liger-kernel-nightly 0.6.4.dev20251202054858__py3-none-any.whl → 0.6.4.dev20260107181130__py3-none-any.whl

liger_kernel/transformers/model/gemma3.py

@@ -235,6 +235,7 @@ def multimodal_forward(
         **lm_kwargs,
     )
 
+    shift_labels = lm_kwargs.pop("shift_labels", None)
     hidden_states = outputs[0]
 
     slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep

liger_kernel/transformers/model/gpt_oss.py

@@ -0,0 +1,211 @@
+from typing import List
+from typing import Optional
+from typing import Union
+
+import torch
+
+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
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerMoeCausalLMOutputWithPast
+
+
+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,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> LigerMoeCausalLMOutputWithPast:
+    r"""
+        Forward pass for causal language modeling with Mixture of Experts (MoE) architecture using Liger Kernel optimizations.
+
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of input sequence tokens in the vocabulary. Indices can be obtained using tokenizers.
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings.
+        past_key_values (`List[torch.FloatTensor]` or `Cache`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks) that can be used to speed up
+            sequential decoding. See `past_key_values` input for more details.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        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]`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+            (see `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        output_router_logits (`bool`, *optional*):
+            Whether or not to return the router logits of all MoE layers. See `router_logits` under returned tensors
+            for more detail.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        logits_to_keep (`int` or `torch.Tensor`, *optional*, defaults to 0):
+            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).
+        skip_logits (`bool`, *optional*):
+            Whether to skip logit computation and directly compute loss. If `None`, defaults to `True` during training
+            when labels are provided (to save memory), and `False` during inference.
+
+    Returns:
+        `LigerMoeCausalLMOutputWithPast`: An output object containing:
+            - loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+                Language modeling loss (for next-token prediction), including the auxiliary load balancing loss.
+            - aux_loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
+                Auxiliary load balancing loss for the sparse MoE modules.
+            - logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`, *optional*):
+                Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+                Note: logits are `None` during training when `skip_logits=True` to save memory.
+            - past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed):
+                Cached key and value projection states for faster sequential decoding.
+            - hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True`):
+                Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for each layer) of shape
+                `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer.
+            - attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True`):
+                Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+                sequence_length)`. Attentions weights after the attention softmax.
+            - router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True`):
+                Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
+                Router logits of the MoE layers, useful to compute the auxiliary loss and z_loss.
+            - token_accuracy (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
+                Token-level prediction accuracy.
+
+    Example:
+
+    ```python
+        >>> from transformers import AutoTokenizer, GptOssForCausalLM
+        >>> from liger_kernel.transformers import apply_liger_kernel_to_gpt_oss
+
+        >>> # Apply Liger Kernel patches for optimized performance
+        >>> apply_liger_kernel_to_gpt_oss()
+
+        >>> model = GptOssForCausalLM.from_pretrained("openai/gpt-oss-20b")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/gpt-oss-20b")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Inference: Forward pass returns logits
+        >>> outputs = model(**inputs)
+        >>> outputs.logits.shape
+        torch.Size([1, 12, 201088])
+
+        >>> # Get next token prediction
+        >>> next_token_logits = outputs.logits[:, -1, :]
+        >>> predicted_token_id = next_token_logits.argmax(dim=-1)
+
+        >>> # Training: Forward pass with labels returns loss
+        >>> labels = inputs.input_ids.clone()
+        >>> outputs = model(**inputs, labels=labels)
+        >>> outputs.loss
+        tensor(2.6454)
+    ```"""
+
+    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,
+        **kwargs,
+    )
+
+    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 = kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+    token_accuracy = None
+
+    if skip_logits is None:
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    if skip_logits:
+        result = 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,
+            **kwargs,
+        )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
+    else:  # if in inference model materialize logits
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.vocab_size,
+                **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 LigerMoeCausalLMOutputWithPast(
+        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,
+        token_accuracy=token_accuracy,
+    )

liger_kernel/transformers/model/paligemma.py

@@ -330,6 +330,7 @@ def lce_forward(
         **lm_kwargs,
     )
 
+    shift_labels = lm_kwargs.pop("shift_labels", None)
     hidden_states = outputs[0]
 
     loss = None

liger_kernel/transformers/monkey_patch.py

@@ -20,6 +20,7 @@ from liger_kernel.transformers.model.gemma import lce_forward as gemma_lce_forwa
 from liger_kernel.transformers.model.gemma import lce_forward_deprecated as gemma_lce_forward_deprecated
 from liger_kernel.transformers.model.gemma2 import lce_forward as gemma2_lce_forward
 from liger_kernel.transformers.model.gemma2 import lce_forward_deprecated as gemma2_lce_forward_deprected
+from liger_kernel.transformers.model.gpt_oss import lce_forward as gpt_oss_lce_forward
 from liger_kernel.transformers.model.llama import lce_forward as llama_lce_forward
 from liger_kernel.transformers.model.llama import lce_forward_deprecated as llama_lce_forward_deprecated
 from liger_kernel.transformers.model.llava import lce_forward as llava_lce_forward
@@ -34,8 +35,7 @@ from liger_kernel.transformers.model.smollm3 import lce_forward as smollm3_lce_f
 from liger_kernel.transformers.qwen2vl_mrope import liger_multimodal_rotary_pos_emb
 from liger_kernel.transformers.rms_norm import LigerRMSNorm
 from liger_kernel.transformers.rope import liger_rotary_pos_emb
-from liger_kernel.transformers.rope import liger_rotary_pos_emb_with_cast
-from liger_kernel.transformers.rope import liger_rotary_pos_emb_with_cast_and_leading_batch
+from liger_kernel.transformers.rope import liger_rotary_pos_emb_vision
 from liger_kernel.transformers.swiglu import LigerBlockSparseTop2MLP
 from liger_kernel.transformers.swiglu import LigerPhi3SwiGLUMLP
 from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
@@ -430,7 +430,7 @@ def apply_liger_kernel_to_llava(
                     f"These parameters are not supported by {text_model_name}. Enter the remaining {list(text_kwargs.keys())} except for {list(remain_params)}\n"
                     f"Parameters accepted by {text_model_name}: {list(accept_params.keys())}"
                 )
-            text_kwargs["model"] = model.language_model
+            text_kwargs["model"] = model.model.language_model
             text_liger_fn(**text_kwargs)
         elif text_model_name not in MODEL_TYPE_TO_APPLY_LIGER_FN:
             logger.warning(f"{text_model_name} is not supported by Liger kernel.")
@@ -445,7 +445,7 @@ def apply_liger_kernel_to_llava(
                     f"These parameters are not supported by {vision_model_name}. Enter the remaining {list(vision_kwargs.keys())} except for {list(remain_params)}\n"
                     f"Parameters accepted by {vision_model_name}: {list(accept_params.keys())}"
                 )
-            vision_kwargs["model"] = model.vision_tower
+            vision_kwargs["model"] = model.model.vision_tower
             vision_liger_fn(**vision_kwargs)
         elif vision_model_name not in MODEL_TYPE_TO_APPLY_LIGER_FN:
             logger.warning(f"{vision_model_name} is not supported by Liger kernel.")
@@ -615,8 +615,8 @@ def apply_liger_kernel_to_mllama(
         # instance variables that reference already-instantiated modules
 
         if isinstance(model, MllamaForConditionalGeneration):
-            language_model: MllamaForCausalLM = model.language_model
-            vision_model: MllamaVisionModel = model.vision_model
+            language_model: MllamaForCausalLM = model.model.language_model
+            vision_model: MllamaVisionModel = model.model.vision_model
             if isinstance(language_model, MllamaForCausalLM):
                 text_model: MllamaTextModel = language_model.model
             else:
@@ -1118,8 +1118,8 @@ def apply_liger_kernel_to_gemma3(
         # instance variables that reference already-instantiated modules
 
         if isinstance(model, Gemma3ForConditionalGeneration):
-            if isinstance(model.vision_tower, SiglipVisionModel):
-                vision_tower = model.vision_tower
+            if isinstance(model.model.vision_tower, SiglipVisionModel):
+                vision_tower = model.model.vision_tower
 
                 _patch_layer_norm_module(vision_tower.vision_model.post_layernorm)
 
@@ -1132,7 +1132,7 @@ def apply_liger_kernel_to_gemma3(
                 raise TypeError("The vision tower must be SiglipVisionModel")
 
             if rms_norm:
-                _patch_rms_norm_module_for_gemma3(model.multi_modal_projector.mm_soft_emb_norm)
+                _patch_rms_norm_module_for_gemma3(model.model.multi_modal_projector.mm_soft_emb_norm)
 
             apply_liger_kernel_to_gemma3_text(
                 rope=rope,
@@ -1140,7 +1140,7 @@ def apply_liger_kernel_to_gemma3(
                 fused_linear_cross_entropy=False,
                 rms_norm=rms_norm,
                 geglu=geglu,
-                model=model.language_model,
+                model=model.model.language_model,
             )
 
         else:
@@ -1228,7 +1228,7 @@ def apply_liger_kernel_to_paligemma(
         if not isinstance(model, PaliGemmaForConditionalGeneration):
             raise TypeError("model have to be of type PaliGemmaForConditionalGeneration")
 
-        vision_tower: SiglipVisionModel = model.vision_tower
+        vision_tower: SiglipVisionModel = model.model.vision_tower
 
         _patch_layer_norm_module(vision_tower.vision_model.post_layernorm)
 
@@ -1238,7 +1238,7 @@ def apply_liger_kernel_to_paligemma(
                 _patch_layer_norm_module(layer.layer_norm1)
                 _patch_layer_norm_module(layer.layer_norm2)
 
-        language_model = model.language_model
+        language_model = model.model.language_model
 
         if isinstance(language_model, (GemmaForCausalLM, GemmaModel)):
             apply_liger_kernel_to_gemma(
@@ -1459,6 +1459,79 @@ def apply_liger_kernel_to_qwen3_moe(
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_gpt_oss(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = False,  # Set to False by default since GPT-OSS has custom expert implementation
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace GPT-OSS models.
+    NOTE: GPT-OSS is supported in transformers >= 4.55.0
+    NOTE: SwiGLU patching is disabled by default for GPT-OSS as it uses a custom expert
+          implementation with clamping and MXFP4 quantization.
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is True.
+        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 MLP. Default is False.
+            Note: GPT-OSS uses a custom expert implementation, so SwiGLU patching is disabled by default.
+        model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
+            loaded. Default is None.
+    """
+    if version.parse(transformers.__version__) < version.parse("4.55.0"):
+        logger.warning("GPT-OSS support requires transformers >= 4.55.0")
+        return
+
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.gpt_oss import modeling_gpt_oss
+    from transformers.models.gpt_oss.modeling_gpt_oss import GptOssModel
+
+    if rope:
+        modeling_gpt_oss.apply_rotary_pos_emb = liger_rotary_pos_emb
+
+    if rms_norm:
+        modeling_gpt_oss.GptOssRMSNorm = LigerRMSNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        if model is not None:
+            model.forward = MethodType(gpt_oss_lce_forward, model)
+        else:
+            modeling_gpt_oss.GptOssForCausalLM.forward = gpt_oss_lce_forward
+
+    # Note: SwiGLU patching is not implemented for GPT-OSS due to custom expert implementation
+    # with clamping (swiglu_limit=7.0) and MXFP4 quantization
+
+    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: GptOssModel = 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 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,
@@ -1520,11 +1593,10 @@ def apply_liger_kernel_to_qwen2_vl(
     if model is not None:
         # The model instance already exists, so we need to additionally patch the
         # instance variables that reference already-instantiated modules
-
-        if isinstance(model, (Qwen2VLForConditionalGeneration, Qwen2VLModel)):
-            # Note: language_model and visual properties can be accessed throught conditional class for BC.
-            # Not sure if it is subject to changes in the future.
-            # Reference: https://github.com/huggingface/transformers/blob/v4.52.4/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py#L1698
+        if isinstance(model, Qwen2VLForConditionalGeneration):
+            text_model: Qwen2VLTextModel = model.model.language_model
+            vision_model: Qwen2VisionTransformerPretrainedModel = model.model.visual
+        elif isinstance(model, Qwen2VLModel):
             text_model: Qwen2VLTextModel = model.language_model
             vision_model: Qwen2VisionTransformerPretrainedModel = model.visual
         elif isinstance(model, Qwen2VLTextModel):
@@ -1611,11 +1683,10 @@ def apply_liger_kernel_to_qwen2_5_vl(
     if model is not None:
         # The model instance already exists, so we need to additionally patch the
         # instance variables that reference already-instantiated modules
-
-        if isinstance(model, (Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLModel)):
-            # Note: language_model and visual properties can be accessed throught conditional class for BC.
-            # Not sure if it is subject to changes in the future.
-            # Reference: https://github.com/huggingface/transformers/blob/v4.52.4/src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py#L1823
+        if isinstance(model, Qwen2_5_VLForConditionalGeneration):
+            text_model: Qwen2_5_VLTextModel = model.model.language_model
+            vision_model: Qwen2_5_VisionTransformerPretrainedModel = model.model.visual
+        elif isinstance(model, Qwen2_5_VLModel):
             text_model: Qwen2_5_VLTextModel = model.language_model
             vision_model: Qwen2_5_VisionTransformerPretrainedModel = model.visual
         elif isinstance(model, Qwen2_5_VLTextModel):
@@ -1629,7 +1700,7 @@ def apply_liger_kernel_to_qwen2_5_vl(
 
         if vision_model is not None:
             # Patch Qwen2_5_VisionTransformerPretrainedModel
-            for vision_block in model.visual.blocks:
+            for vision_block in vision_model.blocks:
                 if rms_norm:
                     _patch_rms_norm_module(vision_block.norm1)
                     _patch_rms_norm_module(vision_block.norm2)
@@ -1680,8 +1751,8 @@ def apply_liger_kernel_to_qwen3_vl(
     from liger_kernel.transformers.model.qwen3_vl import lce_forward as qwen3_vl_lce_forward
 
     if rope:
-        modeling_qwen3_vl.apply_rotary_pos_emb = liger_rotary_pos_emb_with_cast
-        modeling_qwen3_vl.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_with_cast_and_leading_batch
+        modeling_qwen3_vl.apply_rotary_pos_emb = liger_rotary_pos_emb
+        modeling_qwen3_vl.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_vision
 
     if rms_norm:
         modeling_qwen3_vl.Qwen3VLTextRMSNorm = LigerRMSNorm
@@ -1698,7 +1769,9 @@ def apply_liger_kernel_to_qwen3_vl(
             modeling_qwen3_vl.Qwen3VLForConditionalGeneration.forward = qwen3_vl_lce_forward
 
     if model is not None and rms_norm:
-        if isinstance(model, (Qwen3VLForConditionalGeneration, Qwen3VLModel)):
+        if isinstance(model, Qwen3VLForConditionalGeneration):
+            text_model: Qwen3VLTextModel = model.model.language_model
+        elif isinstance(model, Qwen3VLModel):
             text_model: Qwen3VLTextModel = model.language_model
         elif isinstance(model, Qwen3VLTextModel):
             text_model = model
@@ -1755,8 +1828,8 @@ def apply_liger_kernel_to_qwen3_vl_moe(
     from liger_kernel.transformers.model.qwen3_vl_moe import lce_forward as qwen3_vl_moe_lce_forward
 
     if rope:
-        modeling_qwen3_vl_moe.apply_rotary_pos_emb = liger_rotary_pos_emb_with_cast
-        modeling_qwen3_vl_moe.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_with_cast_and_leading_batch
+        modeling_qwen3_vl_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
+        modeling_qwen3_vl_moe.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_vision
 
     if rms_norm:
         modeling_qwen3_vl_moe.Qwen3VLMoeTextRMSNorm = LigerRMSNorm
@@ -1773,7 +1846,9 @@ def apply_liger_kernel_to_qwen3_vl_moe(
             modeling_qwen3_vl_moe.Qwen3VLMoeForConditionalGeneration.forward = qwen3_vl_moe_lce_forward
 
     if model is not None and rms_norm:
-        if isinstance(model, (Qwen3VLMoeForConditionalGeneration, Qwen3VLMoeModel)):
+        if isinstance(model, Qwen3VLMoeForConditionalGeneration):
+            text_model: Qwen3VLMoeTextModel = model.model.language_model
+        elif isinstance(model, Qwen3VLMoeModel):
             text_model: Qwen3VLMoeTextModel = model.language_model
         elif isinstance(model, Qwen3VLMoeTextModel):
             text_model = model
@@ -2118,10 +2193,10 @@ def apply_liger_kernel_to_glm4v(
     if model is not None:
         # The model instance already exists, so we need to additionally patch the
         # instance variables that reference already-instantiated modules
-        if isinstance(model, (Glm4vForConditionalGeneration, Glm4vModel)):
-            # Note: language_model and visual properties can be accessed throught conditional class for BC.
-            # Not sure if it is subject to changes in the future.
-            # Reference: https://github.com/huggingface/transformers/blob/main/src/transformers/models/glm4v/modeling_glm4v.py#L1305
+        if isinstance(model, Glm4vForConditionalGeneration):
+            text_model: Glm4vTextModel = model.model.language_model
+            vision_model: Glm4vVisionModel = model.model.visual
+        elif isinstance(model, Glm4vModel):
             text_model: Glm4vTextModel = model.language_model
             vision_model: Glm4vVisionModel = model.visual
         elif isinstance(model, Glm4vTextModel):
@@ -2208,10 +2283,11 @@ def apply_liger_kernel_to_glm4v_moe(
     if model is not None:
         # The model instance already exists, so we need to additionally patch the
         # instance variables that reference already-instantiated modules
-        if isinstance(model, (Glm4vMoeForConditionalGeneration, Glm4vMoeModel)):
-            # Note: language_model and visual properties can be accessed throught conditional class for BC.
-            # Not sure if it is subject to changes in the future.
-            # Reference: https://github.com/huggingface/transformers/blob/main/src/transformers/models/glm4v_moe/modeling_glm4v_moe.py#L337
+        if isinstance(model, Glm4vMoeForConditionalGeneration):
+            text_model: Glm4vMoeTextModel = model.model.language_model
+            vision_model: Glm4vMoeVisionModel = model.model.visual
+            Glm4vMoeTextMoE = modeling_glm4v_moe.Glm4vMoeTextMoE
+        elif isinstance(model, Glm4vMoeModel):
             text_model: Glm4vMoeTextModel = model.language_model
             vision_model: Glm4vMoeVisionModel = model.visual
             Glm4vMoeTextMoE = modeling_glm4v_moe.Glm4vMoeTextMoE
@@ -2314,8 +2390,10 @@ def apply_liger_kernel_to_internvl(
     if model is not None:
         # The model instance already exists, so we need to additionally patch the
         # instance variables that reference already-instantiated modules
-        if isinstance(model, (InternVLForConditionalGeneration, InternVLModel)):
-            # NOTE: language_model and visual properties can be accessed throught conditional class.
+        if isinstance(model, InternVLForConditionalGeneration):
+            text_model = model.model.language_model
+            vision_model: InternVLVisionModel = model.model.vision_tower
+        elif isinstance(model, InternVLModel):
             text_model = model.language_model
             vision_model: InternVLVisionModel = model.vision_tower
         else:
@@ -2752,6 +2830,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "glm4": apply_liger_kernel_to_glm4,
     "glm4v": apply_liger_kernel_to_glm4v,
     "glm4v_moe": apply_liger_kernel_to_glm4v_moe,
+    "gpt_oss": apply_liger_kernel_to_gpt_oss,
     "internvl": apply_liger_kernel_to_internvl,
     "llama": apply_liger_kernel_to_llama,
     "llama4_text": apply_liger_kernel_to_llama4,

liger_kernel/transformers/multi_token_attention.py

@@ -5,7 +5,7 @@ import torch.nn as nn
 
 from torch.nn.modules.utils import _pair
 
-from liger_kernel.ops.multi_token_attention import LigerMultiTokenAttentionFunction
+from liger_kernel.ops import LigerMultiTokenAttentionFunction
 
 
 class LigerMultiTokenAttention(nn.Module):

liger_kernel/transformers/poly_norm.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 
-from liger_kernel.ops.poly_norm import LigerPolyNormFunction
+from liger_kernel.ops import LigerPolyNormFunction
 
 
 class LigerPolyNorm(nn.Module):

liger_kernel/transformers/qwen2vl_mrope.py

@@ -1,4 +1,4 @@
-from liger_kernel.ops.qwen2vl_mrope import LigerQwen2VLMRopeFunction
+from liger_kernel.ops import LigerQwen2VLMRopeFunction
 
 
 def liger_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):

liger_kernel/transformers/rms_norm.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 
-from liger_kernel.ops.rms_norm import LigerRMSNormFunction
+from liger_kernel.ops import LigerRMSNormFunction
 
 
 class LigerRMSNorm(nn.Module):
@@ -14,13 +14,18 @@ class LigerRMSNorm(nn.Module):
         init_fn="ones",
         in_place=True,
         row_mode=None,
+        elementwise_affine=True,
     ):
         super().__init__()
         assert init_fn in [
             "ones",
             "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.elementwise_affine = elementwise_affine
+        if self.elementwise_affine:
+            self.weight = nn.Parameter(torch.ones(hidden_size) if init_fn == "ones" else torch.zeros(hidden_size))
+        else:
+            self.register_parameter("weight", None)
         self.variance_epsilon, self.offset, self.casting_mode, self.in_place, self.row_mode = (
             eps,
             offset,
@@ -41,7 +46,7 @@ class LigerRMSNorm(nn.Module):
         )
 
     def extra_repr(self):
-        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}, offset={self.offset}, in_place={self.in_place}, row_mode={self.row_mode}"
+        return f"weight_shape={tuple(self.weight.shape) if self.weight is not None else None}, eps={self.variance_epsilon}, offset={self.offset}, in_place={self.in_place}, row_mode={self.row_mode}"
 
 
 class LigerRMSNormForGemma(LigerRMSNorm):