qwen3_lce_forward
+    from liger_kernel.transformers.swiglu import LigerQwen3MoeSwiGLUMLP
+
+    if rope:
+        modeling_qwen3_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
+
+    if rms_norm:
+        modeling_qwen3_moe.Qwen3MoeRMSNorm = LigerRMSNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        modeling_qwen3_moe.Qwen3MoeForCausalLM.forward = qwen3_lce_forward
+
+    if swiglu:
+        modeling_qwen3_moe.Qwen3MoeMLP = LigerQwen3MoeSwiGLUMLP
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        # get the base model from the model instance
+        base_model: Qwen3MoeModel = getattr(model, model.base_model_prefix, model)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _patch_swiglu_module(decoder_layer.mlp, LigerQwen3MoeSwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.input_layernorm)
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+
 def apply_liger_kernel_to_qwen2_vl(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -1319,12 +1481,78 @@ def apply_liger_kernel_to_olmo2(
                 _patch_rms_norm_module(decoder_layer.post_feedforward_layernorm, in_place=False)
 
 
+def apply_liger_kernel_to_glm4(
+    rope: bool = False,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace GLM-4 models.
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is False.
+        cross_entropy (bool): Whether to apply Liger's cross entropy loss. Default is False.
+        fused_linear_cross_entropy (bool):
+            Whether to apply Liger's fused linear cross entropy loss. Default is True.
+            `cross_entropy` and `fused_linear_cross_entropy` cannot both be True.
+            If `fused_linear_cross_entropy` is True, the logits will not be materialized but more memory efficient.
+        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is True.
+        swiglu (bool): Whether to apply Liger's SwiGLU Glm4MLP. Default is True.
+        model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
+        loaded. Default is None.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.glm4 import modeling_glm4
+    from transformers.models.glm4.modeling_glm4 import Glm4Model
+
+    from liger_kernel.transformers.model.glm4 import lce_forward as glm4_lce_forward
+
+    if rope:
+        raise NotImplementedError("liger_rotary_pos_emb is not available for Glm4 models.")
+    if rms_norm:
+        modeling_glm4.Glm4RMSNorm = partial(LigerRMSNorm, in_place=False)
+    if swiglu:
+        modeling_glm4.Glm4MLP = LigerPhi3SwiGLUMLP
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+    if fused_linear_cross_entropy:
+        modeling_glm4.Glm4ForCausalLM.forward = glm4_lce_forward
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        # get the base model from the model instance
+        base_model: Glm4Model = getattr(model, model.base_model_prefix, model)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm, in_place=False)
+
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _patch_swiglu_module(decoder_layer.mlp, LigerPhi3SwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.input_layernorm, in_place=False)
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm, in_place=False)
+                _patch_rms_norm_module(decoder_layer.post_self_attn_layernorm, in_place=False)
+                _patch_rms_norm_module(decoder_layer.post_mlp_layernorm, in_place=False)
+
+
 # Model type corresponds to the keys defined in transformers/models/auto/modeling_auto.py
 MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma": apply_liger_kernel_to_gemma,
     "gemma2": apply_liger_kernel_to_gemma2,
     "gemma3_text": apply_liger_kernel_to_gemma3_text,
     "gemma3": apply_liger_kernel_to_gemma3,
+    "glm4": apply_liger_kernel_to_glm4,
     "llama": apply_liger_kernel_to_llama,
     "llava": apply_liger_kernel_to_llava,
     "granite": apply_liger_kernel_to_granite,
@@ -1334,6 +1562,8 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "mixtral": apply_liger_kernel_to_mixtral,
     "olmo2": apply_liger_kernel_to_olmo2,
     "qwen2": apply_liger_kernel_to_qwen2,
+    "qwen3": apply_liger_kernel_to_qwen3,
+    "qwen3_moe": apply_liger_kernel_to_qwen3_moe,
     "qwen2_vl": apply_liger_kernel_to_qwen2_vl,
     "qwen2_5_vl": apply_liger_kernel_to_qwen2_5_vl,
     "phi3": apply_liger_kernel_to_phi3,

liger_kernel/transformers/sparsemax.py

@@ -0,0 +1,16 @@
+import torch
+import torch.nn as nn
+
+from liger_kernel.ops.sparsemax import LigerSparsemaxFunction
+
+
+class LigerSparsemax(nn.Module):
+    def __init__(self, dim: int = -1):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return LigerSparsemaxFunction.apply(x, self.dim)
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}"

liger_kernel/transformers/swiglu.py

@@ -56,3 +56,24 @@ class LigerPhi3SwiGLUMLP(nn.Module):
         up_states = self.gate_up_proj(x)
         gate, up_states = up_states.chunk(2, dim=-1)
         return self.down_proj(LigerSiLUMulFunction.apply(gate, up_states))
+
+
+class LigerQwen3MoeSwiGLUMLP(nn.Module):
+    """
+    Patch Qwen3MoeMLP to use LigerSiLUMulFunction.
+    https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/qwen3_moe/modular_qwen3_moe.py#L57
+    """
+
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size if intermediate_size is not None else config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        if config.hidden_act not in ["silu", "swish"]:
+            raise ValueError(f"Activation function {config.hidden_act} not supported.")
+
+    def forward(self, x):
+        return self.down_proj(LigerSiLUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))

liger_kernel/transformers/trainer/orpo_trainer.py

@@ -1,5 +1,3 @@
-from typing import Any
-from typing import Callable
 from typing import Dict
 from typing import List
 from typing import Literal
@@ -13,57 +11,7 @@ from torch.distributed.fsdp import FullyShardedDataParallel
 from trl.trainer import ORPOTrainer
 
 from liger_kernel.chunked_loss import LigerFusedLinearORPOLoss
-
-
-class _FSDPForwardRedirection:
-    """
-    Modified based on
-    https://github.com/Lightning-AI/pytorch-lightning/blob/d3f9c83d6efa4f1def36aa6c199600946cdb9117/src/lightning/pytorch/strategies/strategy.py#L601-L648
-    Redirect a method call through FullyShardedDataParallel.forward so that the FSDP module's root pre-forward and
-    post-forward can be properly executed around the method call.
-    This is needed in cases where we call a submodule of a FSDP module. For instance, when we want to call only
-    the `LlamaModel` part out of a FSDP-wrapped `LlamaForCausalLM` to get the hidden states without involving
-    GPU-memory-heavy `lm_head` and cross entropy computation, doing this directly (i.e. `model.model.forward()`)
-    will not work because the first `nn.Embedding` layer is not independently wrapped as a FSDP module (because of
-    the transformer-based wrapping policy), and not calling it through FSDP root module forward will not all-gather
-    its parameter, thus resulting in "RuntimeError: 'weight' must be 2-D" error. Similarly, if we want to call just
-    the `lm_head` part of a model, we need this trick too to properly get its params all-gathered.
-    """
-
-    def __call__(
-        self,
-        wrapper_module: FullyShardedDataParallel,
-        method: Callable,
-        *args: Any,
-        **kwargs: Any,
-    ):
-        """Reroutes a method call through the `wrapper_module`'s `forward` method.
-        Args:
-            wrapper_module: The module that has `original_module` wrapped.
-            original_module: The module that was wrapped inside `wrapper_module`.
-            method_name: The name of the method that should be called on the `original_module` after inputs get
-                redirected through the `wrapper_module`'s `forward` method.
-            *args: The positional arguments to the method `method_name`. They will get passed to a patched
-                `forward` method instead.
-            **kwargs: The keyword arguments to the method `method_name`. They will get passed to a patched
-                `forward` method instead.
-        """
-        assert isinstance(wrapper_module, FullyShardedDataParallel)
-        original_module = wrapper_module._fsdp_wrapped_module
-        original_forward = original_module.forward
-
-        def wrapped_forward(*_args: Any, **_kwargs: Any) -> Any:
-            # Unpatch ourselves immediately before calling the method `method_name`
-            # because itself may want to call the real `forward`
-            original_module.forward = original_forward  # type: ignore[method-assign]
-            # Call the actual method e.g. `.training_step(...)`
-            out = method(*_args, **_kwargs)
-            return out
-
-        # Patch the original_module's forward so we can redirect the arguments back to the real method
-        original_module.forward = wrapped_forward  # type: ignore[method-assign]
-        wrapper_output = wrapper_module(*args, **kwargs)
-        return wrapper_output
+from liger_kernel.transformers.fsdp import _FSDPForwardRedirection
 
 
 class LigerORPOTrainer(ORPOTrainer):

liger_kernel/utils.py

@@ -1,6 +1,17 @@
+try:
+    import peft  # noqa: F401
+
+    PEFT_AVAILABLE = True
+except ImportError:
+    PEFT_AVAILABLE = False
+
 import torch
 
 
+def is_peft_available():
+    return PEFT_AVAILABLE
+
+
 def infer_device():
     """
     Get current device name based on available devices

{liger_kernel-0.5.8.dist-info → liger_kernel-0.5.10.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: liger_kernel
-Version: 0.5.8
+Version: 0.5.10
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -59,7 +59,6 @@ Dynamic: requires-dist
         <th style="padding: 10px;" colspan="2">Stable</th>
         <th style="padding: 10px;" colspan="2">Nightly</th>
         <th style="padding: 10px;">Discord</th>
-        <th style="padding: 10px;">Build</th>
     </tr>
     <tr>
         <td style="padding: 10px;">
@@ -87,23 +86,6 @@ Dynamic: requires-dist
                 <img src="https://dcbadge.vercel.app/api/server/gpumode?style=flat" alt="Join Our Discord">
             </a>
         </td>
-        <td style="padding: 10px;">
-            <div style="display: block;">
-                <a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml">
-                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?event=schedule" alt="Build">
-                </a>
-            </div>
-            <div style="display: block;">
-                <a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml">
-                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?event=schedule" alt="Build">
-                </a>
-            </div>
-            <div style="display: block;">
-                <a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml">
-                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml/badge.svg?event=schedule" alt="Build">
-                </a>
-            </div>
-        </td>
     </tr>
 </table>
 
@@ -320,9 +302,12 @@ loss.backward()
 | Qwen2, Qwen2.5, & QwQ      | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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       |
 | 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 |
+| GLM-4   | `liger_kernel.transformers.apply_liger_kernel_to_glm4`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 
 
 ## Low-level APIs
@@ -340,7 +325,8 @@ loss.backward()
 | SwiGLU                          | `liger_kernel.transformers.LigerSwiGLUMLP`                  |
 | GeGLU                           | `liger_kernel.transformers.LigerGEGLUMLP`                   |
 | CrossEntropy                    | `liger_kernel.transformers.LigerCrossEntropyLoss`           |
-| Fused Linear CrossEntropy         | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+| Fused Linear CrossEntropy       | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+| Sparsemax                       | `liger_kernel.transformers.LigerSparsemax`                  |
 
 
 ### Alignment Kernels
@@ -388,6 +374,36 @@ loss.backward()
 - [Axolotl](https://axolotl.ai/): Integrating Liger Kernel into Axolotl.
 - [Llama-Factory](https://github.com/hiyouga/LLaMA-Factory): Integrating Liger Kernel into Llama-Factory.
 
+
+## CI status
+
+<table style="width: 100%; text-align: center; border-collapse: collapse;">
+    <tr>
+        <th style="padding: 10px;">Build</th>
+    </tr>
+    <tr>
+        <td style="padding: 10px;">
+            <div style="display: block;">
+                <a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?event=schedule" alt="Build">
+                </a>
+            </div>
+            <div style="display: block;">
+                <a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?event=schedule" alt="Build">
+                </a>
+            </div>
+            <div style="display: block;">
+                <a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml/badge.svg?event=schedule" alt="Build">
+                </a>
+            </div>
+        </td>
+    </tr>
+</table>
+
+
+
 ## Contact
 
 - For issues, create a Github ticket in this repository