liger-kernel 0.5.7__py3-none-any.whl → 0.5.9__py3-none-any.whl

liger_kernel/chunked_loss/dpo_loss.py

@@ -68,6 +68,7 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
         compute_nll_loss=False,
         compiled=True,
         use_ref_model=True,
+        average_log_prob=False,
         chunk_size=1,
     ):
         """
@@ -85,6 +86,7 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
             compute_nll_loss (bool): Whether to compute the NLL loss
             compiled (bool): Whether to use torch compile
             use_ref_model (bool): Whether to use a reference model
+            average_log_prob (bool): Whether to average the log probability per non-masked token
             chunk_size (int): Size of chunks for processing.
         Returns:
             torch.Tensor: Computed loss
@@ -104,13 +106,14 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
             ref_input=ref_input,
             ref_weight=ref_weight,
             ref_bias=ref_bias,
+            average_log_prob=average_log_prob,
             chunk_size=chunk_size,
         )
 
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None, None, None, None, None, None
+        return *grads, None, None, None, None, None, None, None, None, None, None
 
 
 class LigerFusedLinearDPOLoss(torch.nn.Module):
@@ -125,6 +128,7 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
         compute_nll_loss: bool = False,
         compiled: bool = True,
         use_ref_model: bool = True,
+        average_log_prob: bool = True,
         chunk_size: int = 1,
     ):
         """
@@ -134,6 +138,7 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
             compute_nll_loss (bool): Whether to compute the NLL loss.
             compiled (bool): Whether to use the torch compiled kernel.
             use_ref_model (bool): Whether to use a reference model for the DPO loss.
+            average_log_prob (bool): Whether to average the log probability per non-masked token.
             chunk_size (int): Size of chunks for processing.
         """
         super().__init__()
@@ -142,6 +147,7 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
         self.compute_nll_loss = compute_nll_loss
         self.compiled = compiled
         self.use_ref_model = use_ref_model
+        self.average_log_prob = average_log_prob
         self.chunk_size = chunk_size
 
     def forward(
@@ -167,5 +173,6 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
             self.compute_nll_loss,
             self.compiled,
             self.use_ref_model,
+            self.average_log_prob,
             self.chunk_size,
         )

liger_kernel/chunked_loss/grpo_loss.py

@@ -1,3 +1,5 @@
+from typing import Optional
+
 import torch
 
 from liger_kernel.chunked_loss.fused_linear_ppo import LigerFusedLinearPPOBase
@@ -204,7 +206,7 @@ class LigerFusedLinearGRPOLoss(torch.nn.Module):
         epsilon_low: float = 0.2,
         epsilon_high: float = 0.2,
         loss_type: str = "bnpo",
-        max_completion_length: int | None = None,
+        max_completion_length: Optional[int] = None,
         temperature: float = 1.0,
     ):
         """

liger_kernel/ops/cross_entropy.py

@@ -351,7 +351,10 @@ def cross_entropy_backward(_input, grad_output):
     # If cross entropy is the last layer, grad_output is 1.0. Skip the mul to save time
     if torch.equal(grad_output, torch.tensor(1.0, device=grad_output.device)):
         pass
-
+    # If reduction is 'none'
+    elif grad_output.ndim > 0:
+        _input = _input * grad_output.unsqueeze(dim=1)
+    # If reduction is ['mean', 'sum'], grad_output is just a scalar
     # We use a Triton kernel instead of a PyTorch operation because modifying inputs in-place
     # for gradient storage and backward multiple times causes anomalies with PyTorch but not with Triton.
     else:

liger_kernel/ops/fused_linear_cross_entropy.py

@@ -143,9 +143,10 @@ def fused_linear_cross_entropy_forward(
                 alpha=1.0,
             )
 
-    if reduction == "none":
-        loss = loss_1d
-        z_loss = z_loss_1d if return_z_loss else None
+    # Need extra calculations for backward if reduction=='none'. Not supporting reduction='none' now.
+    # if reduction == "none":
+    #     loss = loss_1d
+    #     z_loss = z_loss_1d if return_z_loss else None
 
     else:
         loss = torch.sum(loss_1d)

liger_kernel/transformers/__init__.py

@@ -26,6 +26,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma2  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma3_text  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_granite  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llama  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llava  # noqa: F401
@@ -38,6 +39,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_5_vl  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_vl  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3  # noqa: F401
 
 
 # Check if 'transformers' is installed
@@ -79,6 +81,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_gemma2",
         "apply_liger_kernel_to_gemma3",
         "apply_liger_kernel_to_gemma3_text",
+        "apply_liger_kernel_to_glm4",
         "apply_liger_kernel_to_granite",
         "apply_liger_kernel_to_llama",
         "apply_liger_kernel_to_llava",
@@ -91,6 +94,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen2",
         "apply_liger_kernel_to_qwen2_5_vl",
         "apply_liger_kernel_to_qwen2_vl",
+        "apply_liger_kernel_to_qwen3",
     }
 
     if name in monkey_patch_symbols:
@@ -129,6 +133,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_gemma2",
             "apply_liger_kernel_to_gemma3",
             "apply_liger_kernel_to_gemma3_text",
+            "apply_liger_kernel_to_glm4",
             "apply_liger_kernel_to_granite",
             "apply_liger_kernel_to_llama",
             "apply_liger_kernel_to_llava",
@@ -141,5 +146,6 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_qwen2",
             "apply_liger_kernel_to_qwen2_5_vl",
             "apply_liger_kernel_to_qwen2_vl",
+            "apply_liger_kernel_to_qwen3",
         ]
     )

liger_kernel/transformers/fused_linear_cross_entropy.py

@@ -23,8 +23,7 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         assert reduction in {
             "mean",
             "sum",
-            "none",
-        }, f"reduction must be one of 'mean', 'sum', or 'none'. Got: {reduction}"
+        }, f"reduction must be 'mean' or 'sum'. Got: {reduction}"
         assert softcap is None or softcap > 0, f"softcap must greater than 0.0 or None. Got: {softcap}"
         self.ce_weight = ce_weight
         self.ignore_index = ignore_index

liger_kernel/transformers/model/gemma.py

@@ -200,21 +200,25 @@ def lce_forward(
     )
 
     hidden_states = outputs[0]
+    # 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, don't materialize logits
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            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 mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

liger_kernel/transformers/model/gemma2.py

@@ -212,23 +212,27 @@ def lce_forward(
     )
 
     hidden_states = outputs[0]
+    # 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, don't materialize logits
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
+            shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
             final_logit_softcapping=self.config.final_logit_softcapping,
             **loss_kwargs,
         )
 
     else:  # if in inference mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if self.config.final_logit_softcapping is not None:
             logits = logits / self.config.final_logit_softcapping
             logits = torch.tanh(logits)

liger_kernel/transformers/model/gemma3.py

@@ -104,13 +104,15 @@ def causal_forward(
     # 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)
     loss = None
     logits = None
-    if self.training and (labels is not None):
+    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,
             final_logit_softcapping=self.config.final_logit_softcapping,
             **loss_kwargs,

liger_kernel/transformers/model/glm4.py

@@ -0,0 +1,125 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.glm4.modeling_glm4 import _CONFIG_FOR_DOC
+from transformers.models.glm4.modeling_glm4 import GLM4_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(GLM4_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+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,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **loss_kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+    Args:
+        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, Glm4ForCausalLM
+
+    >>> model = Glm4ForCausalLM.from_pretrained("THUDM/GLM-4-9B-0414")
+    >>> tokenizer = AutoTokenizer.from_pretrained("THUDM/GLM-4-9B-0414")
+
+    >>> 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 sure if you’re conscious of this, but I’m'
+    ```
+    """
+    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
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = 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,
+        return_dict=return_dict,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs[0]
+    # 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, don't 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 mode materialize logits
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **loss_kwargs,
+            )
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )

liger_kernel/transformers/model/llama.py

@@ -209,25 +209,29 @@ def lce_forward(
     )
 
     hidden_states = outputs[0]
+    # 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, :]
 
     if self.config.pretraining_tp > 1:
         raise Exception("Liger Kernel does not support pretraining_tp!!")
 
+    shift_labels = loss_kwargs.pop("shift_labels", None)
     logits = None
     loss = None
     # if in training mode, don't materialize logits
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            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 mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

liger_kernel/transformers/model/mistral.py

@@ -91,22 +91,26 @@ def lce_forward(
     )
 
     hidden_states = outputs[0]
+    # 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)
     loss = None
     logits = None
 
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            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:
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
 
         loss = None
         if labels is not None:

liger_kernel/transformers/model/mixtral.py

@@ -225,22 +225,26 @@ def lce_forward(
     )
 
     hidden_states = outputs[0]
+    # 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, don't materialize logits
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            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 mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
 
         loss = None
         if labels is not None:

liger_kernel/transformers/model/mllama.py

@@ -215,22 +215,26 @@ def lce_forward(
     )
 
     hidden_states = outputs[0]
+    # 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, don't materialize logits
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            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 mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

liger_kernel/transformers/model/olmo2.py

@@ -88,22 +88,26 @@ def lce_forward(
     )
 
     hidden_states = outputs[0]
+    # 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, don't materialize logits
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            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 mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

liger_kernel/transformers/model/phi3.py

@@ -213,22 +213,26 @@ def lce_forward(
     )
 
     hidden_states = outputs[0]
+    # 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, don't materialize logits
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            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 mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

liger_kernel/transformers/model/qwen2.py

@@ -199,22 +199,26 @@ def lce_forward(
     )
 
     hidden_states = outputs[0]
+    # 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, don't materialize logits
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            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 mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

liger_kernel/transformers/model/qwen2_5_vl.py

@@ -163,14 +163,16 @@ def lce_forward(
 
     hidden_states = outputs[0]
 
+    shift_labels = loss_kwargs.pop("shift_labels", None)
     loss = None
     logits = None
 
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
             hidden_states=hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
+            shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
             **loss_kwargs,
         )

liger_kernel/transformers/model/qwen2_vl.py

@@ -167,14 +167,16 @@ def lce_forward(
 
     hidden_states = outputs[0]
 
+    shift_labels = loss_kwargs.pop("shift_labels", None)
     loss = None
     logits = None
 
-    if self.training and (labels is not None):
+    if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
             hidden_states=hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
+            shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
             **loss_kwargs,
         )

liger_kernel/transformers/model/qwen3.py

@@ -0,0 +1,118 @@
+from typing import List
+from typing import Optional
+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,
+    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,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **kwargs,
+) -> CausalLMOutputWithPast:
+    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, Qwen3ForCausalLM
+
+    >>> model = Qwen3ForCausalLM.from_pretrained("Qwen/Qwen3-8B")
+    >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
+
+    >>> 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_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 = 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,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    # 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
+    # if in training mode, don't 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,
+            **kwargs,
+        )
+
+    else:  # if in inference mode materialize logits
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )

liger_kernel/transformers/monkey_patch.py

@@ -1048,6 +1048,60 @@ 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_qwen2_vl(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -1319,12 +1373,78 @@ 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 +1454,7 @@ 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,
     "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-0.5.7.dist-info → liger_kernel-0.5.9.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: liger_kernel
-Version: 0.5.7
+Version: 0.5.9
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -320,9 +320,11 @@ 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       |
 | 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

{liger_kernel-0.5.7.dist-info → liger_kernel-0.5.9.dist-info}/RECORD

@@ -4,21 +4,21 @@ liger_kernel/utils.py,sha256=178Hn8uD-VauDT6FjqMyXLbKLod8ObIpaTtapHwfEK0,1861
 liger_kernel/chunked_loss/README.md,sha256=0FmkFC3hKBqyoDT5uTlIYmrvRkF-EOCR1y-EBU1LpWU,2248
 liger_kernel/chunked_loss/__init__.py,sha256=ATu-xX5Fc49Cr6yBOGBRNTo593ZrU5ZCsIuvoIbJWw4,603
 liger_kernel/chunked_loss/cpo_loss.py,sha256=Gzz1eU4kgcbdubFVRy55e8A1Cr-r45UgNicXwZIjmBU,5454
-liger_kernel/chunked_loss/dpo_loss.py,sha256=xZwGqS04si9zXyob95SAdalC-hajZg8fWINqiqffN8k,5855
+liger_kernel/chunked_loss/dpo_loss.py,sha256=Xypt4FoTSmAnJE4SWtsCv4aNHK4ToR1LonUQtCTEuHQ,6258
 liger_kernel/chunked_loss/functional.py,sha256=9G3nKm-Bi7uoZRFkL8wwGMl6juDl4bSzDvTa5GHZPzg,955
 liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=ooR-qnZCyWJN935oHCSWLaKKKyaYERyhNczRGi1VOiw,11935
 liger_kernel/chunked_loss/fused_linear_ppo.py,sha256=AA19cpv6D8mo5RbSK5GRCcZoOSnpxV_Z1eJlAsC5eic,13434
 liger_kernel/chunked_loss/fused_linear_preference.py,sha256=ojB42jYPu0c4ki96Ft-hy7Sf6fh_WikG-aWNrlZzSio,18362
 liger_kernel/chunked_loss/fused_linear_unpaired_preference.py,sha256=RiuK3UtRwH9T6jZ36sA8Urj-TVuOLOO2syLg_JOQapY,13437
-liger_kernel/chunked_loss/grpo_loss.py,sha256=eh6mErFUZsSQrgRRefuXdk-LG0gS7Rg2r-U9CtbH3eU,10834
+liger_kernel/chunked_loss/grpo_loss.py,sha256=kuqHkYV383sUxqJN-DMsfADHi2hxHVyKx5S24TNc8bQ,10866
 liger_kernel/chunked_loss/jsd_loss.py,sha256=u2ahkuHsbhpNaKcpBCz5gCMDk9ou-P04DHji592dIBo,7067
 liger_kernel/chunked_loss/kto_loss.py,sha256=llVCe6DkcpCo57seGWoMikaQVFApx764jsmSbQyqwQY,7529
 liger_kernel/chunked_loss/orpo_loss.py,sha256=nu9UYG16dcMw93lvHi4_hYs3Q0FK1KnlmMRj7OpYU8s,4872
 liger_kernel/chunked_loss/simpo_loss.py,sha256=fy2w8KbhMrBv7b1jdIeH3bBFxY52bPQPZb3KwBvmurM,5385
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/ops/cross_entropy.py,sha256=T5oSsqOS1y-Iea5o9v_BSU-_mIEXqWAT1oX_m59NcA4,18941
+liger_kernel/ops/cross_entropy.py,sha256=e8THGnhOcy_0SbOLABx67HEM7-B8a8pG7nDKbCRpQKM,19123
 liger_kernel/ops/dyt.py,sha256=YD1-buHz9VmIX838VKzLc-lm5CeUQ4LAskGDWBUMQHA,6187
-liger_kernel/ops/fused_linear_cross_entropy.py,sha256=1Y3Uk_TCSjqKgoG2eot1ptnWXJXXQESqGvOmqAW1gsM,10912
+liger_kernel/ops/fused_linear_cross_entropy.py,sha256=5fbGhN85n3zf0uIdJ7PYHWIRzTf0VTFiS0ARtOmqIP0,11020
 liger_kernel/ops/fused_linear_jsd.py,sha256=CSoprxb-YcJy-YUKiTcYkxN8sb9h2kdk_iHuncvSV5c,9683
 liger_kernel/ops/geglu.py,sha256=axGvCIvlBzuluoAIrWTsp2iZM4BFKNInkPov8YVvH9E,4126
 liger_kernel/ops/group_norm.py,sha256=qD4D4lSjSgVtO52EBNLC2iTseALRgPgqXE50U2woggk,10837
@@ -33,12 +33,12 @@ 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=SH30Pt2ZqyQY-mmWQldg_r-5koowuymTIoU4F4e1KHk,6419
+liger_kernel/transformers/__init__.py,sha256=x_3CYHJt-xj4va3N32kfwf000F-DNBtj-YE6OylDAW8,6774
 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=QMqqc14pkE0WhpRZvapfnNAun-6C0C_tHExL2ZJuCUA,648
 liger_kernel/transformers/functional.py,sha256=4h9Pdx_iINBqfv2Zod_c27qOpYXDDwbdVgatQ9_XBmI,5089
-liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=09Rt7FZzLH42VOcIbQ4dlQd0o3Rlb4vk6fqiOQ7WTD8,1778
+liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=O8Sg5BT81nTaY9fSGoOY9dOD9ekibwwiuXhdUHaxntQ,1742
 liger_kernel/transformers/fused_linear_jsd.py,sha256=bZ4otCvWBuOnA5XdQL-FzZVItJlDt-ht9e_pG7PG93E,3999
 liger_kernel/transformers/geglu.py,sha256=mrgqzIUVd6lN7fkDKLkw5YaESDxDtFgbot430WwPVOQ,1107
 liger_kernel/transformers/gema3_rms.py,sha256=LTmZOXe6WEnv6ZroW-kU1TE2B36-z5v8OLmKr3XEVFo,353
@@ -46,7 +46,7 @@ liger_kernel/transformers/group_norm.py,sha256=6qMAWOprr4SzP0YhNVNGQIBpM5aUHplUD
 liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
 liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
 liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
-liger_kernel/transformers/monkey_patch.py,sha256=QpfNU7MmVDGlBWIZ2RLTSyh0vuZ-si7H37SL-qOliUs,64393
+liger_kernel/transformers/monkey_patch.py,sha256=8Q84xxWA7ltgqgGRBxKxPPNeG7k5HYQfgaw1-HFnKGM,69287
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=GqCEJuGt0YdqqlMcToE0Wp4A8YFquDa4UUSyH2uFW2A,1191
 liger_kernel/transformers/rope.py,sha256=ZTrTORSAyfcFIKjk6XEeYmk4ROH7xXED9L4g2NFntlE,999
@@ -55,28 +55,30 @@ liger_kernel/transformers/trainer_integration.py,sha256=W3ON51O5GkyzNJsItz0y5rKx
 liger_kernel/transformers/tvd.py,sha256=XrRfyJIqN6HFxXk8MYyFVZM1OLz3mtSbRZvWfZ_JerQ,450
 liger_kernel/transformers/experimental/embedding.py,sha256=2P0QYdlFyFrG5OqTzTa1wcRgDSyjBMv5i1a7BrDPDQw,881
 liger_kernel/transformers/model/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/transformers/model/gemma.py,sha256=-JoHKWjtYPpxHQa6QbCwnzX_cctRZG2ZTsaUv-dmOt4,9816
-liger_kernel/transformers/model/gemma2.py,sha256=n4MZupFGDMvtnvkvkNhRrxXS3ZF341BVfyLjrOXp10g,10923
-liger_kernel/transformers/model/gemma3.py,sha256=ge3JYchiKvX1G1Zp00jX2zmQK2K7ymJoZAxbb2ggslw,16102
-liger_kernel/transformers/model/llama.py,sha256=UVXQLRW7rCU5vPab54dLNS3ER37eM446peHX00Yz6eA,10493
+liger_kernel/transformers/model/gemma.py,sha256=nMUY2Iw7j6a-fOUqYBlfzIPznpKPKVa2DMBIZqCVfuI,10087
+liger_kernel/transformers/model/gemma2.py,sha256=eulrUbh1DEMpMR6Lupx69kL-FeuRDP19mVoW1gc7keY,11194
+liger_kernel/transformers/model/gemma3.py,sha256=wGSNqaLRRgIGQ_r9esyhDezm2SkAGZflopoWoWR-nYY,16226
+liger_kernel/transformers/model/glm4.py,sha256=rtyMTtzgh_ncZ7DsfNxRJoUUm7xlDMKGzNqlxXjdAJk,5452
+liger_kernel/transformers/model/llama.py,sha256=F8cvDAlf4NeKESdGEFXs8m3ue2F8i0h3aV2LricMqoM,10764
 liger_kernel/transformers/model/llava.py,sha256=b0pEagjUbu2-eS9xegjyfl1DwIXLwZcNpff55ibaMbA,17601
 liger_kernel/transformers/model/loss_utils.py,sha256=WWAMdiONPaXpIvxyOim_0igLrYh0yyOok5Q9_L9xvZw,1787
-liger_kernel/transformers/model/mistral.py,sha256=RacuKcckuDK6oSraCGD0R0bm-fE0K3q-lkYaAC56C2E,5481
-liger_kernel/transformers/model/mixtral.py,sha256=gLcqGabdv1XnuciS9b-TpkTDnGL8K32Hoq9j2vZMBRY,11502
-liger_kernel/transformers/model/mllama.py,sha256=75mxtmMsNd_q8KlKeawj2uMP6v2KjDuUi4nsUKM5jqA,11308
-liger_kernel/transformers/model/olmo2.py,sha256=rSzSALikEGkk0w3PLNQPrqg-ioN8TpWCXkAlg3LtCdI,5189
+liger_kernel/transformers/model/mistral.py,sha256=1AcwJT9WOIpHkpu4Njs35ZryiGyW8ygERYmGqLz2Z4o,5752
+liger_kernel/transformers/model/mixtral.py,sha256=URMzPLU1akf1H4hHXalCyfbVGUldRx8_jqdrZfM7Y-w,11773
+liger_kernel/transformers/model/mllama.py,sha256=v_ayi6m4sC6AVKTrrLHF4W5HVaL86AYQNBqdWuTTOTw,11579
+liger_kernel/transformers/model/olmo2.py,sha256=Kb6sGPsQS970GsYmWoT0DC2DFiXQ9Yjyxr8FRnT_8tQ,5460
 liger_kernel/transformers/model/paligemma.py,sha256=GNReT6tVZt3ON6aaa9ovg8mnu1hYocSx9OhgC7b-_28,19191
-liger_kernel/transformers/model/phi3.py,sha256=ebITCrmwmb4z66CbSrZl1kD6BsP52IcSAR8uwUTp9nc,10455
-liger_kernel/transformers/model/qwen2.py,sha256=QaoTDrJv2wIuAM8QMoeWVvgNl0N5gHzIrew9QGG7kXc,9744
-liger_kernel/transformers/model/qwen2_5_vl.py,sha256=70BnHZjx6eQWTwi3zc5SMwxTeOOA4Tbdkfy6IYRcTaM,9289
-liger_kernel/transformers/model/qwen2_vl.py,sha256=zo4O9fShNHYqSLrzLGqQYWSMtJI6UHaSY7zvMCYWyD8,9685
+liger_kernel/transformers/model/phi3.py,sha256=TSeHK8H0mnS2esJaZI3lxmo5X3-Uwtd_TsrgvJRkm3s,10726
+liger_kernel/transformers/model/qwen2.py,sha256=bEusb6vrVbagtSUHyntpi9j0x79IrZ1NP8iA5GR5Ryw,10015
+liger_kernel/transformers/model/qwen2_5_vl.py,sha256=oACIsTpg9_GdoSvekCyXLhJkuCpQEiFOTzKj7cjgi2E,9413
+liger_kernel/transformers/model/qwen2_vl.py,sha256=F6DeQ65wPtcpeQJZ9a3SJZKkQ-e24SRLdYUgC-_jT-k,9809
+liger_kernel/transformers/model/qwen3.py,sha256=JdIeh0fvDLdGs8nk4_eHrovHCNa09VG15D4aa0X0mwI,5084
 liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7HHWHwku25A-GYL0WU,193
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=pdekW7l6Qg_aqa5SYKYlSWUF8m3lkOFvFLcIMEHrz9s,8338
 liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
 liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
-liger_kernel-0.5.7.dist-info/licenses/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel-0.5.7.dist-info/licenses/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel-0.5.7.dist-info/METADATA,sha256=DJsNdftnn0V8KyBJyssCU0NnPtOvl3J7ldrKWs_aPRs,23340
-liger_kernel-0.5.7.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
-liger_kernel-0.5.7.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel-0.5.7.dist-info/RECORD,,
+liger_kernel-0.5.9.dist-info/licenses/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel-0.5.9.dist-info/licenses/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel-0.5.9.dist-info/METADATA,sha256=Wq3nqeBFdqmOj8uiy7S4ZEL4xA88DVb0ad2b9KDn-qI,23627
+liger_kernel-0.5.9.dist-info/WHEEL,sha256=0CuiUZ_p9E4cD6NyLD6UG80LBXYyiSYZOKDm5lp32xk,91
+liger_kernel-0.5.9.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel-0.5.9.dist-info/RECORD,,

{liger_kernel-0.5.7.dist-info → liger_kernel-0.5.9.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (78.1.0)
+Generator: setuptools (80.3.1)
 Root-Is-Purelib: true
 Tag: py3-none-any