liger-kernel-nightly 0.4.0.dev20241107052928__py3-none-any.whl → 0.6.3.dev20251121010306__py3-none-any.whl

liger_kernel/ops/tvd.py

@@ -0,0 +1,207 @@
+from typing import Literal
+from typing import Optional
+
+import torch
+import triton
+import triton.language as tl
+
+from liger_kernel.ops.utils import ensure_contiguous
+
+MAX_FUSED_SIZE = 65536 // 4
+
+REDUCTION_LITERAL = Literal["none", "sum", "mean", "batchmean"]
+
+_REDUCTION_MODE_NONE = tl.constexpr(0)
+_REDUCTION_MODE_SUM = tl.constexpr(1)
+_REDUCTION_MODE_MEAN = tl.constexpr(2)
+_REDUCTION_MODE_BATCHMEAN = tl.constexpr(3)
+
+_str_to_reduction_mode = {
+    "none": _REDUCTION_MODE_NONE.value,
+    "sum": _REDUCTION_MODE_SUM.value,
+    "mean": _REDUCTION_MODE_MEAN.value,
+    "batchmean": _REDUCTION_MODE_BATCHMEAN.value,
+}
+
+
+def get_num_warps(BLOCK_SIZE):
+    num_warps = 4
+    if BLOCK_SIZE >= 32768:
+        num_warps = 32
+    elif BLOCK_SIZE >= 8192:
+        num_warps = 16
+    elif BLOCK_SIZE >= 2048:
+        num_warps = 8
+
+    return num_warps
+
+
+@triton.jit
+def _tv_distance_kernel(
+    p_ptr,
+    p_stride,
+    q_ptr,
+    q_stride,
+    loss_ptr,
+    loss_stride,
+    grads_ptr,
+    grads_stride,
+    label_ptr,
+    ignore_index: tl.constexpr,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+    HAS_LABEL: tl.constexpr,
+    reduction: tl.constexpr = _REDUCTION_MODE_BATCHMEAN,
+):
+    pid = tl.program_id(0).to(tl.int64)
+    p_ptr += pid * p_stride
+    q_ptr += pid * q_stride
+    loss_ptr += pid * loss_stride
+    grads_ptr += pid * grads_stride
+    label_ptr += pid
+
+    base_offsets = tl.arange(0, BLOCK_SIZE)
+
+    if HAS_LABEL:
+        label = tl.load(label_ptr)
+        if label == ignore_index:
+            for i in range(0, n_cols, BLOCK_SIZE):
+                offsets = i + base_offsets
+                mask = offsets < n_cols
+                tl.store(grads_ptr + offsets, 0.0, mask=mask)
+                if reduction == _REDUCTION_MODE_NONE:
+                    tl.store(loss_ptr + offsets, 0.0, mask=mask)
+            return
+
+    loss_sum = 0.0
+    for i in range(0, n_cols, BLOCK_SIZE):
+        offsets = i + base_offsets
+        mask = offsets < n_cols
+
+        p = tl.load(p_ptr + offsets, mask=mask, other=0.0)
+        q = tl.load(q_ptr + offsets, mask=mask, other=0.0)
+
+        # TVD(P || Q) = 0.5 * |P - Q|
+        tv_loss = 0.5 * tl.abs(p - q)
+
+        grad_res = tl.where(p > q, 0.5, -0.5)
+
+        tl.store(grads_ptr + offsets, grad_res, mask=mask)
+
+        if reduction == _REDUCTION_MODE_NONE:
+            tl.store(loss_ptr + offsets, tv_loss, mask=mask)
+        else:
+            loss_sum += tl.sum(tv_loss, axis=0)
+
+    if reduction != _REDUCTION_MODE_NONE:
+        tl.store(loss_ptr, loss_sum)
+
+
+def tv_distance_forward_triton(p, q, shift_labels, reduction, ignore_index, has_label):
+    BT, V = p.shape
+
+    BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
+    num_warps = get_num_warps(BLOCK_SIZE)
+
+    grid = (BT,)
+
+    reduction = _str_to_reduction_mode[reduction]
+
+    out_size = (BT, V) if reduction == _REDUCTION_MODE_NONE.value else (BT,)
+    output_tensor = torch.zeros(out_size, device=p.device, dtype=torch.float32)
+    grads = torch.empty_like(p)
+
+    n_non_ignore = (shift_labels != ignore_index).sum().item() if has_label else BT
+
+    _tv_distance_kernel[grid](
+        p,
+        p.stride(0),
+        q,
+        q.stride(0),
+        output_tensor,
+        output_tensor.stride(0),
+        grads,
+        grads.stride(0),
+        shift_labels if has_label else torch.empty(1, device=p.device),
+        ignore_index,
+        V,
+        BLOCK_SIZE=BLOCK_SIZE,
+        HAS_LABEL=has_label,
+        num_warps=num_warps,
+        reduction=reduction,
+    )
+
+    if reduction == _REDUCTION_MODE_BATCHMEAN.value:
+        return output_tensor.sum() / n_non_ignore, grads / n_non_ignore
+    elif reduction == _REDUCTION_MODE_SUM.value:
+        return output_tensor.sum(dim=0), grads
+    elif reduction == _REDUCTION_MODE_MEAN.value:
+        return output_tensor.sum() / (n_non_ignore * V), grads / (n_non_ignore * V)
+    else:
+        return output_tensor, grads
+
+
+def tvd_backward_triton(grad_output, grads):
+    # If cross entropy is the last layer, grad_output is 1.0. Skip the mul then.
+    if torch.equal(grad_output, torch.tensor(1.0, device=grad_output.device)):
+        return grads
+
+    return grads * grad_output
+
+
+class LigerTVDLossFunction(torch.autograd.Function):
+    """
+    Class implementing the forward and backward pass for the Total Variation Distance Loss using Triton.
+    """
+
+    @staticmethod
+    @ensure_contiguous
+    def forward(
+        ctx,
+        p: torch.Tensor,
+        q: torch.Tensor,
+        shift_labels: Optional[torch.Tensor] = None,
+        reduction: REDUCTION_LITERAL = "batchmean",
+        ignore_index: int = -100,
+    ) -> torch.Tensor:
+        """A forward pass for the Total Variation Distance Loss.
+
+        Args:
+            ctx: Torch autograd context
+            p (torch.Tensor): A tensor of shape (BT, V) containing the first distribution.
+            q (torch.Tensor): A tensor of shape (BT, V) containing the second distribution.
+            shift_labels (Optional[torch.Tensor]): A tensor of shape (BT,) containing the labels.
+            reduction (REDUCTION_LITERAL, optional): The reduction method to be applied. Defaults to "batchmean".
+            ignore_index (int, optional): The index to ignore during loss calculation. Defaults to -100.
+
+        Returns:
+            torch.Tensor: The computed Total Variation Distance Loss.
+        """
+        has_label = False
+        if shift_labels is not None:
+            assert shift_labels.shape == (p.shape[0],), (
+                f"the shape of shift_labels must be (BT,). Got: {shift_labels.shape}"
+            )
+            shift_labels = shift_labels.contiguous()
+            has_label = True
+
+        loss, grads = tv_distance_forward_triton(p, q, shift_labels, reduction, ignore_index, has_label)
+        ctx.save_for_backward(grads)
+        return loss
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_output: torch.Tensor) -> torch.Tensor:
+        """A backward pass for the Total Variation Distance Loss.
+
+        Args:
+            ctx: Torch autograd context
+            grad_output (torch.Tensor): The gradient of the loss with respect to the output.
+
+        Returns:
+            tuple[torch.Tensor, None, None, None, None]: The gradient of the loss with respect to the inputs.
+        """
+        (grads,) = ctx.saved_tensors
+        grads = tvd_backward_triton(grad_output, grads)
+
+        return grads, None, None, None, None

liger_kernel/ops/utils.py

@@ -13,13 +13,17 @@ Modifications made by Yanning Chen, 2024.
 import functools
 import importlib
 import operator
+
 from typing import Callable
 
 import torch
 import triton
 import triton.language as tl
+
 from packaging.version import Version
 
+from liger_kernel.utils import infer_device
+
 
 def is_hip() -> bool:
     return torch.version.hip is not None
@@ -45,8 +49,7 @@ def calculate_settings(n):
     BLOCK_SIZE = triton.next_power_of_2(n)
     if BLOCK_SIZE > MAX_FUSED_SIZE:
         raise RuntimeError(
-            f"Cannot launch Triton kernel since n = {n} exceeds "
-            f"the recommended Triton blocksize = {MAX_FUSED_SIZE}."
+            f"Cannot launch Triton kernel since n = {n} exceeds the recommended Triton blocksize = {MAX_FUSED_SIZE}."
         )
 
     num_warps = 4
@@ -69,10 +72,11 @@ def compare_version(package: str, operator: Callable, target: str):
 
 
 def get_amp_custom_fwd_bwd() -> Callable:
+    device = infer_device()
     if compare_version("torch", operator.ge, "2.4.0"):
         return (
-            functools.partial(torch.amp.custom_fwd, device_type="cuda"),
-            functools.partial(torch.amp.custom_bwd, device_type="cuda"),
+            functools.partial(torch.amp.custom_fwd, device_type=device),
+            functools.partial(torch.amp.custom_bwd, device_type=device),
         )
     return torch.cuda.amp.custom_fwd, torch.cuda.amp.custom_bwd
 

liger_kernel/transformers/__init__.py

@@ -1,31 +1,206 @@
-from liger_kernel.transformers.auto_model import (  # noqa: F401
-    AutoLigerKernelForCausalLM,
-)
+import importlib
+
+from typing import TYPE_CHECKING
+
+# Always-safe imports (independent of 'transformers')
 from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss  # noqa: F401
-from liger_kernel.transformers.fused_linear_cross_entropy import (  # noqa: F401
-    LigerFusedLinearCrossEntropyLoss,
-)
+from liger_kernel.transformers.dyt import LigerDyT  # noqa: F401
+from liger_kernel.transformers.fused_add_rms_norm import LigerFusedAddRMSNorm  # noqa: F401
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss  # noqa: F401
 from liger_kernel.transformers.fused_linear_jsd import LigerFusedLinearJSD  # noqa: F401
 from liger_kernel.transformers.geglu import LigerGEGLUMLP  # noqa: F401
 from liger_kernel.transformers.jsd import LigerJSD  # noqa: F401
+from liger_kernel.transformers.kl_div import LigerKLDIVLoss  # noqa: F401
 from liger_kernel.transformers.layer_norm import LigerLayerNorm  # noqa: F401
-from liger_kernel.transformers.monkey_patch import (  # noqa: F401
-    _apply_liger_kernel,
-    _apply_liger_kernel_to_instance,
-    apply_liger_kernel_to_gemma,
-    apply_liger_kernel_to_gemma2,
-    apply_liger_kernel_to_llama,
-    apply_liger_kernel_to_mistral,
-    apply_liger_kernel_to_mixtral,
-    apply_liger_kernel_to_mllama,
-    apply_liger_kernel_to_phi3,
-    apply_liger_kernel_to_qwen2,
-    apply_liger_kernel_to_qwen2_vl,
-)
+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.swiglu import (  # noqa: F401
-    LigerBlockSparseTop2MLP,
-    LigerPhi3SwiGLUMLP,
-    LigerSwiGLUMLP,
-)
+from liger_kernel.transformers.softmax import LigerSoftmax  # noqa: F401
+from liger_kernel.transformers.sparsemax import LigerSparsemax  # noqa: F401
+from liger_kernel.transformers.swiglu import LigerBlockSparseTop2MLP  # noqa: F401
+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
+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_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
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma3  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma3_text  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_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_granite  # 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
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llava  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mistral  # noqa: F401
+    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_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
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_5_vl  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_vl  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3  # noqa: F401
+    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
+try:
+    import transformers  # noqa: F401
+
+    _TRANSFORMERS_AVAILABLE = True
+except ImportError:
+    _TRANSFORMERS_AVAILABLE = False
+
+
+def is_transformers_available() -> bool:
+    """
+    Returns True if the 'transformers' package is available.
+    Useful for conditional logic in downstream code.
+    """
+    return _TRANSFORMERS_AVAILABLE
+
+
+def __getattr__(name: str):
+    """
+    Handles lazy access to transformer-dependent attributes.
+    If 'transformers' is not installed, raises a user-friendly ImportError.
+    """
+    if not _TRANSFORMERS_AVAILABLE:
+        raise ImportError(
+            f"The attribute '{name}' requires the 'transformers' library, which is not installed.\n"
+            f"Please install it with `pip install transformers` to use this functionality."
+        )
+
+    if name == "AutoLigerKernelForCausalLM":
+        module = importlib.import_module("liger_kernel.transformers.auto_model")
+        return getattr(module, name)
+
+    monkey_patch_symbols = {
+        "_apply_liger_kernel",
+        "_apply_liger_kernel_to_instance",
+        "apply_liger_kernel_to_falcon_h1",
+        "apply_liger_kernel_to_gemma",
+        "apply_liger_kernel_to_gemma2",
+        "apply_liger_kernel_to_gemma3",
+        "apply_liger_kernel_to_gemma3_text",
+        "apply_liger_kernel_to_glm4",
+        "apply_liger_kernel_to_glm4v",
+        "apply_liger_kernel_to_glm4v_moe",
+        "apply_liger_kernel_to_granite",
+        "apply_liger_kernel_to_internvl",
+        "apply_liger_kernel_to_llama",
+        "apply_liger_kernel_to_llava",
+        "apply_liger_kernel_to_llama4",
+        "apply_liger_kernel_to_mistral",
+        "apply_liger_kernel_to_mixtral",
+        "apply_liger_kernel_to_mllama",
+        "apply_liger_kernel_to_olmo2",
+        "apply_liger_kernel_to_paligemma",
+        "apply_liger_kernel_to_phi3",
+        "apply_liger_kernel_to_qwen2",
+        "apply_liger_kernel_to_qwen2_5_vl",
+        "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",
+    }
+
+    if name in monkey_patch_symbols:
+        module = importlib.import_module("liger_kernel.transformers.monkey_patch")
+        return getattr(module, name)
+
+    raise AttributeError(f"module {__name__} has no attribute {name}")
+
+
+# Shared symbols in all environments
+__all__ = [
+    "is_transformers_available",
+    "LigerCrossEntropyLoss",
+    "LigerDyT",
+    "LigerFusedLinearCrossEntropyLoss",
+    "LigerFusedLinearJSD",
+    "LigerGEGLUMLP",
+    "LigerJSD",
+    "LigerLayerNorm",
+    "LigerFusedAddRMSNorm",
+    "LigerPolyNorm",
+    "LigerRMSNorm",
+    "liger_rotary_pos_emb",
+    "liger_llama4_text_rotary_pos_emb",
+    "liger_llama4_vision_rotary_pos_emb",
+    "LigerBlockSparseTop2MLP",
+    "LigerPhi3SwiGLUMLP",
+    "LigerQwen3MoeSwiGLUMLP",
+    "LigerSwiGLUMLP",
+    "LigerTiledGEGLUMLP",
+    "LigerTiledSwiGLUMLP",
+    "LigerTVDLoss",
+    "LigerKLDIVLoss",
+    "LigerMultiTokenAttention",
+    "LigerSoftmax",
+    "LigerSparsemax",
+]
+
+# Add transformer-dependent symbols only if available
+if _TRANSFORMERS_AVAILABLE:
+    __all__.extend(
+        [
+            "AutoLigerKernelForCausalLM",
+            "_apply_liger_kernel",
+            "_apply_liger_kernel_to_instance",
+            "apply_liger_kernel_to_falcon_h1",
+            "apply_liger_kernel_to_gemma",
+            "apply_liger_kernel_to_gemma2",
+            "apply_liger_kernel_to_gemma3",
+            "apply_liger_kernel_to_gemma3_text",
+            "apply_liger_kernel_to_glm4",
+            "apply_liger_kernel_to_glm4v",
+            "apply_liger_kernel_to_glm4v_moe",
+            "apply_liger_kernel_to_granite",
+            "apply_liger_kernel_to_internvl",
+            "apply_liger_kernel_to_llama",
+            "apply_liger_kernel_to_llava",
+            "apply_liger_kernel_to_llama4",
+            "apply_liger_kernel_to_mistral",
+            "apply_liger_kernel_to_mixtral",
+            "apply_liger_kernel_to_mllama",
+            "apply_liger_kernel_to_olmo2",
+            "apply_liger_kernel_to_paligemma",
+            "apply_liger_kernel_to_phi3",
+            "apply_liger_kernel_to_qwen2",
+            "apply_liger_kernel_to_qwen2_5_vl",
+            "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",
+        ]
+    )

liger_kernel/transformers/auto_model.py

@@ -1,11 +1,10 @@
 import inspect
 
-from transformers import AutoConfig, AutoModelForCausalLM
+from transformers import AutoConfig
+from transformers import AutoModelForCausalLM
 
-from liger_kernel.transformers.monkey_patch import (
-    MODEL_TYPE_TO_APPLY_LIGER_FN,
-    _apply_liger_kernel,
-)
+from liger_kernel.transformers.monkey_patch import MODEL_TYPE_TO_APPLY_LIGER_FN
+from liger_kernel.transformers.monkey_patch import _apply_liger_kernel
 
 
 def _get_model_config(model_dir, **model_init_kwargs):
@@ -34,12 +33,6 @@ class AutoLigerKernelForCausalLM(AutoModelForCausalLM):
         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
-        }
+        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
-        )
+        return super().from_pretrained(pretrained_model_name_or_path, *model_args, **applicable_kwargs)

liger_kernel/transformers/cross_entropy.py

@@ -1,43 +1,56 @@
-import torch.nn as nn
+from typing import Optional
+
+import torch
 
 from liger_kernel.ops.cross_entropy import LigerCrossEntropyFunction
+from liger_kernel.transformers.functional import CrossEntropyOutput
 
 
-class LigerCrossEntropyLoss(nn.Module):
+class LigerCrossEntropyLoss(torch.nn.Module):
     def __init__(
         self,
-        ignore_index=-100,
-        lse_square_scale=0.0,
-        label_smoothing=0.0,
-        reduction="mean",
-        return_z_loss=False,
+        weight: Optional[torch.FloatTensor] = None,
+        ignore_index: int = -100,
+        lse_square_scale: float = 0.0,
+        label_smoothing: float = 0.0,
+        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), (
+            f"label_smoothing must be between 0.0 and 1.0. Got: {label_smoothing}"
+        )
+        assert reduction in {
+            "mean",
+            "sum",
+            "none",
+        }, f"reduction must be one of 'mean', 'sum', or 'none'. Got: {reduction}"
+        assert softcap is None or softcap > 0, f"softcap must greater than 0.0 or None. Got: {softcap}"
+        self.weight = weight
         self.ignore_index = ignore_index
         self.lse_square_scale = lse_square_scale
         self.label_smoothing = label_smoothing
         self.reduction = reduction
+        self.softcap = softcap
         self.return_z_loss = return_z_loss
+        self.return_token_accuracy = return_token_accuracy
 
-        assert (self.label_smoothing >= 0) and (
-            self.label_smoothing <= 1
-        ), f"label_smoothing must be between 0.0 and 1.0. Got: {self.label_smoothing}"
-        assert self.reduction in {
-            "mean",
-            "sum",
-            "none",
-        }, f"reduction must be one of 'mean', 'sum', or 'none'. Got: {self.reduction}"
-
-    def forward(self, _input, target):
-        loss, z_loss = LigerCrossEntropyFunction.apply(
+    def forward(self, _input: torch.Tensor, target: torch.Tensor):
+        loss, z_loss, token_accuracy = LigerCrossEntropyFunction.apply(
             _input,
             target,
+            self.weight,
             self.ignore_index,
             self.lse_square_scale,
             self.label_smoothing,
             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

@@ -0,0 +1,22 @@
+import torch
+import torch.nn as nn
+
+from liger_kernel.ops.dyt import LigerDyTFunction
+
+
+class LigerDyT(nn.Module):
+    def __init__(self, hidden_size, beta=True, init_alpha=0.5):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.init_alpha = init_alpha
+        self.alpha = nn.Parameter(torch.ones(1) * init_alpha)
+        self.gamma = nn.Parameter(torch.ones(hidden_size))
+        self.beta = None
+        if beta:
+            self.beta = nn.Parameter(torch.zeros(hidden_size))
+
+    def forward(self, x):
+        return LigerDyTFunction.apply(x, self.alpha, self.gamma, self.beta)
+
+    def extra_repr(self):
+        return f"{self.hidden_size}, init_alpha={self.init_alpha}, beta={self.beta}"

liger_kernel/transformers/experimental/__init__.py

@@ -0,0 +1,5 @@
+from liger_kernel.transformers.experimental.embedding import LigerEmbedding  # noqa: F401
+
+__all__ = [
+    "LigerEmbedding",
+]

liger_kernel/transformers/experimental/embedding.py

@@ -7,9 +7,7 @@ from liger_kernel.ops.experimental.embedding import LigerEmbeddingFunction
 
 
 class LigerEmbedding(nn.Module):
-    def __init__(
-        self, num_embeddings, embedding_dim, padding_idx: Optional[int] = None
-    ):
+    def __init__(self, num_embeddings, embedding_dim, padding_idx: Optional[int] = None):
         super().__init__()
         self.num_embeddings = num_embeddings
         self.embedding_dim = embedding_dim