liger-kernel-nightly 0.5.4.dev20250225020243__py3-none-any.whl → 0.5.4.dev20250304205249__py3-none-any.whl

liger_kernel/chunked_loss/cpo_loss.py

@@ -39,8 +39,9 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
 
         return loss, chosen_rewards, rejected_rewards
 
-    @staticmethod
+    @classmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
@@ -53,13 +54,13 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
         compute_nll_loss=True,
         compiled=True,
     ):
-        return LigerFusedLinearPreferenceBase.forward(
-            ctx,
-            _input,
-            weight,
-            target,
-            bias,
-            loss_fn=LigerFusedLinearCPOFunction.preference_loss_fn,
+        return super().forward(
+            cls=cls,
+            ctx=ctx,
+            _input=_input,
+            weight=weight,
+            target=target,
+            bias=bias,
             ignore_index=ignore_index,
             alpha=alpha,
             beta=beta,

liger_kernel/chunked_loss/dpo_loss.py

@@ -52,8 +52,9 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
         loss = -F.logsigmoid(logits_diff).sum() / (full_target.shape[0] // 2)
         return loss, chosen_rewards, rejected_rewards
 
-    @staticmethod
+    @classmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
@@ -68,13 +69,13 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
         compiled=True,
         use_ref_model=True,
     ):
-        return LigerFusedLinearPreferenceBase.forward(
+        return super().forward(
+            cls=cls,
             ctx=ctx,
             _input=_input,
             weight=weight,
             target=target,
             bias=bias,
-            loss_fn=LigerFusedLinearDPOFunction.preference_loss_fn,
             ignore_index=ignore_index,
             beta=beta,
             compute_nll_loss=compute_nll_loss,

liger_kernel/chunked_loss/fused_linear_distillation.py

@@ -125,6 +125,7 @@ class LigerFusedLinearDistillationBase(torch.autograd.Function):
 
     @staticmethod
     def forward(
+        cls,
         ctx,
         student_input,
         student_weight,
@@ -133,7 +134,6 @@ class LigerFusedLinearDistillationBase(torch.autograd.Function):
         target,
         student_bias=None,
         teacher_bias=None,
-        loss_fn=None,
         chunk_size=1024,
         ignore_index=-100,
         weight_hard_loss=0.5,
@@ -175,7 +175,7 @@ class LigerFusedLinearDistillationBase(torch.autograd.Function):
 
         loss_func_to_call = partial(
             LigerFusedLinearDistillationBase._compute_loss,
-            distillation_loss_fn=loss_fn,
+            distillation_loss_fn=cls.distillation_loss_fn,
             full_target=target,
             ignore_index=ignore_index,
             weight_hard_loss=weight_hard_loss,
@@ -263,4 +263,4 @@ class LigerFusedLinearDistillationBase(torch.autograd.Function):
             grad_weight = grad_weight * grad_output
             grad_bias = grad_bias * grad_output if grad_bias is not None else None
 
-        return grad_input, grad_weight, None, grad_bias
+        return grad_input, grad_weight, None, None, None, grad_bias

liger_kernel/chunked_loss/fused_linear_preference.py

@@ -16,12 +16,12 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
 
     @staticmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
         target,
         bias=None,
-        loss_fn=None,
         chunk_size=1,
         ignore_index=-100,
         alpha=1.0,
@@ -89,7 +89,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
 
         compute_loss = partial(
             LigerFusedLinearPreferenceBase._compute_loss,
-            preference_loss_fn=loss_fn,
+            preference_loss_fn=cls.preference_loss_fn,
             ignore_index=ignore_index,
             alpha=alpha,
             beta=beta,

liger_kernel/chunked_loss/fused_linear_rlhf.py

@@ -1,3 +1,4 @@
+from abc import abstractmethod
 from functools import partial
 
 import torch
@@ -5,15 +6,22 @@ import torch.nn.functional as F
 
 
 class LigerFusedLinearRLHFBase(torch.autograd.Function):
+    @abstractmethod
+    def rlhf_loss_fn(*args, **kwargs):
+        """
+        To be extended by subclasses.
+        """
+        raise NotImplementedError("RLHF loss function must be implemented.")
+
     @staticmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
         attention_mask,
         rewards,
         bias=None,
-        loss_fn=None,
         num_generations=4,
         beta=0.1,
         compiled=True,
@@ -41,7 +49,7 @@ class LigerFusedLinearRLHFBase(torch.autograd.Function):
             use_ref_model=use_ref_model,
             ref_weight=ref_weight,
             ref_bias=ref_bias,
-            rlhf_loss_fn=loss_fn,
+            rlhf_loss_fn=cls.rlhf_loss_fn,
         )
 
         def fused_fwd_bwd(input_chunk, attention_mask_chunk, rewards_chunk, ref_input_chunk):
@@ -202,7 +210,6 @@ class LigerFusedLinearRLHFBase(torch.autograd.Function):
             None,  # grad_attention_mask
             None,  # grad_rewards
             grad_bias,
-            None,  # grad_loss_fn
             None,  # grad_chunk_size
             None,  # grad_beta
             None,  # grad_compiled

liger_kernel/chunked_loss/fused_linear_unpaired_preference.py

@@ -16,13 +16,13 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
 
     @staticmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
         target,
         preference_labels,
         bias=None,
-        loss_fn=None,
         chunk_size=1,
         ignore_index=-100,
         compiled=True,
@@ -30,6 +30,7 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
         ref_input=None,
         ref_weight=None,
         ref_bias=None,
+        average_log_prob=False,
         **loss_kwargs,
     ):
         """
@@ -59,6 +60,7 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
                 Shape: (batch_size,).
             ref_weight (torch.Tensor): Reference weight tensor. Shape: (vocab_size, hidden_size).
             ref_bias (torch.Tensor, optional): Reference bias tensor. Shape: (vocab_size,).
+            average_log_prob (bool): Whether to average the log probability per non-masked token.
             loss_kwargs (dict): Other possible arguments that a loss function might need
         """
         # TODO: Tune CHUNK_SIZE to fully utilize the GPU
@@ -72,14 +74,22 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
         # Loss to be accumulated
         loss_acc = torch.zeros((), device=_input.device)
 
+        # Metrics to be recorded
+        chosen_logps_sum = torch.zeros((), device=_input.device)
+        rejected_logps_sum = torch.zeros((), device=_input.device)
+        chosen_logits_sum = torch.zeros((), device=_input.device)
+        rejected_logits_sum = torch.zeros((), device=_input.device)
+        aggregated_aux_outputs = []
+
         compute_loss = partial(
             LigerFusedLinearUnpairedPreferenceBase._compute_loss,
-            preference_loss_fn=loss_fn,
+            preference_loss_fn=cls.preference_loss_fn,
             full_target=target,
             ignore_index=ignore_index,
             use_ref_model=use_ref_model,
             ref_weight=ref_weight,
             ref_bias=ref_bias,
+            average_log_prob=average_log_prob,
             **loss_kwargs,
         )
 
@@ -88,7 +98,7 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
             Fused forward and backward pass for a chunk of input and target.
             """
             argnums = (0, 1, 4) if bias is not None else (0, 1)
-            return torch.func.grad_and_value(compute_loss, argnums=argnums, has_aux=False)(
+            return torch.func.grad_and_value(compute_loss, argnums=argnums, has_aux=True)(
                 input_chunk,
                 weight,
                 target_chunk,
@@ -103,9 +113,19 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
             preference_labels_chunk=None,
             ref_input_chunk=None,
         ):
-            (chunk_grad_input, chunk_grad_weight, *chunk_grad_bias), (chunk_loss) = fused_fwd_bwd(
-                input_chunk, target_chunk, preference_labels_chunk, ref_input_chunk
-            )
+            (
+                (chunk_grad_input, chunk_grad_weight, *chunk_grad_bias),
+                (
+                    chunk_loss,
+                    (
+                        chunk_chosen_logps_sum,
+                        chunk_rejected_logps_sum,
+                        chunk_chosen_logits_sum,
+                        chunk_rejected_logits_sum,
+                        *aux_outputs,
+                    ),
+                ),
+            ) = fused_fwd_bwd(input_chunk, target_chunk, preference_labels_chunk, ref_input_chunk)
             if bias is not None:
                 grad_bias.add_(chunk_grad_bias[0])  # accumulate bias gradient
 
@@ -116,6 +136,23 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
             # Accumulate loss
             loss_acc.add_(chunk_loss)
 
+            # Accumulate metrics
+            chosen_logps_sum.add_(chunk_chosen_logps_sum)
+            rejected_logps_sum.add_(chunk_rejected_logps_sum)
+            chosen_logits_sum.add_(chunk_chosen_logits_sum)
+            rejected_logits_sum.add_(chunk_rejected_logits_sum)
+
+            # aux_outputs
+            # Initialize storage for aux_outputs
+            if len(aggregated_aux_outputs) == 0:
+                for aux in aux_outputs:
+                    aggregated_aux_outputs.append(torch.zeros((), device=aux.device))
+
+            # Process each aux_output
+            for i, aux in enumerate(aux_outputs):
+                if aux.ndim == 0:
+                    aggregated_aux_outputs[i].add_(aux)
+
         if compiled:
             fused_fwd_bwd = torch.compile(fused_fwd_bwd)
 
@@ -151,12 +188,25 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
             # accumulate loss, gradients, and metrics
             accumulate_chunk(input_chunk, target_chunk, preference_labels_chunk, ref_input_chunk)
 
+        # Aggregate aux outputs lists into tensors
+        for i, aux in enumerate(aggregated_aux_outputs):
+            if isinstance(aux, list):
+                aggregated_aux_outputs[i] = torch.cat(aux, dim=0)
+
         ctx.save_for_backward(
             torch.cat(grad_inputs, dim=0),
             grad_weight,
             grad_bias,
         )
-        return loss_acc
+
+        return_vars = (
+            chosen_logps_sum,
+            rejected_logps_sum,
+            chosen_logits_sum,
+            rejected_logits_sum,
+        )
+
+        return loss_acc, (*return_vars, *aggregated_aux_outputs)
 
     @staticmethod
     def backward(ctx, *grad_output):
@@ -173,21 +223,37 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
         input_chunk,
         weight,
         target_chunk,
+        preference_labels_chunk,
         bias=None,
         ignore_index=-100,
+        average_log_prob=False,
     ):
         logits_chunk = input_chunk @ weight.t()
         if bias is not None:
             logits_chunk = logits_chunk + bias
         log_probs_chunk = F.log_softmax(logits_chunk.float(), dim=-1)
-
         loss_mask_chunk = target_chunk != ignore_index
         label_chunk = torch.where(loss_mask_chunk, target_chunk, 0)
 
         per_token_logps_chunk = log_probs_chunk.gather(-1, label_chunk.unsqueeze(-1)).squeeze(-1)
-        average_log_prob_chunk = (per_token_logps_chunk * loss_mask_chunk).sum(-1) / loss_mask_chunk.sum(-1)
-
-        return average_log_prob_chunk
+        if average_log_prob:
+            log_probs = (per_token_logps_chunk * loss_mask_chunk).sum(-1) / loss_mask_chunk.sum(-1)
+        else:
+            log_probs = (per_token_logps_chunk * loss_mask_chunk).sum(-1)
+
+        chosen_logps_sum = (log_probs * preference_labels_chunk.unsqueeze(1)).sum()
+        rejected_logps_sum = (log_probs * (~preference_labels_chunk).unsqueeze(1)).sum()
+
+        chosen_logits_sum = (logits_chunk * preference_labels_chunk.unsqueeze(1)).sum()
+        rejected_logits_sum = (logits_chunk * (~preference_labels_chunk).unsqueeze(1)).sum()
+
+        return (
+            log_probs,
+            chosen_logps_sum,
+            rejected_logps_sum,
+            chosen_logits_sum,
+            rejected_logits_sum,
+        )
 
     @staticmethod
     def _compute_loss(
@@ -203,6 +269,7 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
         ref_input_chunk=None,
         ref_weight=None,
         ref_bias=None,
+        average_log_prob=False,
         **loss_kwargs,
     ):
         """
@@ -218,29 +285,57 @@ class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
             use_ref_model (bool): Whether to use a reference model for the alignment loss.
             ref_weight (torch.Tensor): Reference weight tensor. Shape: (vocab_size, hidden_size).
             ref_bias (torch.Tensor, optional): Reference bias tensor. Shape: (vocab_size,).
+            average_log_prob (bool): Whether to average the log probability per non-masked token.
             loss_kwargs (dict): Additional arguments for the loss function.
         """
-        average_log_prob_chunk = LigerFusedLinearUnpairedPreferenceBase.chunk_forward(
+        (
+            log_prob_chunk,
+            chosen_logps_sum,
+            rejected_logps_sum,
+            chosen_logits_sum,
+            rejected_logits_sum,
+        ) = LigerFusedLinearUnpairedPreferenceBase.chunk_forward(
             input_chunk,
             weight,
             target_chunk,
+            preference_labels_chunk,
             bias=bias,
             ignore_index=ignore_index,
+            average_log_prob=average_log_prob,
         )
 
         if use_ref_model:
             with torch.no_grad():
-                ref_average_log_prob_chunk = LigerFusedLinearUnpairedPreferenceBase.chunk_forward(
+                (
+                    ref_log_prob_chunk,
+                    _,
+                    _,
+                    _,
+                    _,
+                ) = LigerFusedLinearUnpairedPreferenceBase.chunk_forward(
                     ref_input_chunk,
                     ref_weight,
                     target_chunk,
+                    preference_labels_chunk,
                     ref_bias,
                     ignore_index=ignore_index,
+                    average_log_prob=average_log_prob,
                 )
-            loss_kwargs["ref_average_log_prob_chunk"] = ref_average_log_prob_chunk
+            loss_kwargs["ref_log_prob_chunk"] = ref_log_prob_chunk
 
-        preference_loss_chunk = preference_loss_fn(
-            average_log_prob_chunk, preference_labels_chunk, full_target, **loss_kwargs
+        preference_loss_outputs = preference_loss_fn(
+            log_prob_chunk, preference_labels_chunk, full_target, **loss_kwargs
+        )
+        if isinstance(preference_loss_outputs, tuple):
+            preference_loss_chunk, *aux_outputs = preference_loss_outputs
+        else:
+            preference_loss_chunk, aux_outputs = preference_loss_outputs, []
+
+        return_vars = (
+            chosen_logps_sum,
+            rejected_logps_sum,
+            chosen_logits_sum,
+            rejected_logits_sum,
         )
 
-        return preference_loss_chunk
+        return preference_loss_chunk, (*return_vars, *aux_outputs)

liger_kernel/chunked_loss/grpo_loss.py

@@ -63,8 +63,9 @@ class LigerFusedLinearGRPOFunction(LigerFusedLinearRLHFBase):
 
         return loss, metrics
 
-    @staticmethod
+    @classmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
@@ -79,12 +80,12 @@ class LigerFusedLinearGRPOFunction(LigerFusedLinearRLHFBase):
         use_ref_model=True,
         num_generations=1,
     ):
-        return LigerFusedLinearRLHFBase.forward(
+        return super().forward(
+            cls=cls,
             ctx=ctx,
             _input=_input,
             weight=weight,
             attention_mask=attention_mask,
-            loss_fn=LigerFusedLinearGRPOFunction.rlhf_loss_fn,
             rewards=rewards,
             bias=bias,
             ref_input=ref_input,

liger_kernel/chunked_loss/jsd_loss.py

@@ -30,14 +30,17 @@ class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
         jsd_loss = beta * teacher_kl + (1 - beta) * student_kl
         return jsd_loss
 
-    @staticmethod
+    @classmethod
     def forward(
+        cls,
         ctx,
         student_input: torch.Tensor,
         student_weight: torch.Tensor,
         teacher_input: torch.Tensor,
         teacher_weight: torch.Tensor,
         true_labels: torch.LongTensor,
+        student_bias: torch.Tensor,
+        teacher_bias: torch.Tensor,
         weight_hard_loss: float = 0.5,
         weight_soft_loss: float = 0.5,
         beta: float = 0.5,
@@ -62,15 +65,17 @@ class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
         Returns:
             torch.Tensor: Computed loss
         """
-        return LigerFusedLinearDistillationBase.forward(
+        return super().forward(
+            cls=cls,
             ctx=ctx,
             student_input=student_input,
             student_weight=student_weight,
             teacher_input=teacher_input,
             teacher_weight=teacher_weight,
             target=true_labels,
-            loss_fn=LigerFusedLinearJSDFunction.distillation_loss_fn,
-            chunk_size=1,
+            student_bias=student_bias,
+            teacher_bias=teacher_bias,
+            chunk_size=1024,
             weight_hard_loss=weight_hard_loss,
             weight_soft_loss=weight_soft_loss,
             beta=beta,
@@ -81,9 +86,18 @@ class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
 
     @staticmethod
     def backward(ctx, grad_output):
-        grads = LigerFusedLinearDistillationBase.backward(ctx, grad_output)[:4]
+        grads = LigerFusedLinearDistillationBase.backward(ctx, grad_output)[:6]
 
-        return (*grads, None, None, None, None, None, None, None)
+        return (
+            *grads,
+            None,  # teacher_bias
+            None,  # weight_hard_loss
+            None,  # weight_soft_loss
+            None,  # beta
+            None,  # ignore_index
+            None,  # temperature
+            None,  # compiled
+        )
 
 
 class LigerFusedLinearJSDLoss(torch.nn.Module):
@@ -125,6 +139,8 @@ class LigerFusedLinearJSDLoss(torch.nn.Module):
         teacher_input: torch.Tensor,
         teacher_weight: torch.Tensor,
         true_labels: torch.LongTensor,
+        student_bias: torch.Tensor,
+        teacher_bias: torch.Tensor,
     ) -> torch.Tensor:
         """
         Compute the JSD distillation loss.
@@ -145,6 +161,8 @@ class LigerFusedLinearJSDLoss(torch.nn.Module):
             teacher_input,
             teacher_weight,
             true_labels,
+            student_bias,
+            teacher_bias,
             self.weight_hard_loss,
             self.weight_soft_loss,
             self.beta,

liger_kernel/chunked_loss/kto_loss.py

@@ -7,10 +7,10 @@ from liger_kernel.chunked_loss.fused_linear_unpaired_preference import LigerFuse
 class LigerFusedLinearKTOFunction(LigerFusedLinearUnpairedPreferenceBase):
     @staticmethod
     def preference_loss_fn(
-        average_log_prob_chunk,
+        log_prob_chunk,
         preference_labels_chunk,
         full_target,
-        ref_average_log_prob_chunk=None,
+        ref_log_prob_chunk=None,
         beta=0.1,
         kl=None,
     ):
@@ -43,30 +43,34 @@ class LigerFusedLinearKTOFunction(LigerFusedLinearUnpairedPreferenceBase):
         3. Maintain reasonable distance from the reference model
 
         Args:
-            average_log_prob_chunk: Log probabilities for the chunk (batch_size,)
+            log_prob_chunk: Log probabilities for the chunk (batch_size,)
             preference_labels_chunk: Preference labels for the chunk (batch_size,)
             full_target: Non chunked full target tensor
-            ref_average_log_prob_chunk: Reference log probs for the chunk (batch_size,)
+            ref_log_prob_chunk: Reference log probs for the chunk (batch_size,)
             beta: Weight for the KTO loss
             kl: KL divergence between the policy model and the reference model for the chosen responses. Shape: (batch_size,)
         Returns:
             - loss: The KTO loss value
         """
-        if ref_average_log_prob_chunk is not None:
-            logratios_chunk = average_log_prob_chunk - ref_average_log_prob_chunk
+        if ref_log_prob_chunk is not None:
+            logratios_chunk = log_prob_chunk - ref_log_prob_chunk
         else:
-            logratios_chunk = average_log_prob_chunk
-
+            logratios_chunk = log_prob_chunk
         multiplier_chunk = torch.where(preference_labels_chunk, 1, -1)
         if kl is not None:
             losses = 1 - F.sigmoid(beta * (logratios_chunk - kl) * multiplier_chunk)
         else:
             losses = 1 - F.sigmoid(beta * logratios_chunk * multiplier_chunk)
 
-        return losses.sum() / (full_target.shape[0])
+        rewards = beta * logratios_chunk
+        chosen_rewards_sum = (rewards * preference_labels_chunk.unsqueeze(1)).sum()
+        rejected_rewards_sum = (rewards * (~preference_labels_chunk).unsqueeze(1)).sum()
 
-    @staticmethod
+        return losses.sum() / (full_target.shape[0]), chosen_rewards_sum, rejected_rewards_sum
+
+    @classmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
@@ -81,15 +85,16 @@ class LigerFusedLinearKTOFunction(LigerFusedLinearUnpairedPreferenceBase):
         beta=0.1,
         compiled=True,
         use_ref_model=True,
+        average_log_prob=False,
     ):
-        return LigerFusedLinearUnpairedPreferenceBase.forward(
+        return super().forward(
+            cls=cls,
             ctx=ctx,
             _input=_input,
             weight=weight,
             target=target,
             preference_labels=preference_labels,
             bias=bias,
-            loss_fn=LigerFusedLinearKTOFunction.preference_loss_fn,
             ignore_index=ignore_index,
             beta=beta,
             compiled=compiled,
@@ -97,6 +102,7 @@ class LigerFusedLinearKTOFunction(LigerFusedLinearUnpairedPreferenceBase):
             ref_input=ref_input,
             ref_weight=ref_weight,
             ref_bias=ref_bias,
+            average_log_prob=average_log_prob,
             kl=kl,
         )
 
@@ -129,6 +135,7 @@ class LigerFusedLinearKTOLoss(torch.nn.Module):
         beta: float = 0.1,
         compiled: bool = True,
         use_ref_model: bool = False,
+        average_log_prob: bool = False,
     ):
         """
         Args:
@@ -136,12 +143,14 @@ class LigerFusedLinearKTOLoss(torch.nn.Module):
             beta (float): Temperature parameter for the KTO loss
             compiled (bool): Whether to use compiled operations
             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.
         """
         super().__init__()
         self.ignore_index = ignore_index
         self.beta = beta
         self.compiled = compiled
         self.use_ref_model = use_ref_model
+        self.average_log_prob = average_log_prob
 
     def forward(
         self,
@@ -169,4 +178,5 @@ class LigerFusedLinearKTOLoss(torch.nn.Module):
             self.beta,
             self.compiled,
             self.use_ref_model,
+            self.average_log_prob,
         )

liger_kernel/chunked_loss/orpo_loss.py

@@ -42,8 +42,9 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
 
         return loss, chosen_rewards, rejected_rewards, log_odds_ratio, log_odds_chosen
 
-    @staticmethod
+    @classmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
@@ -55,13 +56,13 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
         nll_target=None,
         compiled=True,
     ):
-        return LigerFusedLinearPreferenceBase.forward(
+        return super().forward(
+            cls=cls,
             ctx=ctx,
             _input=_input,
             weight=weight,
             target=target,
             bias=bias,
-            loss_fn=LigerFusedLinearORPOFunction.preference_loss_fn,
             ignore_index=ignore_index,
             beta=beta,
             compute_nll_loss=compute_nll_loss,

liger_kernel/chunked_loss/simpo_loss.py

@@ -47,8 +47,9 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
 
         return loss, chosen_rewards, rejected_rewards
 
-    @staticmethod
+    @classmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
@@ -62,13 +63,13 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
         compiled=True,
         gamma=0.5,
     ):
-        return LigerFusedLinearPreferenceBase.forward(
+        return super().forward(
+            cls,
             ctx,
             _input,
             weight,
             target,
             bias,
-            loss_fn=LigerFusedLinearSimPOFunction.preference_loss_fn,
             compute_nll_loss=compute_nll_loss,
             ignore_index=ignore_index,
             alpha=alpha,

liger_kernel/transformers/__init__.py

@@ -17,6 +17,7 @@ from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mllama
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_olmo2  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_phi3  # noqa: F401
 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.rms_norm import LigerRMSNorm  # noqa: F401
 from liger_kernel.transformers.rope import liger_rotary_pos_emb  # noqa: F401

liger_kernel/transformers/model/qwen2_5_vl.py

@@ -0,0 +1,205 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from torch.nn import CrossEntropyLoss
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import _CONFIG_FOR_DOC
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import QWEN2_5_VL_INPUTS_DOCSTRING
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLCausalLMOutputWithPast
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
+
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+
+
+@add_start_docstrings_to_model_forward(QWEN2_5_VL_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=Qwen2_5_VLCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    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,
+) -> Union[Tuple, Qwen2_5_VLCausalLMOutputWithPast]:
+    r"""
+    Copy paste Qwen2_5_VL's forward but replace torch cross entropy with liger fused linear cross entropy
+    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]`.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
+
+    >>> model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+    >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+
+    >>> 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
+
+    if inputs_embeds is None:
+        inputs_embeds = self.model.embed_tokens(input_ids)
+        if pixel_values is not None:
+            pixel_values = pixel_values.type(self.visual.dtype)
+            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+            n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
+            n_image_features = image_embeds.shape[0]
+            if n_image_tokens != n_image_features:
+                raise ValueError(
+                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+                )
+
+            mask = input_ids == self.config.image_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            image_mask = mask_expanded.to(inputs_embeds.device)
+
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
+
+        if pixel_values_videos is not None:
+            pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
+            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+            n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
+            n_video_features = video_embeds.shape[0]
+            if n_video_tokens != n_video_features:
+                raise ValueError(
+                    f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
+                )
+
+            mask = input_ids == self.config.video_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            video_mask = mask_expanded.to(inputs_embeds.device)
+
+            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
+
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(inputs_embeds.device)
+
+    # if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
+    if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
+        # calculate RoPE index once per generation in the pre-fill stage only
+        if (cache_position is not None and cache_position[0] == 0) or self.rope_deltas is None:
+            position_ids, rope_deltas = self.get_rope_index(
+                input_ids,
+                image_grid_thw,
+                video_grid_thw,
+                second_per_grid_ts,
+                attention_mask,
+            )
+            self.rope_deltas = rope_deltas
+        # then use the prev pre-calculated rope-deltas to get the correct position ids
+        else:
+            batch_size, seq_length, _ = inputs_embeds.shape
+            delta = (cache_position[0] + self.rope_deltas).to(inputs_embeds.device) if cache_position is not None else 0
+            position_ids = torch.arange(seq_length, device=inputs_embeds.device)
+            position_ids = position_ids.view(1, -1).expand(batch_size, -1)
+            if cache_position is not None:  # otherwise `deltas` is an int `0`
+                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
+            position_ids = position_ids.add(delta)
+            position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)
+
+    outputs = self.model(
+        input_ids=None,
+        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,
+    )
+
+    hidden_states = outputs[0]
+
+    loss = None
+    logits = None
+
+    if self.training and (labels is not None):
+        shift_hidden_states = hidden_states[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+
+        # Flatten tokens
+        shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
+        shift_labels = shift_labels.view(-1)
+
+        lce = LigerFusedLinearCrossEntropyLoss()
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+    else:
+        logits = self.lm_head(hidden_states)
+        if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Qwen2_5_VLCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=rope_deltas,
+    )

liger_kernel/transformers/monkey_patch.py

@@ -745,6 +745,73 @@ def apply_liger_kernel_to_qwen2_vl(
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_qwen2_5_vl(
+    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 Qwen2.5-VL models.
+    NOTE: Qwen2.5-VL is not available in transformers<4.48.2
+
+    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 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.qwen2_5_vl import modeling_qwen2_5_vl
+    from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLModel
+
+    from liger_kernel.transformers.model.qwen2_5_vl import lce_forward as qwen2_5_vl_lce_forward
+
+    if rope:
+        modeling_qwen2_5_vl.apply_multimodal_rotary_pos_emb = liger_multimodal_rotary_pos_emb
+    if rms_norm:
+        modeling_qwen2_5_vl.Qwen2RMSNorm = LigerRMSNorm
+    if cross_entropy:
+        modeling_qwen2_5_vl.CrossEntropyLoss = LigerCrossEntropyLoss
+    if fused_linear_cross_entropy:
+        modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration.forward = qwen2_5_vl_lce_forward
+    if swiglu:
+        modeling_qwen2_5_vl.Qwen2MLP = 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: Qwen2_5_VLModel = getattr(model, model.base_model_prefix, model)
+
+        if hasattr(model, "visual"):
+            # Patch Qwen2_5_VisionTransformerPretrainedModel
+            for vision_block in model.visual.blocks:
+                if rms_norm:
+                    _patch_rms_norm_module(vision_block.norm1)
+                    _patch_rms_norm_module(vision_block.norm2)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _bind_method_to_module(decoder_layer.mlp, "forward", LigerSwiGLUMLP.forward)
+            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_phi3(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -890,6 +957,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "olmo2": apply_liger_kernel_to_olmo2,
     "qwen2": apply_liger_kernel_to_qwen2,
     "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_nightly-0.5.4.dev20250225020243.dist-info → liger_kernel_nightly-0.5.4.dev20250304205249.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.4.dev20250225020243
+Version: 0.5.4.dev20250304205249
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -312,6 +312,7 @@ loss.backward()
 | Gemma2      | `liger_kernel.transformers.apply_liger_kernel_to_gemma2`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | 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        |
 | 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 |

{liger_kernel_nightly-0.5.4.dev20250225020243.dist-info → liger_kernel_nightly-0.5.4.dev20250304205249.dist-info}/RECORD

@@ -3,18 +3,18 @@ liger_kernel/env_report.py,sha256=uhdEC8OydxoZlb7B6YYcAaBF3crGFdIck-4cxaW4NJY,17
 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=OdBR8WYdHTKpLI_c9DcuwqKSWPeAAeTyREz46Vu_cAY,3682
-liger_kernel/chunked_loss/dpo_loss.py,sha256=wgjnwzLfrMUwV5mXgrq6G1YfQKWnbiFJegaP48BGJHY,4509
+liger_kernel/chunked_loss/cpo_loss.py,sha256=LZr2mVe2EUQuKQ9uB30vxmuCfuK_yxJZraG62kH-qSY,3654
+liger_kernel/chunked_loss/dpo_loss.py,sha256=MMTiHviJUg7htStKevMgqpyXkQfaKcqahaBz1dzXM1A,4451
 liger_kernel/chunked_loss/functional.py,sha256=THWWpCnRVhTVfnPnyvQjdBvo1JDtxhwLmtZE_yiBBqM,817
-liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=5V8rdva89WyHVbmJ8JOmC4DYNOR6ByXfx3qlUieOZkI,11002
-liger_kernel/chunked_loss/fused_linear_preference.py,sha256=idK9V9NivoVITqVpiG0fEGUHSvinYWkn9-EYXZjR-KQ,18356
-liger_kernel/chunked_loss/fused_linear_rlhf.py,sha256=sAApL4GQ3YL2F-ymIAF61GCpFfBgFcWF5LB4Gzd7LgY,8044
-liger_kernel/chunked_loss/fused_linear_unpaired_preference.py,sha256=ZqYlXXhIphkJPxOS7iI70avgrr6x0skEtgpckZTYau0,9819
-liger_kernel/chunked_loss/grpo_loss.py,sha256=M5qlQR-v5Rh8N3P3dPGNhOKygDFJ4516_rJaVPzU_-c,4980
-liger_kernel/chunked_loss/jsd_loss.py,sha256=yRCQdvd3ruTWP4A_BfU8VcZ6LepSUfO0Ob7stGnueQY,6052
-liger_kernel/chunked_loss/kto_loss.py,sha256=b3ffJyk97e-6XdXd4HFrYyx8wW4A-CU4gOaJSimKLtA,5476
-liger_kernel/chunked_loss/orpo_loss.py,sha256=yjcrrbVeemLYodoSKT-FMSnaPtyKAZ3aOrvPD6tTY6Y,3617
-liger_kernel/chunked_loss/simpo_loss.py,sha256=3TTc7U79Orjgi-Wu81WZkWk5MgsdqKXIOBHgIvDazPw,3865
+liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=FJh7k3sry-fqnBApLSngf7h-lHQEiXtOY_tiRDVanPM,11022
+liger_kernel/chunked_loss/fused_linear_preference.py,sha256=ojB42jYPu0c4ki96Ft-hy7Sf6fh_WikG-aWNrlZzSio,18362
+liger_kernel/chunked_loss/fused_linear_rlhf.py,sha256=Dxtj6fOfZ8jGoX5uvC9z3An6np39Zdz3c2B2fF39WAA,8240
+liger_kernel/chunked_loss/fused_linear_unpaired_preference.py,sha256=RiuK3UtRwH9T6jZ36sA8Urj-TVuOLOO2syLg_JOQapY,13437
+liger_kernel/chunked_loss/grpo_loss.py,sha256=6RPYSIzTgjmCSIkULvpmbU_jsEM4awU9OgScBkWsXj0,4933
+liger_kernel/chunked_loss/jsd_loss.py,sha256=NbT_D0ybYL68p7Rh9hdKeRR8HkHKEyX2ab5n_RxWz-0,6481
+liger_kernel/chunked_loss/kto_loss.py,sha256=4NHloq8gAwkD4Q8-PNa4Yz_uNyu80PokOdFTOnH1138,5903
+liger_kernel/chunked_loss/orpo_loss.py,sha256=WXX_eCSSIDoi9bHOK8-DYFgx3-vEZhvDvLyXK3mVW6U,3558
+liger_kernel/chunked_loss/simpo_loss.py,sha256=V2YN7aB7o-5rMplgYTon04JKYhKNYxLMqZezUOG5OPc,3801
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 liger_kernel/ops/cross_entropy.py,sha256=yKKhN63I7r9NxJye4wTLBvvKAyrXQt6jf4nBo3lJyVg,18860
 liger_kernel/ops/fused_linear_cross_entropy.py,sha256=1Y3Uk_TCSjqKgoG2eot1ptnWXJXXQESqGvOmqAW1gsM,10912
@@ -32,7 +32,7 @@ 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=6v_VcV1GQ9ISgNCd-ZxtmEg_s5GTBQ9F-s1KrFkYzPQ,2265
+liger_kernel/transformers/__init__.py,sha256=4bwMPQhGHxmZ-WTFAMD9m-s0PYyfcvIRxhq_h3b0Wz0,2363
 liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
 liger_kernel/transformers/cross_entropy.py,sha256=z3KTWQnFxr_IZaVjtYt0ZNEWQdDdYThN35xWkHlDGH0,1683
 liger_kernel/transformers/functional.py,sha256=ShLD3eb--XKNtllznCrOYTbo4f-1KVwzi0KLMICdrn4,4942
@@ -43,7 +43,7 @@ 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=g3i3q5McBg23A3Mnviw-Eb32le1hvN7jByzONa9ngcs,44000
+liger_kernel/transformers/monkey_patch.py,sha256=9ud9tv1LI9WIa9UDu0abGIiusIIkayO1fjAUMWgwwT0,47096
 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
@@ -61,14 +61,15 @@ liger_kernel/transformers/model/mllama.py,sha256=qWexBdskuN3gPJvPUwt4J0nU675tGD6
 liger_kernel/transformers/model/olmo2.py,sha256=yyksS6E4fuWd8asEW8rEDBKqZpFmP4ITCM_bjIDZaoY,5124
 liger_kernel/transformers/model/phi3.py,sha256=biRa8fph9qdnQmkD9I21t5XIjpIt1i6UKU4uk8Up8pU,10292
 liger_kernel/transformers/model/qwen2.py,sha256=14UuPjxB-tjqWn85Tn4fqBFvVhVsth5iPEt8kJSMiew,9581
+liger_kernel/transformers/model/qwen2_5_vl.py,sha256=l71WBfX0ptrisoURIRwXJH7MQ2vGKOvcRYMNsrydwlQ,9455
 liger_kernel/transformers/model/qwen2_vl.py,sha256=yMLqsfSYcvhClUpTUjGoADiOxfLB2B8240VdrPP0c8s,9851
 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_nightly-0.5.4.dev20250225020243.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.5.4.dev20250225020243.dist-info/METADATA,sha256=8MIiL5w8zFdcQMmctVN8Ca-A1pk2UG2e-aKDtlRaKf4,22234
-liger_kernel_nightly-0.5.4.dev20250225020243.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.5.4.dev20250225020243.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-liger_kernel_nightly-0.5.4.dev20250225020243.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.5.4.dev20250225020243.dist-info/RECORD,,
+liger_kernel_nightly-0.5.4.dev20250304205249.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.5.4.dev20250304205249.dist-info/METADATA,sha256=AFWOuf67EudNiAgUU006ngrmfAzKB3bYTnjYtUCYZb0,22389
+liger_kernel_nightly-0.5.4.dev20250304205249.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.5.4.dev20250304205249.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+liger_kernel_nightly-0.5.4.dev20250304205249.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.5.4.dev20250304205249.dist-info/RECORD,,