liger-kernel-nightly 0.6.3.dev20251031170118__py3-none-any.whl → 0.6.3.dev20251105012545__py3-none-any.whl

liger_kernel/transformers/__init__.py

@@ -56,6 +56,8 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_moe  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_next  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_vl  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_vl_moe  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_smollm3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_smolvlm  # noqa: F401
 
@@ -120,6 +122,8 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen3",
         "apply_liger_kernel_to_qwen3_moe",
         "apply_liger_kernel_to_qwen3_next",
+        "apply_liger_kernel_to_qwen3_vl",
+        "apply_liger_kernel_to_qwen3_vl_moe",
         "apply_liger_kernel_to_smollm3",
         "apply_liger_kernel_to_smolvlm",
     }
@@ -190,6 +194,8 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_qwen3",
             "apply_liger_kernel_to_qwen3_moe",
             "apply_liger_kernel_to_qwen3_next",
+            "apply_liger_kernel_to_qwen3_vl",
+            "apply_liger_kernel_to_qwen3_vl_moe",
             "apply_liger_kernel_to_smollm3",
             "apply_liger_kernel_to_smolvlm",
         ]

liger_kernel/transformers/model/qwen3_vl.py

@@ -0,0 +1,144 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLCausalLMOutputWithPast
+from transformers.utils import can_return_tuple
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+@can_return_tuple
+def lce_forward(
+    self,
+    input_ids: 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,
+    return_dict: Optional[bool] = None,
+    pixel_values: Optional[torch.Tensor] = None,
+    pixel_values_videos: Optional[torch.FloatTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    video_grid_thw: Optional[torch.LongTensor] = None,
+    rope_deltas: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    second_per_grid_ts: Optional[torch.Tensor] = None,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, Qwen3VLCausalLMOutputWithPast]:
+    """
+    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]`.
+    pixel_values_videos (`torch.FloatTensor` of shape `(seq_length, num_channels * temporal_size * image_size * image_size)):
+        The tensors corresponding to the input videos. Pixel values can be obtained using
+        [`AutoImageProcessor`]. See [`Qwen2_5_VLImageProcessor.__call__`] for details. [`Qwen2_5_VLProcessor`] uses
+        [`Qwen2_5_VLImageProcessor`] for processing videos.
+    image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+        The temporal, height and width of feature shape of each image in LLM.
+    video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+        The temporal, height and width of feature shape of each video in LLM.
+    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+        The rope index difference between sequence length and multimodal rope.
+    second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
+        The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
+    Example:
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
+    >>> model = Qwen3VLForConditionalGeneration.from_pretrained("Qwen/Qwen3-VL")
+    >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL")
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image"},
+                {"type": "text", "text": "What is shown in this image?"},
+            ],
+        },
+    ]
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+    >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])
+    >>> # 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]
+    "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
+    ```"""
+
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    outputs = self.model(
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        pixel_values_videos=pixel_values_videos,
+        image_grid_thw=image_grid_thw,
+        video_grid_thw=video_grid_thw,
+        second_per_grid_ts=second_per_grid_ts,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+
+    shift_labels = kwargs.pop("shift_labels", None)
+    loss = None
+    logits = None
+
+    if skip_logits and labels is None and shift_labels is None:
+        raise ValueError("skip_logits is True, but labels and shift_labels are None")
+
+    if skip_logits is None:
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    if skip_logits:
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.text_config.hidden_size,
+            **kwargs,
+        )
+    else:
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Qwen3VLCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=outputs.rope_deltas,
+    )

liger_kernel/transformers/model/qwen3_vl_moe.py

@@ -0,0 +1,119 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import Qwen3VLMoeCausalLMOutputWithPast
+from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import load_balancing_loss_func
+from transformers.utils import can_return_tuple
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+@can_return_tuple
+def lce_forward(
+    self,
+    input_ids: 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,
+    return_dict: Optional[bool] = None,
+    pixel_values: Optional[torch.Tensor] = None,
+    pixel_values_videos: Optional[torch.FloatTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    video_grid_thw: Optional[torch.LongTensor] = None,
+    rope_deltas: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    second_per_grid_ts: Optional[torch.Tensor] = None,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, Qwen3VLMoeCausalLMOutputWithPast]:
+    """
+    Qwen3-VL-MoE forward with fused linear cross entropy support mirroring Qwen3-VL behaviour.
+    """
+
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    outputs = self.model(
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        pixel_values_videos=pixel_values_videos,
+        image_grid_thw=image_grid_thw,
+        video_grid_thw=video_grid_thw,
+        second_per_grid_ts=second_per_grid_ts,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+
+    shift_labels = kwargs.pop("shift_labels", None)
+    loss = None
+    logits = None
+
+    if skip_logits and labels is None and shift_labels is None:
+        raise ValueError("skip_logits is True, but labels and shift_labels are None")
+
+    if skip_logits is None:
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    if skip_logits:
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.text_config.hidden_size,
+            **kwargs,
+        )
+    else:
+        logits = self.lm_head(hidden_states)
+
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size)
+
+    # Compute auxiliary load-balancing loss for MoE when requested
+    aux_loss = None
+    if kwargs.get("output_router_logits", False):
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            self.config.text_config.num_experts,
+            self.config.text_config.num_experts_per_tok,
+            attention_mask,
+        )
+        # If we computed training loss, add the scaled aux loss to it
+        if loss is not None and aux_loss is not None:
+            loss = loss + self.config.text_config.router_aux_loss_coef * aux_loss.to(loss.device)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Qwen3VLMoeCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=outputs.rope_deltas,
+        aux_loss=aux_loss,
+    )

liger_kernel/transformers/monkey_patch.py

@@ -30,10 +30,14 @@ from liger_kernel.transformers.model.mixtral import lce_forward_deprecated as mi
 from liger_kernel.transformers.model.phi3 import lce_forward as phi3_lce_forward
 from liger_kernel.transformers.model.qwen2 import lce_forward as qwen2_lce_forward
 from liger_kernel.transformers.model.qwen2 import lce_forward_deprecated as qwen2_lce_forward_deprecated
+from liger_kernel.transformers.model.qwen3_vl import lce_forward as qwen3_vl_lce_forward
+from liger_kernel.transformers.model.qwen3_vl_moe import lce_forward as qwen3_vl_moe_lce_forward
 from liger_kernel.transformers.model.smollm3 import lce_forward as smollm3_lce_forward
 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.swiglu import LigerBlockSparseTop2MLP
 from liger_kernel.transformers.swiglu import LigerPhi3SwiGLUMLP
 from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
@@ -1643,6 +1647,154 @@ def apply_liger_kernel_to_qwen2_5_vl(
                     _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_qwen3_vl(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = False,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Qwen3-VL models.
+
+    Args:
+        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.
+        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.qwen3_vl import modeling_qwen3_vl
+    from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLForConditionalGeneration
+    from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLModel
+    from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLTextModel
+
+    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
+
+    if rms_norm:
+        modeling_qwen3_vl.Qwen3VLTextRMSNorm = 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(qwen3_vl_lce_forward, model)
+        else:
+            modeling_qwen3_vl.Qwen3VLForConditionalGeneration.forward = qwen3_vl_lce_forward
+
+    if model is not None and rms_norm:
+        if isinstance(model, (Qwen3VLForConditionalGeneration, Qwen3VLModel)):
+            text_model: Qwen3VLTextModel = model.language_model
+        elif isinstance(model, Qwen3VLTextModel):
+            text_model = model
+        else:
+            raise TypeError(
+                f"Unsupported Qwen3VL model type. `model` must be `Qwen3VLForConditionalGeneration`, `Qwen3VLModel` or `Qwen3VLTextModel`. Got: {type(model)}"
+            )
+
+        _patch_qwen3_vl_rms_norm = partial(_patch_rms_norm_module, offset=0.0, casting_mode="llama")
+
+        if text_model is not None:
+            _patch_qwen3_vl_rms_norm(text_model.norm)
+            for decoder_layer in text_model.layers:
+                _patch_qwen3_vl_rms_norm(decoder_layer.input_layernorm)
+                _patch_qwen3_vl_rms_norm(decoder_layer.post_attention_layernorm)
+                self_attn = getattr(decoder_layer, "self_attn", None)
+                if self_attn is not None:
+                    if hasattr(self_attn, "q_norm") and self_attn.q_norm is not None:
+                        _patch_qwen3_vl_rms_norm(self_attn.q_norm)
+                    if hasattr(self_attn, "k_norm") and self_attn.k_norm is not None:
+                        _patch_qwen3_vl_rms_norm(self_attn.k_norm)
+
+
+def apply_liger_kernel_to_qwen3_vl_moe(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = False,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Qwen3-VL MoE models.
+
+    Args:
+        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 False.
+        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is True.
+        swiglu (bool): Whether to apply Liger's SwiGLU MLP. Default is False.
+        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.qwen3_vl_moe import modeling_qwen3_vl_moe
+    from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import Qwen3VLMoeForConditionalGeneration
+    from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import Qwen3VLMoeModel
+    from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import Qwen3VLMoeTextModel
+
+    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
+
+    if rms_norm:
+        modeling_qwen3_vl_moe.Qwen3VLMoeTextRMSNorm = 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(qwen3_vl_moe_lce_forward, model)
+        else:
+            modeling_qwen3_vl_moe.Qwen3VLMoeForConditionalGeneration.forward = qwen3_vl_moe_lce_forward
+
+    if model is not None and rms_norm:
+        if isinstance(model, (Qwen3VLMoeForConditionalGeneration, Qwen3VLMoeModel)):
+            text_model: Qwen3VLMoeTextModel = model.language_model
+        elif isinstance(model, Qwen3VLMoeTextModel):
+            text_model = model
+        else:
+            raise TypeError(
+                f"Unsupported Qwen3VLMoe model type. `model` must be `Qwen3VLMoeForConditionalGeneration`, `Qwen3VLMoeModel` or `Qwen3VLMoeTextModel`. Got: {type(model)}"
+            )
+
+        _patch_qwen3_vl_moe_rms_norm = partial(_patch_rms_norm_module, offset=0.0, casting_mode="llama")
+
+        if text_model is not None:
+            _patch_qwen3_vl_moe_rms_norm(text_model.norm)
+            for decoder_layer in text_model.layers:
+                _patch_qwen3_vl_moe_rms_norm(decoder_layer.input_layernorm)
+                _patch_qwen3_vl_moe_rms_norm(decoder_layer.post_attention_layernorm)
+                self_attn = getattr(decoder_layer, "self_attn", None)
+                if self_attn is not None:
+                    if hasattr(self_attn, "q_norm") and self_attn.q_norm is not None:
+                        _patch_qwen3_vl_moe_rms_norm(self_attn.q_norm)
+                    if hasattr(self_attn, "k_norm") and self_attn.k_norm is not None:
+                        _patch_qwen3_vl_moe_rms_norm(self_attn.k_norm)
+
+
 def apply_liger_kernel_to_phi3(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -2432,6 +2584,10 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "qwen2_5_vl": apply_liger_kernel_to_qwen2_5_vl,
     "qwen2_5_vl_text": apply_liger_kernel_to_qwen2_5_vl,
     "qwen3_next": apply_liger_kernel_to_qwen3_next,
+    "qwen3_vl": apply_liger_kernel_to_qwen3_vl,
+    "qwen3_vl_text": apply_liger_kernel_to_qwen3_vl,
+    "qwen3_vl_moe": apply_liger_kernel_to_qwen3_vl_moe,
+    "qwen3_vl_moe_text": apply_liger_kernel_to_qwen3_vl_moe,
     "smollm3": apply_liger_kernel_to_smollm3,
     "phi3": apply_liger_kernel_to_phi3,
     "paligemma": apply_liger_kernel_to_paligemma,

liger_kernel/transformers/rope.py

@@ -1,3 +1,8 @@
+from typing import Optional
+from typing import Tuple
+
+import torch
+
 from liger_kernel.ops.rope import LigerRopeFunction
 
 
@@ -18,3 +23,41 @@ def liger_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     """
 
     return LigerRopeFunction.apply(q, k, cos, sin, position_ids, unsqueeze_dim)
+
+
+def liger_rotary_pos_emb_with_cast(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    orig_q_dtype, orig_k_dtype = q.dtype, k.dtype
+
+    q32 = q.to(torch.float32)
+    k32 = k.to(torch.float32)
+    cos32 = cos.to(torch.float32)
+    sin32 = sin.to(torch.float32)
+
+    q_out, k_out = liger_rotary_pos_emb(q32, k32, cos32, sin32, position_ids=position_ids, unsqueeze_dim=unsqueeze_dim)
+    return q_out.to(orig_q_dtype), k_out.to(orig_k_dtype)
+
+
+def liger_rotary_pos_emb_with_cast_and_leading_batch(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    orig_q_dtype, orig_k_dtype = q.dtype, k.dtype
+
+    q32 = q.to(torch.float32).unsqueeze(0)
+    k32 = k.to(torch.float32).unsqueeze(0)
+    cos32 = cos.to(torch.float32).unsqueeze(0)
+    sin32 = sin.to(torch.float32).unsqueeze(0)
+
+    q_out, k_out = liger_rotary_pos_emb(q32, k32, cos32, sin32, position_ids=position_ids, unsqueeze_dim=unsqueeze_dim)
+    return q_out.to(orig_q_dtype).squeeze(0), k_out.to(orig_k_dtype).squeeze(0)

{liger_kernel_nightly-0.6.3.dev20251031170118.dist-info → liger_kernel_nightly-0.6.3.dev20251105012545.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.3.dev20251031170118
+Version: 0.6.3.dev20251105012545
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.3.dev20251031170118.dist-info → liger_kernel_nightly-0.6.3.dev20251105012545.dist-info}/RECORD

@@ -42,7 +42,7 @@ 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=MAAd-YqPdG-j_sbrIE43nrICpA4xTg-dx6M06KWLMFU,9486
+liger_kernel/transformers/__init__.py,sha256=iV1X0gH1JXwgeb7AeY8Ryv7q3r44MLQvSvn79yIVDzw,9874
 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
@@ -59,12 +59,12 @@ liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCc
 liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
 liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
 liger_kernel/transformers/llama4_rope.py,sha256=kS6PSHEwf3dS7hD7C7p8S0geugx2EMCiP0h0F7LsUoY,3639
-liger_kernel/transformers/monkey_patch.py,sha256=3DLFMn2VusVcR6C5YElfpHJBRoJxvho0a2JoVdGqxHA,117266
+liger_kernel/transformers/monkey_patch.py,sha256=O_kl0l56oHinVv-bwl1LU5nKPm6nA0YBjKTYmmwgRbk,124732
 liger_kernel/transformers/multi_token_attention.py,sha256=K3NIY9_5TPgZ4_Rahn0xnkMXxD_fmlJHK4CWGYvGQp0,1752
 liger_kernel/transformers/poly_norm.py,sha256=g5tC75i3qy1_N26ZUP-jfpct7ivQAEdJfIfx8IXzeyE,1377
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=HwddVqrqS58jE-M2_4NkFGARtCDBhGnkKyjBN9b3FYI,3004
-liger_kernel/transformers/rope.py,sha256=ZTrTORSAyfcFIKjk6XEeYmk4ROH7xXED9L4g2NFntlE,999
+liger_kernel/transformers/rope.py,sha256=VMlDZI6zss9mLaLcN5XCE_ktmYRwAi_Eh4TIgO6NrIQ,2361
 liger_kernel/transformers/softmax.py,sha256=yadlAgE4V2JByMwrDDa2s5SUBp8Jgd57xwnVvAWoBaI,264
 liger_kernel/transformers/sparsemax.py,sha256=0lQA0UEOs4mu8CMruZ3VLhImxQVXJWhPsAKUsYA7vj8,403
 liger_kernel/transformers/swiglu.py,sha256=LZ8YeLIdv2k46JleZMjzubGk98smt6t780kSgcVLsQk,3454
@@ -97,15 +97,17 @@ liger_kernel/transformers/model/qwen2_vl.py,sha256=ZeasFPGs-bxm2Y_E15mo0YNx5wwtK
 liger_kernel/transformers/model/qwen3.py,sha256=Q2aOg5erPrgVgRcqJm8sefLSDtvU1AD5B7aJnP7mRMM,4956
 liger_kernel/transformers/model/qwen3_moe.py,sha256=1CwTMCNFDYsjGoa_aHFBagtC5HuJTV-s0__5UvcjD3A,5686
 liger_kernel/transformers/model/qwen3_next.py,sha256=7To7azriAogxeE7oEvByKztH9154dnDiDVNHHm7PZK4,5632
+liger_kernel/transformers/model/qwen3_vl.py,sha256=YU76HJ0A9kG5CUaZM4i9Bzci4eeXcNl_VSC2tsPWA3k,6301
+liger_kernel/transformers/model/qwen3_vl_moe.py,sha256=ykNIvGBtmcTkn236lhmJHzU1IHVR1Kq1YYYlJ5ynhw4,4445
 liger_kernel/transformers/model/smollm3.py,sha256=0KWVkDtXbjsBKhJnaquV6vUUYyLtfmNwYH0sxJt-qTk,7667
 liger_kernel/transformers/model/smolvlm.py,sha256=yFpPKawLVo3zXzLjM7Y_T8FyRrPxVyp-YPFMM8m3k0c,6734
 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_nightly-0.6.3.dev20251031170118.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.6.3.dev20251031170118.dist-info/METADATA,sha256=tIRv5lazhwtKsdhSattKCeY8GFJaJgIXFrPQXIXNd6E,24777
-liger_kernel_nightly-0.6.3.dev20251031170118.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.6.3.dev20251031170118.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.6.3.dev20251031170118.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.6.3.dev20251031170118.dist-info/RECORD,,
+liger_kernel_nightly-0.6.3.dev20251105012545.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.6.3.dev20251105012545.dist-info/METADATA,sha256=MKC5NuGeIkIrDXRVDM3wv-p0cyVbwya5NujVcmSz-mQ,24777
+liger_kernel_nightly-0.6.3.dev20251105012545.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.6.3.dev20251105012545.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.6.3.dev20251105012545.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.6.3.dev20251105012545.dist-info/RECORD,,