liger-kernel 0.6.3__py3-none-any.whl → 0.6.5__py3-none-any.whl

liger_kernel/ops/tiled_mlp.py

@@ -0,0 +1,136 @@
+import math
+
+from typing import Callable
+from typing import List
+from typing import Optional
+
+import torch
+
+from liger_kernel.ops.utils import ensure_contiguous
+
+
+class LigerTiledMLPFunction(torch.autograd.Function):
+    """
+    Based on DeepSpeed's TiledMLP:
+    https://github.com/deepspeedai/DeepSpeed/blob/v0.18.2/deepspeed/runtime/sequence_parallel/ulysses_sp.py#L838
+
+    Perform a tiled MLP computation to massively reduce memory usage needed to compute MLP
+    when using very long sequence lengths.
+
+    This module re-computes `forward` in the `backward`. So the `forward` occurs twice each iteration.
+    And if you're using activation checkpointing it then occurs thrice.
+
+    Args:
+        fn: the function to call on sharded inputs (e.g., mlp.forward)
+        mlp_module: the MLP nn.Module object
+        x: the input to MLP.forward (hidden_states)
+        shards: how many shards to use
+        compute_params: a list of weights engaged in the compute
+
+    Returns:
+        the computed hidden_states
+    """
+
+    @staticmethod
+    @ensure_contiguous
+    def forward(
+        ctx,
+        fn: Callable,
+        mlp_module: torch.nn.Module,
+        x: torch.Tensor,
+        shards: int,
+        compute_params: Optional[List[torch.nn.Parameter]] = None,
+    ) -> torch.Tensor:
+        ctx.fn = fn
+        ctx.mlp_module = mlp_module
+        ctx.shards = shards
+        ctx.save_for_backward(x)
+
+        # x.shape could be [bs, seqlen, hidden_size] or [seqlen, hidden_size] (moe experts)
+        x_shards = list(torch.chunk(x, chunks=shards, dim=-2))
+        with torch.no_grad():
+            output_shards = [fn(mlp_module, x_shard) for x_shard in x_shards]
+        output_unsharded = torch.cat(output_shards, dim=-2)
+
+        return output_unsharded
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, *grads) -> tuple:
+        fn = ctx.fn
+        (x,) = ctx.saved_tensors
+        mlp_module = ctx.mlp_module
+        shards = ctx.shards
+
+        x_requires_grad = x.requires_grad
+        x = x.detach()
+        # detach() unsets x.requires_grad, so restore it
+        x.requires_grad_(x_requires_grad)
+
+        # x.shape could be [bs, seqlen, hidden_size] or [seqlen, hidden_size] (moe experts)
+        hidden_size = x.shape[-1]
+        x_shape_orig = x.shape
+
+        # flatten bs+seqlen to avoid having stride issues when narrowing into seqlen w/ bs>1
+        x = x.view(-1, hidden_size)
+        incoming_grad = grads[0].view(-1, hidden_size)
+        x_grad = torch.zeros_like(x)
+
+        x_shards = list(torch.chunk(x, chunks=shards, dim=0))
+
+        for i, x_shard in enumerate(x_shards):
+            x_shard.requires_grad_(x_requires_grad)
+
+            # if seqlen is not exactly divisible by shards the last step will be shorter than shard_step
+            shard_step = x_shards[i].shape[0]
+            shard_offset = i * x_shards[0].shape[0]
+
+            x_shard.grad = x_grad.narrow(0, shard_offset, shard_step).view_as(x_shard)
+            incoming_grad_shard = incoming_grad.narrow(0, shard_offset, shard_step).view_as(x_shard)
+
+            with torch.enable_grad():
+                output = fn(mlp_module, x_shard)
+            torch.autograd.backward(output, incoming_grad_shard)
+
+        # unflatten
+        x_grad = x_grad.view(x_shape_orig)
+
+        return (None, None, x_grad, None, None)
+
+
+def apply_tiled_mlp(
+    fn: Callable,
+    mlp_module: torch.nn.Module,
+    x: torch.Tensor,
+    num_shards: Optional[int] = None,
+    compute_params: Optional[List[torch.nn.Parameter]] = None,
+) -> torch.Tensor:
+    """
+    Apply tiled MLP computation for memory efficiency.
+
+    Args:
+        fn: the function to call on sharded inputs (e.g., lambda module, x: module(x))
+        mlp_module: the MLP nn.Module object
+        x: the input tensor with shape [bs, seqlen, hidden_size] or [seqlen, hidden_size]
+        num_shards: number of shards to use. If None, automatically calculated as ceil(seqlen / hidden_size)
+        compute_params: list of parameters for DeepSpeed ZeRO optimization
+
+    Returns:
+        output tensor with the same shape as input
+    """
+    if num_shards is None:
+        # x.shape could be [bs, seqlen, hidden_size] or [seqlen, hidden_size]
+        hidden_size = x.shape[-1]
+        seqlen = x.shape[-2]
+        num_shards = math.ceil(seqlen / hidden_size)
+
+    # Ensure num_shards is at least 1
+    num_shards = max(1, num_shards)
+
+    return LigerTiledMLPFunction.apply(
+        fn,
+        mlp_module,
+        x,
+        num_shards,
+        compute_params,
+    )

liger_kernel/ops/utils.py

@@ -78,6 +78,8 @@ def get_amp_custom_fwd_bwd() -> Callable:
             functools.partial(torch.amp.custom_fwd, device_type=device),
             functools.partial(torch.amp.custom_bwd, device_type=device),
         )
+    if hasattr(torch, "npu") and getattr(torch.npu, "amp", None) is not None:
+        return torch.npu.amp.custom_fwd, torch.npu.amp.custom_bwd
     return torch.cuda.amp.custom_fwd, torch.cuda.amp.custom_bwd
 
 
@@ -125,3 +127,26 @@ def element_mul_kernel(
         X_offsets = i + tl.arange(0, BLOCK_SIZE)
         X_block = tl.load(X_ptr + X_offsets, mask=X_offsets < n_cols)
         tl.store(X_ptr + X_offsets, X_block * grad_output, mask=X_offsets < n_cols)
+
+
+def get_npu_core_count(default: int = 20) -> int:
+    """Return NPU vector core count.
+    Fallback to `default` if Triton runtime or NPU device is unavailable.
+    """
+    try:
+        utils = triton.runtime.driver.active.utils
+        props = utils.get_device_properties(0)
+        return int(props.get("num_vectorcore", default))
+    except Exception:
+        return default
+
+
+def set_large_grf_mode(kernel_args: dict):
+    """Set large GRF mode for XPU devices."""
+    # On XPU triton installed along with pytorch-xpu will be called `pytorch-triton-xpu`,
+    # triton XPU installed from source will be called `triton`.
+    if compare_version("pytorch-triton-xpu", operator.ge, "3.6.0") or compare_version("triton", operator.ge, "3.6.0"):
+        kernel_args["grf_mode"] = "256"
+    else:
+        # API was changed in https://github.com/intel/intel-xpu-backend-for-triton/pull/5430
+        kernel_args["grf_mode"] = "large"

liger_kernel/transformers/__init__.py

@@ -24,6 +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
@@ -31,6 +33,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.auto_model import AutoLigerKernelForCausalLM  # noqa: F401
     from liger_kernel.transformers.monkey_patch import _apply_liger_kernel  # noqa: F401
     from liger_kernel.transformers.monkey_patch import _apply_liger_kernel_to_instance  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_exaone4  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_falcon_h1  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma2  # noqa: F401
@@ -39,7 +42,10 @@ 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
@@ -48,6 +54,7 @@ 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
@@ -56,6 +63,8 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_moe  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_next  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_vl  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_vl_moe  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_smollm3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_smolvlm  # noqa: F401
 
@@ -103,6 +112,7 @@ 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",
@@ -112,6 +122,7 @@ 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",
@@ -120,8 +131,13 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen3",
         "apply_liger_kernel_to_qwen3_moe",
         "apply_liger_kernel_to_qwen3_next",
+        "apply_liger_kernel_to_qwen3_vl",
+        "apply_liger_kernel_to_qwen3_vl_moe",
         "apply_liger_kernel_to_smollm3",
         "apply_liger_kernel_to_smolvlm",
+        "apply_liger_kernel_to_hunyuan_v1_dense",
+        "apply_liger_kernel_to_hunyuan_v1_moe",
+        "apply_liger_kernel_to_exaone4",
     }
 
     if name in monkey_patch_symbols:
@@ -151,6 +167,8 @@ __all__ = [
     "LigerPhi3SwiGLUMLP",
     "LigerQwen3MoeSwiGLUMLP",
     "LigerSwiGLUMLP",
+    "LigerTiledGEGLUMLP",
+    "LigerTiledSwiGLUMLP",
     "LigerTVDLoss",
     "LigerKLDIVLoss",
     "LigerMultiTokenAttention",
@@ -173,6 +191,7 @@ 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",
@@ -182,6 +201,7 @@ 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",
@@ -190,7 +210,12 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_qwen3",
             "apply_liger_kernel_to_qwen3_moe",
             "apply_liger_kernel_to_qwen3_next",
+            "apply_liger_kernel_to_qwen3_vl",
+            "apply_liger_kernel_to_qwen3_vl_moe",
             "apply_liger_kernel_to_smollm3",
             "apply_liger_kernel_to_smolvlm",
+            "apply_liger_kernel_to_hunyuan_v1_dense",
+            "apply_liger_kernel_to_hunyuan_v1_moe",
+            "apply_liger_kernel_to_exaone4",
         ]
     )

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

@@ -1,25 +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.poly_norm import LigerPolyNormFunction
-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
@@ -36,8 +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,
@@ -47,10 +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(
@@ -67,8 +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,
@@ -82,10 +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(

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):