liger-kernel-nightly 0.6.2.dev20251011154427__py3-none-any.whl → 0.6.4.dev20260107111351__py3-none-any.whl

liger_kernel/transformers/__init__.py

@@ -15,6 +15,7 @@ from liger_kernel.transformers.layer_norm import LigerLayerNorm  # noqa: F401
 from liger_kernel.transformers.llama4_rope import liger_llama4_text_rotary_pos_emb  # noqa: F401
 from liger_kernel.transformers.llama4_rope import liger_llama4_vision_rotary_pos_emb  # noqa: F401
 from liger_kernel.transformers.multi_token_attention import LigerMultiTokenAttention  # noqa: F401
+from liger_kernel.transformers.poly_norm import LigerPolyNorm  # noqa: F401
 from liger_kernel.transformers.rms_norm import LigerRMSNorm  # noqa: F401
 from liger_kernel.transformers.rope import liger_rotary_pos_emb  # noqa: F401
 from liger_kernel.transformers.softmax import LigerSoftmax  # noqa: F401
@@ -23,6 +24,8 @@ from liger_kernel.transformers.swiglu import LigerBlockSparseTop2MLP  # noqa: F4
 from liger_kernel.transformers.swiglu import LigerPhi3SwiGLUMLP  # noqa: F401
 from liger_kernel.transformers.swiglu import LigerQwen3MoeSwiGLUMLP  # noqa: F401
 from liger_kernel.transformers.swiglu import LigerSwiGLUMLP  # noqa: F401
+from liger_kernel.transformers.tiled_mlp import LigerTiledGEGLUMLP  # noqa: F401
+from liger_kernel.transformers.tiled_mlp import LigerTiledSwiGLUMLP  # noqa: F401
 from liger_kernel.transformers.tvd import LigerTVDLoss  # noqa: F401
 
 # Static-only imports for IDEs and type checkers
@@ -38,7 +41,10 @@ if TYPE_CHECKING:
     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_glm4v  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4v_moe  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gpt_oss  # 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_hunyuan_v1_dense  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_hunyuan_v1_moe  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_internvl  # 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_llama4  # noqa: F401
@@ -47,6 +53,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mixtral  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mllama  # noqa: F401
     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_olmo3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_paligemma  # 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
@@ -54,7 +61,11 @@ if TYPE_CHECKING:
     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
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_moe  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_next  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_vl  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_vl_moe  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_smollm3  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_smolvlm  # noqa: F401
 
 
 # Check if 'transformers' is installed
@@ -100,6 +111,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_glm4",
         "apply_liger_kernel_to_glm4v",
         "apply_liger_kernel_to_glm4v_moe",
+        "apply_liger_kernel_to_gpt_oss",
         "apply_liger_kernel_to_granite",
         "apply_liger_kernel_to_internvl",
         "apply_liger_kernel_to_llama",
@@ -109,6 +121,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_mixtral",
         "apply_liger_kernel_to_mllama",
         "apply_liger_kernel_to_olmo2",
+        "apply_liger_kernel_to_olmo3",
         "apply_liger_kernel_to_paligemma",
         "apply_liger_kernel_to_phi3",
         "apply_liger_kernel_to_qwen2",
@@ -116,7 +129,13 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen2_vl",
         "apply_liger_kernel_to_qwen3",
         "apply_liger_kernel_to_qwen3_moe",
+        "apply_liger_kernel_to_qwen3_next",
+        "apply_liger_kernel_to_qwen3_vl",
+        "apply_liger_kernel_to_qwen3_vl_moe",
         "apply_liger_kernel_to_smollm3",
+        "apply_liger_kernel_to_smolvlm",
+        "apply_liger_kernel_to_hunyuan_v1_dense",
+        "apply_liger_kernel_to_hunyuan_v1_moe",
     }
 
     if name in monkey_patch_symbols:
@@ -137,6 +156,7 @@ __all__ = [
     "LigerJSD",
     "LigerLayerNorm",
     "LigerFusedAddRMSNorm",
+    "LigerPolyNorm",
     "LigerRMSNorm",
     "liger_rotary_pos_emb",
     "liger_llama4_text_rotary_pos_emb",
@@ -145,6 +165,8 @@ __all__ = [
     "LigerPhi3SwiGLUMLP",
     "LigerQwen3MoeSwiGLUMLP",
     "LigerSwiGLUMLP",
+    "LigerTiledGEGLUMLP",
+    "LigerTiledSwiGLUMLP",
     "LigerTVDLoss",
     "LigerKLDIVLoss",
     "LigerMultiTokenAttention",
@@ -167,6 +189,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_glm4",
             "apply_liger_kernel_to_glm4v",
             "apply_liger_kernel_to_glm4v_moe",
+            "apply_liger_kernel_to_gpt_oss",
             "apply_liger_kernel_to_granite",
             "apply_liger_kernel_to_internvl",
             "apply_liger_kernel_to_llama",
@@ -176,6 +199,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_mixtral",
             "apply_liger_kernel_to_mllama",
             "apply_liger_kernel_to_olmo2",
+            "apply_liger_kernel_to_olmo3",
             "apply_liger_kernel_to_paligemma",
             "apply_liger_kernel_to_phi3",
             "apply_liger_kernel_to_qwen2",
@@ -183,6 +207,12 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_qwen2_vl",
             "apply_liger_kernel_to_qwen3",
             "apply_liger_kernel_to_qwen3_moe",
+            "apply_liger_kernel_to_qwen3_next",
+            "apply_liger_kernel_to_qwen3_vl",
+            "apply_liger_kernel_to_qwen3_vl_moe",
             "apply_liger_kernel_to_smollm3",
+            "apply_liger_kernel_to_smolvlm",
+            "apply_liger_kernel_to_hunyuan_v1_dense",
+            "apply_liger_kernel_to_hunyuan_v1_moe",
         ]
     )

liger_kernel/transformers/auto_model.py

@@ -1,4 +1,5 @@
 import inspect
+import logging
 
 from transformers import AutoConfig
 from transformers import AutoModelForCausalLM
@@ -6,6 +7,8 @@ from transformers import AutoModelForCausalLM
 from liger_kernel.transformers.monkey_patch import MODEL_TYPE_TO_APPLY_LIGER_FN
 from liger_kernel.transformers.monkey_patch import _apply_liger_kernel
 
+logger = logging.getLogger(__name__)
+
 
 def _get_model_config(model_dir, **model_init_kwargs):
     config = AutoConfig.from_pretrained(model_dir, **model_init_kwargs)
@@ -36,3 +39,21 @@ class AutoLigerKernelForCausalLM(AutoModelForCausalLM):
         applicable_kwargs = {key: value for key, value in kwargs.items() if key not in apply_fn_signature.parameters}
 
         return super().from_pretrained(pretrained_model_name_or_path, *model_args, **applicable_kwargs)
+
+    @classmethod
+    def from_config(cls, config, **kwargs):
+        model_type = getattr(config, "model_type", None)
+        if not model_type:
+            logger.info("Model type could not be determined from model config. No Liger kernels will be applied.")
+            return
+        model_type = config.model_type
+
+        _apply_liger_kernel(model_type, **kwargs)
+
+        # Filter out kwargs that were passed to the apply_liger_* function, which will cause
+        # model initialization errors otherwise
+        apply_fn = MODEL_TYPE_TO_APPLY_LIGER_FN[model_type]
+        apply_fn_signature = inspect.signature(apply_fn)
+        applicable_kwargs = {key: value for key, value in kwargs.items() if key not in apply_fn_signature.parameters}
+
+        return super().from_config(config, **applicable_kwargs)

liger_kernel/transformers/cross_entropy.py

@@ -2,7 +2,8 @@ from typing import Optional
 
 import torch
 
-from liger_kernel.ops.cross_entropy import LigerCrossEntropyFunction
+from liger_kernel.ops import LigerCrossEntropyFunction
+from liger_kernel.transformers.functional import CrossEntropyOutput
 
 
 class LigerCrossEntropyLoss(torch.nn.Module):
@@ -15,6 +16,7 @@ class LigerCrossEntropyLoss(torch.nn.Module):
         reduction: str = "mean",
         softcap: Optional[float] = None,
         return_z_loss: bool = False,
+        return_token_accuracy: bool = False,
     ):
         super().__init__()
         assert (label_smoothing >= 0) and (label_smoothing <= 1), (
@@ -33,9 +35,10 @@ class LigerCrossEntropyLoss(torch.nn.Module):
         self.reduction = reduction
         self.softcap = softcap
         self.return_z_loss = return_z_loss
+        self.return_token_accuracy = return_token_accuracy
 
     def forward(self, _input: torch.Tensor, target: torch.Tensor):
-        loss, z_loss = LigerCrossEntropyFunction.apply(
+        loss, z_loss, token_accuracy = LigerCrossEntropyFunction.apply(
             _input,
             target,
             self.weight,
@@ -45,7 +48,9 @@ class LigerCrossEntropyLoss(torch.nn.Module):
             self.reduction,
             self.softcap,
             self.return_z_loss,
+            self.return_token_accuracy,
         )
-        if not self.return_z_loss:
+        if not self.return_z_loss and not self.return_token_accuracy:
             return loss
-        return loss, z_loss
+
+        return CrossEntropyOutput(loss=loss, z_loss=z_loss, token_accuracy=token_accuracy)

liger_kernel/transformers/dyt.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 
-from liger_kernel.ops.dyt import LigerDyTFunction
+from liger_kernel.ops import LigerDyTFunction
 
 
 class LigerDyT(nn.Module):

liger_kernel/transformers/experimental/embedding.py

@@ -3,7 +3,7 @@ from typing import Optional
 import torch
 import torch.nn as nn
 
-from liger_kernel.ops.experimental.embedding import LigerEmbeddingFunction
+from liger_kernel.ops import LigerEmbeddingFunction
 
 
 class LigerEmbedding(nn.Module):

liger_kernel/transformers/functional.py

@@ -1,24 +1,35 @@
+from dataclasses import dataclass
 from typing import Optional
 
-from liger_kernel.ops.cross_entropy import LigerCrossEntropyFunction
-from liger_kernel.ops.dyt import LigerDyTFunction
-from liger_kernel.ops.fused_add_rms_norm import LigerFusedAddRMSNormFunction
-from liger_kernel.ops.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyFunction
-from liger_kernel.ops.fused_linear_jsd import LigerFusedLinearJSDFunction
-from liger_kernel.ops.fused_neighborhood_attention import LigerFusedNeighborhoodAttentionFunction
-from liger_kernel.ops.geglu import LigerGELUMulFunction
-from liger_kernel.ops.group_norm import LigerGroupNormFunction
-from liger_kernel.ops.jsd import LigerJSDFunction
-from liger_kernel.ops.kl_div import LigerKLDivLossFunction
-from liger_kernel.ops.layer_norm import LigerLayerNormFunction
-from liger_kernel.ops.multi_token_attention import LigerMultiTokenAttentionFunction
-from liger_kernel.ops.qwen2vl_mrope import LigerQwen2VLMRopeFunction
-from liger_kernel.ops.rms_norm import LigerRMSNormFunction
-from liger_kernel.ops.rope import LigerRopeFunction
-from liger_kernel.ops.softmax import LigerSoftmaxFunction
-from liger_kernel.ops.sparsemax import LigerSparsemaxFunction
-from liger_kernel.ops.swiglu import LigerSiLUMulFunction
-from liger_kernel.ops.tvd import LigerTVDLossFunction
+import torch
+
+from liger_kernel.ops import LigerCrossEntropyFunction
+from liger_kernel.ops import LigerDyTFunction
+from liger_kernel.ops import LigerFusedAddRMSNormFunction
+from liger_kernel.ops import LigerFusedLinearCrossEntropyFunction
+from liger_kernel.ops import LigerFusedLinearJSDFunction
+from liger_kernel.ops import LigerFusedNeighborhoodAttentionFunction
+from liger_kernel.ops import LigerGELUMulFunction
+from liger_kernel.ops import LigerGroupNormFunction
+from liger_kernel.ops import LigerJSDFunction
+from liger_kernel.ops import LigerKLDivLossFunction
+from liger_kernel.ops import LigerLayerNormFunction
+from liger_kernel.ops import LigerMultiTokenAttentionFunction
+from liger_kernel.ops import LigerPolyNormFunction
+from liger_kernel.ops import LigerQwen2VLMRopeFunction
+from liger_kernel.ops import LigerRMSNormFunction
+from liger_kernel.ops import LigerRopeFunction
+from liger_kernel.ops import LigerSiLUMulFunction
+from liger_kernel.ops import LigerSoftmaxFunction
+from liger_kernel.ops import LigerSparsemaxFunction
+from liger_kernel.ops import LigerTVDLossFunction
+
+
+@dataclass
+class CrossEntropyOutput:
+    loss: torch.Tensor
+    z_loss: Optional[torch.Tensor] = None
+    token_accuracy: Optional[torch.Tensor] = None
 
 
 # conform to the function signature in https://pytorch.org/docs/stable/generated/torch.nn.functional.cross_entropy.html
@@ -35,8 +46,9 @@ def liger_cross_entropy(
     lse_square_scale: float = 0.0,
     softcap: Optional[float] = None,
     return_z_loss: bool = False,
+    return_token_accuracy: bool = False,
 ):
-    loss, z_loss = LigerCrossEntropyFunction.apply(
+    loss, z_loss, token_accuracy = LigerCrossEntropyFunction.apply(
         input,
         target,
         weight,
@@ -46,10 +58,13 @@ def liger_cross_entropy(
         reduction,
         softcap,
         return_z_loss,
+        return_token_accuracy,
     )
-    if not return_z_loss:
+
+    if not return_z_loss and not return_token_accuracy:
         return loss
-    return loss, z_loss
+
+    return CrossEntropyOutput(loss=loss, z_loss=z_loss, token_accuracy=token_accuracy)
 
 
 def liger_fused_linear_cross_entropy(
@@ -66,8 +81,9 @@ def liger_fused_linear_cross_entropy(
     return_z_loss: bool = False,
     accum_dtype=None,
     use_token_scaling: bool = False,
+    return_token_accuracy: bool = False,
 ):
-    loss, z_loss = LigerFusedLinearCrossEntropyFunction.apply(
+    loss, z_loss, token_accuracy = LigerFusedLinearCrossEntropyFunction.apply(
         input,
         weight,
         target,
@@ -81,10 +97,13 @@ def liger_fused_linear_cross_entropy(
         return_z_loss,
         accum_dtype,
         use_token_scaling,
+        return_token_accuracy,
     )
-    if not return_z_loss:
+
+    if not return_z_loss and not return_token_accuracy:
         return loss
-    return loss, z_loss
+
+    return CrossEntropyOutput(loss=loss, z_loss=z_loss, token_accuracy=token_accuracy)
 
 
 def liger_fused_linear_jsd(
@@ -258,6 +277,10 @@ def liger_rms_norm(X, W, eps, offset: float = 0.0, casting_mode: str = "llama",
     return LigerRMSNormFunction.apply(X, W, eps, offset, casting_mode, in_place)
 
 
+def liger_poly_norm(X, W, B, eps=1e-6, in_place=True):
+    return LigerPolyNormFunction.apply(X, W, B, eps, in_place)
+
+
 def liger_fused_add_rms_norm(X, R, W, eps, offset: float = 0.0, casting_mode: str = "llama", in_place: bool = True):
     return LigerFusedAddRMSNormFunction.apply(X, R, W, eps, offset, casting_mode, in_place)
 

liger_kernel/transformers/fused_add_rms_norm.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 
-from liger_kernel.ops.fused_add_rms_norm import LigerFusedAddRMSNormFunction
+from liger_kernel.ops import LigerFusedAddRMSNormFunction
 
 
 class LigerFusedAddRMSNorm(nn.Module):

liger_kernel/transformers/fused_linear_cross_entropy.py

@@ -2,7 +2,8 @@ from typing import Optional
 
 import torch
 
-from liger_kernel.ops.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyFunction
+from liger_kernel.ops import LigerFusedLinearCrossEntropyFunction
+from liger_kernel.transformers.functional import CrossEntropyOutput
 
 
 class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
@@ -17,6 +18,7 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         return_z_loss: bool = False,
         accum_dtype: Optional[torch.dtype] = None,
         use_token_scaling: bool = False,
+        return_token_accuracy: bool = False,
     ):
         super().__init__()
         assert (label_smoothing >= 0) and (label_smoothing <= 1), (
@@ -37,9 +39,10 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         self.return_z_loss = return_z_loss
         self.accum_dtype = accum_dtype
         self.use_token_scaling = use_token_scaling
+        self.return_token_accuracy = return_token_accuracy
 
     def forward(self, lin_weight, _input, target, bias=None):
-        loss, z_loss = LigerFusedLinearCrossEntropyFunction.apply(
+        loss, z_loss, token_accuracy = LigerFusedLinearCrossEntropyFunction.apply(
             _input,
             lin_weight,
             target,
@@ -53,7 +56,9 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
             self.return_z_loss,
             self.accum_dtype,
             self.use_token_scaling,
+            self.return_token_accuracy,
         )
-        if not self.return_z_loss:
+        if not self.return_z_loss and not self.return_token_accuracy:
             return loss
-        return loss, z_loss
+
+        return CrossEntropyOutput(loss=loss, z_loss=z_loss, token_accuracy=token_accuracy)

liger_kernel/transformers/fused_linear_jsd.py

@@ -2,7 +2,7 @@ from typing import Optional
 
 import torch
 
-from liger_kernel.ops.fused_linear_jsd import LigerFusedLinearJSDFunction
+from liger_kernel.ops import LigerFusedLinearJSDFunction
 
 
 class LigerFusedLinearJSD(torch.nn.Module):

liger_kernel/transformers/fused_neighborhood_attention.py

@@ -5,7 +5,7 @@ from typing import Optional
 import torch
 import torch.nn as nn
 
-from liger_kernel.ops.fused_neighborhood_attention import LigerFusedNeighborhoodAttentionFunction
+from liger_kernel.ops import LigerFusedNeighborhoodAttentionFunction
 
 
 class LigerFusedNeighborhoodAttention(nn.Module):

liger_kernel/transformers/geglu.py

@@ -1,6 +1,6 @@
 import torch.nn as nn
 
-from liger_kernel.ops.geglu import LigerGELUMulFunction
+from liger_kernel.ops import LigerGELUMulFunction
 
 
 class LigerGEGLUMLP(nn.Module):

liger_kernel/transformers/group_norm.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 
-from liger_kernel.ops.group_norm import LigerGroupNormFunction
+from liger_kernel.ops import LigerGroupNormFunction
 
 
 class LigerGroupNorm(nn.Module):

liger_kernel/transformers/grpo_loss.py

@@ -1,4 +1,7 @@
-from liger_kernel.ops.grpo_loss import GrpoLossFunction
+import torch
+
+from liger_kernel.chunked_loss.fused_linear_ppo import LigerFusedLinearPPOBase
+from liger_kernel.ops import GrpoLossFunction
 
 
 def triton_grpo_loss(
@@ -13,12 +16,20 @@ def triton_grpo_loss(
     eps_low=0.2,
     eps_high=0.4,
     inplace=True,
+    loss_type="dapo",
+    max_completion_length=None,
+    importance_sampling_level="token",
+    reduce=False,
 ):
     assert logits is not None and completion_ids is not None and advantages is not None, (
         "must provide logits、completion_ids and advantages"
     )
+    if importance_sampling_level != "token":
+        raise ValueError(
+            f"Triton GRPO loss only supports token-level importance sampling. Got {importance_sampling_level}."
+        )
 
-    return GrpoLossFunction.apply(
+    per_token_loss, per_token_kl, is_clipped = GrpoLossFunction.apply(
         logits,
         old_logp,
         ref_logp,
@@ -31,6 +42,50 @@ def triton_grpo_loss(
         eps_high,
         inplace,
     )
+    if not reduce:
+        return per_token_loss, per_token_kl, is_clipped
+
+    loss = _reduce_grpo_loss(
+        per_token_loss,
+        completion_mask,
+        loss_type=loss_type,
+        max_completion_length=max_completion_length,
+    )
+
+    metrics = []
+    if beta != 0.0 and per_token_kl is not None:
+        metrics.append(_masked_mean(per_token_kl, completion_mask))
+    metrics.append(_masked_mean(is_clipped.float(), completion_mask))
+    return loss, metrics
+
+
+def _reduce_grpo_loss(per_token_loss, completion_mask, loss_type, max_completion_length):
+    mask = completion_mask
+    if mask is None:
+        mask = torch.ones_like(per_token_loss, dtype=per_token_loss.dtype, device=per_token_loss.device)
+    mask = mask.to(per_token_loss.dtype)
+
+    if loss_type == "grpo":
+        per_seq = (per_token_loss * mask).sum(-1) / mask.sum(-1).clamp(min=1.0)
+        return per_seq.mean()
+    if loss_type == "bnpo":
+        return (per_token_loss * mask).sum() / mask.sum().clamp(min=1.0)
+    if loss_type == "dr_grpo":
+        if max_completion_length is None:
+            raise ValueError("max_completion_length must be provided when using loss_type='dr_grpo'")
+        batch = per_token_loss.shape[0]
+        return (per_token_loss * mask).sum() / (batch * max_completion_length)
+    if loss_type == "dapo":
+        normalizer = LigerFusedLinearPPOBase._compute_dapo_normalizer(mask)
+        return (per_token_loss * mask).sum() / normalizer
+    raise ValueError(f"Unsupported loss_type '{loss_type}' for Triton GRPO loss.")
+
+
+def _masked_mean(values, mask):
+    if mask is None:
+        mask = torch.ones_like(values, dtype=values.dtype, device=values.device)
+    mask = mask.to(values.dtype)
+    return (values * mask).sum() / mask.sum().clamp(min=1.0)
 
 
 # This is a demo how to use grpo_loss in GRPOTrainer. The Trl version must be 0.16

liger_kernel/transformers/jsd.py

@@ -2,7 +2,7 @@ from typing import Optional
 
 import torch
 
-from liger_kernel.ops.jsd import LigerJSDFunction
+from liger_kernel.ops import LigerJSDFunction
 
 
 class LigerJSD(torch.nn.Module):

liger_kernel/transformers/kl_div.py

@@ -1,6 +1,6 @@
 import torch.nn as nn
 
-from liger_kernel.ops.kl_div import LigerKLDivLossFunction
+from liger_kernel.ops import LigerKLDivLossFunction
 
 
 class LigerKLDIVLoss(nn.KLDivLoss):

liger_kernel/transformers/layer_norm.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 
-from liger_kernel.ops.layer_norm import LigerLayerNormFunction
+from liger_kernel.ops import LigerLayerNormFunction
 
 
 class LigerLayerNorm(nn.Module):

liger_kernel/transformers/llama4_rope.py

@@ -5,7 +5,7 @@ Supports both text and vision RoPE variants with fused operations for optimal pe
 
 import torch
 
-from liger_kernel.ops.llama4_rope import LigerLlama4RopeFunction
+from liger_kernel.ops import LigerLlama4RopeFunction
 
 
 def liger_llama4_text_rotary_pos_emb(

liger_kernel/transformers/model/falcon_h1.py

@@ -4,12 +4,12 @@ from typing import Union
 
 import torch
 
-from transformers.modeling_outputs import CausalLMOutputWithPast
-
 if TYPE_CHECKING:
     from transformers.models.falcon_h1.modeling_falcon_h1 import FalconHybridMambaAttentionDynamicCache
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
 
 
 def lce_forward(
@@ -26,8 +26,9 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
     **kwargs,
-) -> Union[tuple, CausalLMOutputWithPast]:
+) -> Union[tuple, LigerCausalLMOutputWithPast]:
     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, ...,
@@ -54,6 +55,7 @@ def lce_forward(
     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(
@@ -77,6 +79,8 @@ def lce_forward(
     shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
+    token_accuracy = None
+
     # if in training mode, don't materialize logits
     if skip_logits and labels is None:
         raise ValueError("skip_logits is True, but labels and shift_labels are None")
@@ -85,8 +89,9 @@ def lce_forward(
         # By default, if in training mode, don't materialize logits
         skip_logits = self.training and labels is not None
 
+    # Compute loss
     if skip_logits:
-        loss = LigerForCausalLMLoss(
+        result = LigerForCausalLMLoss(
             hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
@@ -94,15 +99,24 @@ def lce_forward(
             hidden_size=self.config.hidden_size,
             **kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
     else:
         logits = self.lm_head(kept_hidden_states)
         if labels is not None or shift_labels is not None:
             loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
 
-    return CausalLMOutputWithPast(
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        output = ((loss,) + output) if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
+
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
         loss=loss,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
     )