liger-kernel-nightly 0.5.10.dev20250605210201__py3-none-any.whl → 0.5.10.dev20250605223455__py3-none-any.whl

liger_kernel/transformers/model/gemma.py

@@ -138,7 +138,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Args:
@@ -190,6 +190,7 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -197,7 +198,7 @@ def lce_forward(
     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)
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
 
@@ -215,7 +216,7 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
     else:
         logits = self.lm_head(kept_hidden_states)
@@ -224,7 +225,7 @@ def lce_forward(
                 logits=logits,
                 labels=labels,
                 vocab_size=self.config.vocab_size,
-                **loss_kwargs,
+                **kwargs,
             )
 
     if not return_dict:

liger_kernel/transformers/model/gemma2.py

@@ -30,6 +30,7 @@ def lce_forward_deprecated(
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     cache_position: Optional[torch.LongTensor] = None,
+    **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Args:
@@ -76,6 +77,7 @@ def lce_forward_deprecated(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -147,7 +149,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Args:
@@ -204,6 +206,7 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -211,7 +214,7 @@ def lce_forward(
     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)
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
 
@@ -230,7 +233,7 @@ def lce_forward(
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
             final_logit_softcapping=self.config.final_logit_softcapping,
-            **loss_kwargs,
+            **kwargs,
         )
 
     else:
@@ -242,7 +245,7 @@ def lce_forward(
 
         loss = None
         if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
 
     if not return_dict:
         output = (logits,) + outputs[1:]

liger_kernel/transformers/model/glm4.py

@@ -27,7 +27,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Args:
@@ -80,6 +80,7 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -87,7 +88,7 @@ def lce_forward(
     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)
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
 
@@ -105,7 +106,7 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
 
     else:
@@ -115,7 +116,7 @@ def lce_forward(
                 logits=logits,
                 labels=labels,
                 vocab_size=self.config.vocab_size,
-                **loss_kwargs,
+                **kwargs,
             )
 
     return CausalLMOutputWithPast(

liger_kernel/transformers/model/llama.py

@@ -152,7 +152,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Args:
@@ -205,6 +205,7 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -215,7 +216,7 @@ def lce_forward(
     if self.config.pretraining_tp > 1:
         raise Exception("Liger Kernel does not support pretraining_tp!!")
 
-    shift_labels = loss_kwargs.pop("shift_labels", None)
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
     # if in training mode, don't materialize logits
@@ -233,7 +234,7 @@ def lce_forward(
             hidden_size=self.config.hidden_size,
             labels=labels,
             shift_labels=shift_labels,
-            **loss_kwargs,
+            **kwargs,
         )
 
     else:
@@ -243,7 +244,7 @@ def lce_forward(
                 logits=logits,
                 labels=labels,
                 vocab_size=self.config.vocab_size,
-                **loss_kwargs,
+                **kwargs,
             )
 
     if not return_dict:

liger_kernel/transformers/model/mistral.py

@@ -28,7 +28,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Copy paste Mistral's forward but replace torch cross entropy with liger fused linear cross entropy
@@ -83,6 +83,7 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -90,7 +91,7 @@ def lce_forward(
     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)
+    shift_labels = kwargs.pop("shift_labels", None)
     loss = None
     logits = None
 
@@ -107,7 +108,7 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
 
     else:
@@ -119,7 +120,7 @@ def lce_forward(
                 logits=logits,
                 labels=labels,
                 vocab_size=self.config.vocab_size,
-                **loss_kwargs,
+                **kwargs,
             )
     if not return_dict:
         output = (logits,) + outputs[1:]

liger_kernel/transformers/model/mixtral.py

@@ -157,7 +157,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
     r"""
     Args:
@@ -215,6 +215,7 @@ def lce_forward(
         output_router_logits=output_router_logits,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -222,7 +223,7 @@ def lce_forward(
     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)
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
 
@@ -240,7 +241,7 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
 
     else:
@@ -248,7 +249,7 @@ def lce_forward(
 
         loss = None
         if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
     aux_loss = None
     if output_router_logits:
         aux_loss = load_balancing_loss_func(

liger_kernel/transformers/model/mllama.py

@@ -148,7 +148,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Args:
@@ -206,6 +206,7 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -213,7 +214,7 @@ def lce_forward(
     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)
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
 
@@ -231,7 +232,7 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
 
     else:
@@ -241,7 +242,7 @@ def lce_forward(
                 logits=logits,
                 labels=labels,
                 vocab_size=self.config.vocab_size,
-                **loss_kwargs,
+                **kwargs,
             )
 
     if not return_dict:

liger_kernel/transformers/model/olmo2.py

@@ -27,7 +27,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Args:
@@ -80,6 +80,7 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -87,7 +88,7 @@ def lce_forward(
     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)
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
 
@@ -105,7 +106,7 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
 
     else:
@@ -115,7 +116,7 @@ def lce_forward(
                 logits=logits,
                 labels=labels,
                 vocab_size=self.config.vocab_size,
-                **loss_kwargs,
+                **kwargs,
             )
 
     return CausalLMOutputWithPast(

liger_kernel/transformers/model/phi3.py

@@ -137,7 +137,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Args:
@@ -203,6 +203,7 @@ def lce_forward(
         output_attentions=output_attentions,
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -210,7 +211,7 @@ def lce_forward(
     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)
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
 
@@ -228,7 +229,7 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
 
     else:
@@ -238,7 +239,7 @@ def lce_forward(
                 logits=logits,
                 labels=labels,
                 vocab_size=self.config.vocab_size,
-                **loss_kwargs,
+                **kwargs,
             )
 
     if not return_dict:

liger_kernel/transformers/model/qwen2.py

@@ -136,7 +136,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Args:
@@ -189,6 +189,7 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -196,7 +197,7 @@ def lce_forward(
     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)
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
 
@@ -214,7 +215,7 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
 
     else:
@@ -224,7 +225,7 @@ def lce_forward(
                 logits=logits,
                 labels=labels,
                 vocab_size=self.config.vocab_size,
-                **loss_kwargs,
+                **kwargs,
             )
 
     return CausalLMOutputWithPast(

liger_kernel/transformers/model/qwen2_5_vl.py

@@ -31,7 +31,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     second_per_grid_ts: Optional[torch.Tensor] = None,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, Qwen2_5_VLCausalLMOutputWithPast]:
     r"""
     Copy paste Qwen2_5_VL's forward but replace torch cross entropy with liger fused linear cross entropy
@@ -154,11 +154,12 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
 
-    shift_labels = loss_kwargs.pop("shift_labels", None)
+    shift_labels = kwargs.pop("shift_labels", None)
     loss = None
     logits = None
 
@@ -175,7 +176,7 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
     else:
         logits = self.lm_head(hidden_states)

liger_kernel/transformers/model/qwen2_vl.py

@@ -32,7 +32,7 @@ def lce_forward(
     rope_deltas: Optional[torch.LongTensor] = None,
     cache_position: Optional[torch.LongTensor] = None,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> Union[Tuple, Qwen2VLCausalLMOutputWithPast]:
     r"""
     Copy paste Qwen2VL's forward but replace torch cross entropy with liger fused linear cross entropy
@@ -158,11 +158,12 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs[0]
 
-    shift_labels = loss_kwargs.pop("shift_labels", None)
+    shift_labels = kwargs.pop("shift_labels", None)
     loss = None
     logits = None
 
@@ -179,7 +180,7 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
     else:
         logits = self.lm_head(hidden_states)

liger_kernel/transformers/model/qwen3_moe.py

@@ -26,7 +26,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
-    **loss_kwargs,
+    **kwargs,
 ) -> MoeCausalLMOutputWithPast:
     r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -81,6 +81,7 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         output_router_logits=output_router_logits,
         cache_position=cache_position,
+        **kwargs,
     )
 
     hidden_states = outputs.last_hidden_state
@@ -88,7 +89,7 @@ def lce_forward(
     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)
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
 
@@ -102,12 +103,12 @@ def lce_forward(
             labels=labels,
             shift_labels=shift_labels,
             hidden_size=self.config.hidden_size,
-            **loss_kwargs,
+            **kwargs,
         )
     else:  # if in inference model materialize logits
         logits = self.lm_head(kept_hidden_states)
         if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
 
     aux_loss = None
     if output_router_logits:

{liger_kernel_nightly-0.5.10.dev20250605210201.dist-info → liger_kernel_nightly-0.5.10.dev20250605223455.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.10.dev20250605210201
+Version: 0.5.10.dev20250605223455
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.10.dev20250605210201.dist-info → liger_kernel_nightly-0.5.10.dev20250605223455.dist-info}/RECORD

@@ -63,31 +63,31 @@ 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=LUifPVeiVpadKwAoi0g0lplhaE5OMdx-k4pSg4g_y8A,9732
-liger_kernel/transformers/model/gemma2.py,sha256=JxPTXVkuFtiaZYkaBM8bZF-ObyatHmAiOG_gzRe_ElU,10989
+liger_kernel/transformers/model/gemma.py,sha256=gvP-9zZ1e-DQD06qltWmRhiJClJDtkMQL1UrPMMZZGQ,9730
+liger_kernel/transformers/model/gemma2.py,sha256=ORmzklEAMpk93nToRo4d_ZJbM4ScVE2szczsEL4hw7w,11019
 liger_kernel/transformers/model/gemma3.py,sha256=JI4jj9K660HeRsofB6cpkCHBQ0OsazElArRtKUehUmw,15945
-liger_kernel/transformers/model/glm4.py,sha256=3YJiGdZ0nNSdZidPFlXdUad8mlFwyfq44yd11OcdNns,5259
-liger_kernel/transformers/model/llama.py,sha256=cAWTCY0bk67lFXNtAVEXIWl9WNgn4JyU25Q7nhpKjE0,12505
+liger_kernel/transformers/model/glm4.py,sha256=GlnEhdGJuDIqp2R9qC54biY3HwV1tWmfpJm6ijoAsrM,5257
+liger_kernel/transformers/model/llama.py,sha256=LcIxVfF0PXXWHBVJa6Ody_5fAtIpxQcI4jC_j-o51fU,12503
 liger_kernel/transformers/model/llava.py,sha256=ONdpx96AVbbL8QDQvHSm08jMJPz3tzkbeO92IRbAb1A,19270
 liger_kernel/transformers/model/loss_utils.py,sha256=WWAMdiONPaXpIvxyOim_0igLrYh0yyOok5Q9_L9xvZw,1787
-liger_kernel/transformers/model/mistral.py,sha256=vFFZD5VAwpx6Bs4gXoXDRmyU9-7Dp50w3jIcj0q0sIo,5567
-liger_kernel/transformers/model/mixtral.py,sha256=vSmgBc91WMu9_iWkAHUJPzo0-WDkTJK5SEVYNaDRT_Y,11398
-liger_kernel/transformers/model/mllama.py,sha256=XhxU5r9v5TfTa4NJpg1EeYC999Q1e0CZwkVE86iaymU,11282
-liger_kernel/transformers/model/olmo2.py,sha256=4IwhP-TBck0dijY5gdLvoQnXO2M80gfLUV2fsK6wpiY,5261
+liger_kernel/transformers/model/mistral.py,sha256=okKkyashfFLfhjIT--f3JY6JHOslOtDI8U1dlpBC2Zs,5565
+liger_kernel/transformers/model/mixtral.py,sha256=VY-y73IyjcCyWyI7ahxXLw0fJrhgjYfr1xwRYtsHX0o,11396
+liger_kernel/transformers/model/mllama.py,sha256=my29NXk-p6ckQaP8qDIN8e318yI_9mQZHt38MV3SqLY,11280
+liger_kernel/transformers/model/olmo2.py,sha256=6L_bo-ZUgO1lYppdJneOtYxNIylQKS6BiGp13g7Uq9E,5259
 liger_kernel/transformers/model/paligemma.py,sha256=xuIx3oOwTgftU3jqLfWOxUxgCLBNJh0yNC21an9qDjo,18773
-liger_kernel/transformers/model/phi3.py,sha256=UslJ1gbyRhVmj5fzq_uizhDY5wYEoK_EwPamNotsUVs,10378
-liger_kernel/transformers/model/qwen2.py,sha256=WTKFPAp_R4aSRLQgMKygX6pmptcHeLWGCdjH42SxXVk,9660
-liger_kernel/transformers/model/qwen2_5_vl.py,sha256=OFPaELlVi1UdkjSXxVWGnNc32CdcQ74KV_3Dc8-uCe4,9200
-liger_kernel/transformers/model/qwen2_vl.py,sha256=s973gNrFT74FYAYiRvorxtK15CpZJnlbbhfk_wk-tag,9611
+liger_kernel/transformers/model/phi3.py,sha256=m-MD_OuTaYMGZhHOvl-RHOVEObrL8tL5cBv3VTNd4F0,10376
+liger_kernel/transformers/model/qwen2.py,sha256=SdN7V-MI3eX9s2DAFRvC1g-G146uG_5n1fnNdY9QwYk,9658
+liger_kernel/transformers/model/qwen2_5_vl.py,sha256=k6jt1bTCJsKsZVGhTxqIbDzmnL8-B3CpWJOjLazswbo,9203
+liger_kernel/transformers/model/qwen2_vl.py,sha256=Cgs7-nPlKFifiDO9gqSI6np4vRUVCKiqoospT_vIi_M,9614
 liger_kernel/transformers/model/qwen3.py,sha256=w2jBHuK9kK9EmOr5dnEIXNQXUgUSV_sJUkXSEwxLPHs,4885
-liger_kernel/transformers/model/qwen3_moe.py,sha256=CbLP4eltlmPPkcSJ2WMe61P7_n-ksKF-bzyWfEMNXFg,5513
+liger_kernel/transformers/model/qwen3_moe.py,sha256=BkpfFH3fOH0yRfA7LF-AoHTLut2GV0Y4MOlkiIYewfU,5511
 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-0.5.10.dev20250605210201.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.5.10.dev20250605210201.dist-info/METADATA,sha256=AHzrxsgoWvM8SMrpFXF1LHayfA--Wmjdh9PdVUcpCSE,24309
-liger_kernel_nightly-0.5.10.dev20250605210201.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.5.10.dev20250605210201.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.5.10.dev20250605210201.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.5.10.dev20250605210201.dist-info/RECORD,,
+liger_kernel_nightly-0.5.10.dev20250605223455.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.5.10.dev20250605223455.dist-info/METADATA,sha256=jtKbBFfhtiyDQ7ZfpSZ1EwxGFNTYt0ND_4jL8Xr_pmc,24309
+liger_kernel_nightly-0.5.10.dev20250605223455.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.5.10.dev20250605223455.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.5.10.dev20250605223455.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.5.10.dev20250605223455.dist-info/RECORD,,