liger-kernel 0.5.9__py3-none-any.whl → 0.5.10__py3-none-any.whl

liger_kernel/transformers/model/mixtral.py

@@ -7,19 +7,13 @@ import torch
 
 from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import MoeCausalLMOutputWithPast
-from transformers.models.mixtral.modeling_mixtral import _CONFIG_FOR_DOC
-from transformers.models.mixtral.modeling_mixtral import MIXTRAL_INPUTS_DOCSTRING
 from transformers.models.mixtral.modeling_mixtral import load_balancing_loss_func
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-@add_start_docstrings_to_model_forward(MIXTRAL_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -146,8 +140,6 @@ def lce_forward_deprecated(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(MIXTRAL_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 # Ignore copy
 def lce_forward(
     self,

liger_kernel/transformers/model/mllama.py

@@ -8,17 +8,12 @@ import torch
 from torch.nn import CrossEntropyLoss
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.mllama.modeling_mllama import MLLAMA_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-@add_start_docstrings_to_model_forward(MLLAMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class="MllamaTextConfig")
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -135,8 +130,6 @@ def lce_forward_deprecated(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(MLLAMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class="MllamaTextConfig")
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/olmo2.py

@@ -6,18 +6,12 @@ from typing import Union
 import torch
 
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.olmo2.modeling_olmo2 import _CONFIG_FOR_DOC
-from transformers.models.olmo2.modeling_olmo2 import OLMO2_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(OLMO2_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/paligemma.py

@@ -7,13 +7,9 @@ import torch
 
 from torch.nn import CrossEntropyLoss
 from transformers.cache_utils import Cache
-from transformers.models.paligemma.modeling_paligemma import _CONFIG_FOR_DOC
-from transformers.models.paligemma.modeling_paligemma import PALIGEMMA_INPUTS_DOCSTRING
 from transformers.models.paligemma.modeling_paligemma import PaliGemmaCausalLMOutputWithPast
-from transformers.utils import add_start_docstrings_to_model_forward
 from transformers.utils import is_torchdynamo_compiling
 from transformers.utils import logging
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
@@ -21,8 +17,6 @@ from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinea
 logger = logging.get_logger(__name__)
 
 
-@add_start_docstrings_to_model_forward(PALIGEMMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=PaliGemmaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -206,8 +200,6 @@ def lce_forward_deprecated(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(PALIGEMMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=PaliGemmaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/phi3.py

@@ -7,18 +7,12 @@ import torch
 
 from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.phi3.modeling_phi3 import _CONFIG_FOR_DOC
-from transformers.models.phi3.modeling_phi3 import PHI3_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-@add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -128,8 +122,6 @@ def lce_forward_deprecated(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/qwen2.py

@@ -7,18 +7,12 @@ import torch
 
 from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.qwen2.modeling_qwen2 import _CONFIG_FOR_DOC
-from transformers.models.qwen2.modeling_qwen2 import QWEN2_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-@add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -127,8 +121,6 @@ def lce_forward_deprecated(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/qwen2_5_vl.py

@@ -6,17 +6,11 @@ from typing import Union
 import torch
 
 from torch.nn import CrossEntropyLoss
-from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import _CONFIG_FOR_DOC
-from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import QWEN2_5_VL_INPUTS_DOCSTRING
 from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLCausalLMOutputWithPast
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-@add_start_docstrings_to_model_forward(QWEN2_5_VL_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=Qwen2_5_VLCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/qwen2_vl.py

@@ -8,17 +8,11 @@ import torch
 from packaging import version
 from torch.nn import CrossEntropyLoss
 from transformers import __version__ as transformers_version
-from transformers.models.qwen2_vl.modeling_qwen2_vl import _CONFIG_FOR_DOC
-from transformers.models.qwen2_vl.modeling_qwen2_vl import QWEN2_VL_INPUTS_DOCSTRING
 from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLCausalLMOutputWithPast
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-@add_start_docstrings_to_model_forward(QWEN2_VL_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=Qwen2VLCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/qwen3.py

@@ -5,16 +5,10 @@ from typing import Union
 import torch
 
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.qwen3.modeling_qwen3 import _CONFIG_FOR_DOC
-from transformers.models.qwen3.modeling_qwen3 import QWEN3_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-@add_start_docstrings_to_model_forward(QWEN3_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: Optional[torch.LongTensor] = None,

liger_kernel/transformers/model/qwen3_moe.py

@@ -0,0 +1,128 @@
+from typing import List
+from typing import Optional
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import MoeCausalLMOutputWithPast
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.models.mixtral.modeling_mixtral import load_balancing_loss_func
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+def lce_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **loss_kwargs,
+) -> MoeCausalLMOutputWithPast:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Qwen3MoeForCausalLM
+
+    >>> model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+    >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_router_logits = (
+        output_router_logits if output_router_logits is not None else self.config.output_router_logits
+    )
+
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: MoeModelOutputWithPast = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+    kept_hidden_states = hidden_states[:, slice_indices, :]
+
+    shift_labels = loss_kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+
+    # if in training mode, do not materialize logits
+    if self.training and (labels is not None or shift_labels is not None):
+        loss = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.hidden_size,
+            **loss_kwargs,
+        )
+    else:  # if in inference model materialize logits
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            self.num_experts,
+            self.num_experts_per_tok,
+            attention_mask,
+        )
+        if labels is not None:
+            loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+    return MoeCausalLMOutputWithPast(
+        loss=loss,
+        aux_loss=aux_loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        router_logits=outputs.router_logits,
+    )

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):
@@ -1102,6 +1155,61 @@ def apply_liger_kernel_to_qwen3(
                 _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,
@@ -1455,6 +1563,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "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