liger-kernel 0.5.5__py3-none-any.whl → 0.5.6__py3-none-any.whl

liger_kernel/chunked_loss/jsd_loss.py

@@ -19,15 +19,20 @@ class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
         student_log_probs = F.log_softmax(student_logits, dim=-1)
         teacher_log_probs = F.log_softmax(teacher_logits, dim=-1)
 
-        # Compute probabilities (only required for mean calculation)
-        mean_probs = beta * student_log_probs.exp() + (1 - beta) * teacher_log_probs.exp()
-        log_mean_probs = mean_probs.log()
+        if beta == 0:
+            jsd_loss = F.kl_div(student_log_probs, teacher_log_probs, reduction="sum", log_target=True)
+        elif beta == 1:
+            jsd_loss = F.kl_div(teacher_log_probs, student_log_probs, reduction="sum", log_target=True)
+        else:
+            # Compute probabilities (only required for mean calculation)
+            mean_probs = (1 - beta) * student_log_probs.exp() + beta * teacher_log_probs.exp()
+            log_mean_probs = mean_probs.log()
 
-        student_kl = F.kl_div(log_mean_probs, student_log_probs, reduction="sum", log_target=True)
-        teacher_kl = F.kl_div(log_mean_probs, teacher_log_probs, reduction="sum", log_target=True)
+            student_kl = F.kl_div(log_mean_probs, student_log_probs, reduction="sum", log_target=True)
+            teacher_kl = F.kl_div(log_mean_probs, teacher_log_probs, reduction="sum", log_target=True)
 
-        # JSD is the weighted average of the KL divergences
-        jsd_loss = beta * teacher_kl + (1 - beta) * student_kl
+            # JSD is the weighted average of the KL divergences
+            jsd_loss = beta * teacher_kl + (1 - beta) * student_kl
         return jsd_loss
 
     @classmethod

liger_kernel/ops/cross_entropy.py

@@ -9,6 +9,7 @@ import triton.language as tl
 from liger_kernel.ops.utils import compare_version
 from liger_kernel.ops.utils import element_mul_kernel
 from liger_kernel.ops.utils import is_hip
+from liger_kernel.utils import infer_device
 
 if compare_version("triton", operator.ge, "3.0.0"):
     try:
@@ -59,7 +60,7 @@ def liger_cross_entropy_kernel(
     z_loss_ptr: Pointer to tensor to store the z loss. No operation if RETURN_Z_LOSS is 0.
     loss_stride (int): The stride of the loss tensor.
     n_cols (int): The number of columns in the input tensor.
-    n_non_ignore (flaot): The number of non-ignored elements in the batch.
+    n_non_ignore (float): The number of non-ignored elements in the batch.
     sum_non_ignore_weight (float): The sum of non-ignored target's weights in the batch.
     weight_sum (float): The sum of weight tensor.
     ignore_index (int): The index to ignore in the target.
@@ -258,7 +259,7 @@ def liger_cross_entropy_kernel(
 # The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576 https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
 # However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
 # The optimal maximum block size depends on your hardware, your kernel, and your dtype
-MAX_FUSED_SIZE = 65536 // 2  # the best size we found by manually tuning
+MAX_FUSED_SIZE = 4096 if infer_device() == "xpu" else 65536 // 2  # the best size we found by manually tuning
 
 
 def cross_entropy_forward(

liger_kernel/ops/dyt.py

@@ -0,0 +1,225 @@
+import operator
+
+import torch
+import triton
+import triton.language as tl
+
+from liger_kernel.ops.utils import calculate_settings
+from liger_kernel.ops.utils import compare_version
+from liger_kernel.ops.utils import ensure_contiguous
+from liger_kernel.ops.utils import infer_device
+
+if compare_version("triton", operator.ge, "3.0.0"):
+    try:
+        # typical import path with dispatch available
+        from triton.language.extra.libdevice import tanh
+    except ModuleNotFoundError:
+        # for working with NGC containers
+        from triton.language.extra.cuda.libdevice import tanh
+else:
+    from triton.language.math import tanh
+
+
+@triton.jit
+def _dyt_fwd_kernel(
+    x_ptr,
+    x_row_stride,
+    alpha_ptr,
+    gamma_ptr,
+    beta_ptr,
+    y_ptr,
+    y_row_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    Reference:
+    https://arxiv.org/abs/2503.10622
+
+    Shapes:
+        - x: (BT, C)
+        - alpha: (1)
+        - gamma: (C)
+        - beta: (C)
+    """
+    row_idx = tl.program_id(0)
+    offsets = tl.arange(0, BLOCK_SIZE)
+    mask = offsets < n_cols
+
+    x_ptr += row_idx * x_row_stride
+    y_ptr += row_idx * y_row_stride
+
+    alpha = tl.load(alpha_ptr)
+    gamma = tl.load(gamma_ptr + offsets, mask=mask)
+    beta = tl.load(beta_ptr + offsets, mask=mask)
+    x = tl.load(x_ptr + offsets, mask=mask)
+    y = gamma * tanh((alpha * x).cast(tl.float32)) + beta
+    tl.store(y_ptr + offsets, y, mask=mask)
+
+
+@triton.jit
+def _dyt_bwd_kernel(
+    x_ptr,
+    x_row_stride,
+    dy_ptr,
+    dy_row_stride,
+    dx_ptr,
+    dx_row_stride,
+    alpha_ptr,
+    dalpha_ptr,
+    gamma_ptr,
+    dgamma_ptr,
+    dgamma_row_stride,
+    n_cols,
+    n_rows,
+    ROWS_PER_PROGRAM: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    Reference:
+    https://arxiv.org/abs/2503.10622
+
+    Shapes:
+        - x: (BT, C)
+        - alpha: (1)
+        - gamma: (C)
+        - dx: (BT, C)
+        - dy: (BT, C)
+        - dgamma: (sm_count, C)
+        - dalpha: (sm_count,)
+    """
+    # d(gamma * tanh(alpha * x) + beta) / dx
+    # = gamma * (1 - tanh^2(alpha * x)) * alpha
+    # d(gamma * tanh(alpha * x) + beta) / dalpha
+    # = gamma * (1 - tanh^2(alpha * x)) * x
+    # d(gamma * tanh(alpha * x) + beta) / dgamma
+    # = tanh(alpha * x)
+    # d(gamma * tanh(alpha * x)) / dbeta = 1
+    pid = tl.program_id(0)
+
+    row_start = pid * ROWS_PER_PROGRAM
+    row_end = min((pid + 1) * ROWS_PER_PROGRAM, n_rows)
+    offsets = tl.arange(0, BLOCK_SIZE)
+    mask = offsets < n_cols
+
+    dalpha = 0.0
+    dgamma = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+
+    x_ptr += row_start * x_row_stride
+    dx_ptr += row_start * dx_row_stride
+    dy_ptr += row_start * dy_row_stride
+    alpha = tl.load(alpha_ptr)
+    gamma = tl.load(gamma_ptr + offsets, mask=mask, other=0.0)
+
+    for _ in tl.range(row_start, row_end):
+        dy = tl.load(dy_ptr + offsets, mask=mask, other=0.0)
+        x = tl.load(x_ptr + offsets, mask=mask, other=0.0)
+        tanh_ax = tanh((alpha * x).cast(tl.float32))
+        sech2_ax = 1 - tanh_ax * tanh_ax
+
+        dx = dy * gamma * sech2_ax * alpha
+        dalpha += tl.sum(dy * gamma * sech2_ax * x)
+        dgamma += dy * tanh_ax
+        tl.store(dx_ptr + offsets, dx, mask=mask)
+
+        dy_ptr += dy_row_stride
+        x_ptr += x_row_stride
+        dx_ptr += dx_row_stride
+
+    tl.store(dgamma_ptr + pid * dgamma_row_stride + offsets, dgamma, mask=mask)
+    tl.store(dalpha_ptr + pid, dalpha)
+
+    pass
+
+
+def liger_dyt_fwd(x, alpha, gamma, beta):
+    shape = x.shape
+    dim = shape[-1]
+    x = x.view(-1, dim)
+    n_rows, n_cols = x.shape
+    y = torch.empty_like(x)
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    _dyt_fwd_kernel[(n_rows,)](
+        x_ptr=x,
+        alpha_ptr=alpha,
+        gamma_ptr=gamma,
+        beta_ptr=beta,
+        y_ptr=y,
+        x_row_stride=x.stride(0),
+        y_row_stride=y.stride(0),
+        n_cols=n_cols,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+    )
+    return y.view(*shape)
+
+
+def liger_dyt_bwd(dy, x, alpha, gamma):
+    shape = dy.shape
+    dtype = x.dtype
+    dim = shape[-1]
+    dy = dy.view(-1, dim)
+    x = x.view(-1, dim)
+    n_rows, n_cols = dy.shape
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    sm_count = 1
+    device = infer_device()
+    if device == "cuda":
+        sm_count = torch.cuda.get_device_properties(x.device).multi_processor_count
+    elif device == "xpu":
+        sm_count = torch.xpu.get_device_properties(x.device).gpu_subslice_count
+    if n_cols > BLOCK_SIZE:
+        raise RuntimeError(
+            f"Feature dimension {dim} exceeds maximum supported size of {BLOCK_SIZE}. Consider using a smaller feature dimension."
+        )
+
+    dx = torch.empty_like(x, dtype=torch.float32)
+    _dalpha = torch.empty((sm_count,), dtype=torch.float32, device=x.device)
+    _dgamma = torch.empty((sm_count, n_cols), dtype=torch.float32, device=x.device)
+
+    grid = (sm_count,)
+    rows_per_program = triton.cdiv(n_rows, sm_count)
+    _dyt_bwd_kernel[grid](
+        x_ptr=x,
+        x_row_stride=x.stride(0),
+        dy_ptr=dy,
+        dy_row_stride=dy.stride(0),
+        dx_ptr=dx,
+        dx_row_stride=dx.stride(0),
+        alpha_ptr=alpha,
+        dalpha_ptr=_dalpha,
+        gamma_ptr=gamma,
+        dgamma_ptr=_dgamma,
+        dgamma_row_stride=_dgamma.stride(0),
+        n_cols=n_cols,
+        n_rows=n_rows,
+        ROWS_PER_PROGRAM=rows_per_program,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+    )
+    dalpha = _dalpha.sum(dim=0, keepdim=True).to(dtype)
+    dgamma = _dgamma.sum(dim=0).to(dtype)
+    dbeta = dy.sum(dim=0).to(dtype)
+    return dx.view(*shape), dalpha, dgamma, dbeta
+
+
+class LigerDyTFunction(torch.autograd.Function):
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, x, alpha, gamma, beta):
+        y = liger_dyt_fwd(x, alpha, gamma, beta)
+        ctx.save_for_backward(x, alpha, gamma)
+        return y
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_output):
+        x, alpha, gamma = ctx.saved_tensors
+        dx, dalpha, dgamma, dbeta = liger_dyt_bwd(
+            grad_output,
+            x,
+            alpha,
+            gamma,
+        )
+
+        return (dx, dalpha, dgamma, dbeta)

liger_kernel/ops/fused_linear_jsd.py

@@ -8,11 +8,12 @@ from liger_kernel.ops.utils import amp_custom_bwd
 from liger_kernel.ops.utils import amp_custom_fwd
 from liger_kernel.ops.utils import element_mul_kernel
 from liger_kernel.ops.utils import is_hip
+from liger_kernel.utils import infer_device
 
 # The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576 https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
 # However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
 # The optimal maximum block size depends on your hardware, your kernel, and your dtype
-MAX_FUSED_SIZE = 65536 // 2
+MAX_FUSED_SIZE = 4096 if infer_device() == "xpu" else 65536 // 2
 
 
 def fused_linear_jsd_forward(

liger_kernel/ops/jsd.py

@@ -51,24 +51,43 @@ def _jsd_kernel(
         Y = tl.load(Y_ptr + offsets, mask=mask, other=float("-inf")).to(tl.float32)
 
         if beta == 0.0:  # forward KL
-            Y_prob = tl.exp(Y)
+            Y_max = tl.max(Y, axis=0)
+            Y_shifted = Y - Y_max
+            Y_prob = tl.exp(Y_shifted) * tl.exp(Y_max)  # Compensate for the shift
             loss = Y_prob * (Y - X)
             dX = -Y_prob
-        elif beta == 1.0:
-            X_prob = tl.exp(X)
+        elif beta == 1.0:  # reverse KL
+            X_max = tl.max(X, axis=0)
+            X_shifted = X - X_max
+            X_prob = tl.exp(X_shifted) * tl.exp(X_max)  # Compensate for the shift
             loss = X_prob * (X - Y)
             dX = loss + X_prob
         else:
-            Q = tl.exp(X)
-            P = tl.exp(Y)
-            M = beta * P + (1 - beta) * Q
-            log_M = tl.log(M)
+            max_val = tl.maximum(tl.max(X, axis=0), tl.max(Y, axis=0))
+            X_shifted = X - max_val
+            Y_shifted = Y - max_val
 
-            loss = beta * P * Y + (1 - beta) * Q * X - M * log_M
-            dX = (1 - beta) * Q * (X - log_M)
+            # Pre-compute exp(max_val) since it's used twice
+            exp_max = tl.exp(max_val)
+
+            # Compute exp terms with compensation
+            Q = tl.exp(X_shifted) * exp_max  # = exp(X)
+            P = tl.exp(Y_shifted) * exp_max  # = exp(Y)
+
+            # Pre-compute common terms
+            beta_P = beta * P
+            one_minus_beta_Q = (1 - beta) * Q
+            M = beta_P + one_minus_beta_Q
+            log_M = tl.log(M)  # No need to compensate as M is already in original scale
+
+            loss = beta_P * Y + one_minus_beta_Q * X - M * log_M
+            dX = one_minus_beta_Q * (X - log_M)
+
+        # Pre-compute scaling factor
+        scale = 1.0 / n_non_ignore
+        loss = loss * scale
+        dX = dX * scale
 
-        loss = loss / n_non_ignore
-        dX = dX / n_non_ignore
         tl.store(loss_ptr + offsets, loss, mask=mask)
         tl.store(dX_ptr + offsets, dX, mask=mask)
 

liger_kernel/ops/kl_div.py

@@ -185,9 +185,9 @@ class LigerKLDivLossFunction(torch.autograd.Function):
     Class implementing the forward and backward pass for the KL Divergence Loss using Triton, as defined by the following formula:
     ```python
     if log_target:
-        loss = target * (target.log() - input)
-    else:
         loss = target.exp() * (target - input)
+    else:
+        loss = target * (target.log() - input)
     ```,
     then the loss is reduced according to the `reduction` parameter.
     as defined in the PyTorch documentation: https://pytorch.org/docs/stable/generated/torch.nn.KLDivLoss.html

liger_kernel/transformers/__init__.py

@@ -1,5 +1,6 @@
 from liger_kernel.transformers.auto_model import AutoLigerKernelForCausalLM  # noqa: F401
 from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss  # noqa: F401
+from liger_kernel.transformers.dyt import LigerDyT  # 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
@@ -11,10 +12,12 @@ from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma
 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_granite  # 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_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

liger_kernel/transformers/dyt.py

@@ -0,0 +1,20 @@
+import torch
+import torch.nn as nn
+
+from liger_kernel.ops.dyt import LigerDyTFunction
+
+
+class LigerDyT(nn.Module):
+    def __init__(self, hidden_size, 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 = 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}"

liger_kernel/transformers/functional.py

@@ -1,6 +1,7 @@
 from typing import Optional
 
 from liger_kernel.ops.cross_entropy import LigerCrossEntropyFunction
+from liger_kernel.ops.dyt import LigerDyTFunction
 from liger_kernel.ops.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyFunction
 from liger_kernel.ops.fused_linear_jsd import LigerFusedLinearJSDFunction
 from liger_kernel.ops.geglu import LigerGELUMulFunction
@@ -192,3 +193,7 @@ def liger_rope(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
 
 def liger_swiglu(a, b):
     return LigerSiLUMulFunction.apply(a, b)
+
+
+def liger_dyt(x, alpha, gamma, beta):
+    return LigerDyTFunction.apply(x, alpha, gamma, beta)

liger_kernel/transformers/model/gemma.py

@@ -14,6 +14,7 @@ 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
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
 @add_start_docstrings_to_model_forward(GEMMA_INPUTS_DOCSTRING)
@@ -200,22 +201,13 @@ def lce_forward(
     loss = None
     # if in training mode, don't materialize logits
     if self.training and (labels is not None):
-        # We do the same thing as ForCausalLMLoss but using Liger FLCE
-
-        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)
-
-        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
-        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
-
-        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
-        if reduction == "sum":
-            loss /= loss_kwargs["num_items_in_batch"]
-
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            hidden_size=self.config.hidden_size,
+            **loss_kwargs,
+        )
     else:  # if in inference mode materialize logits
         logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
         if labels is not None:

liger_kernel/transformers/model/gemma2.py

@@ -15,6 +15,7 @@ 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
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 logger = logging.getLogger(__name__)
 
@@ -212,25 +213,15 @@ def lce_forward(
     loss = None
     # if in training mode, don't materialize logits
     if self.training and (labels is not None):
-        # We do the same thing as ForCausalLMLoss but using Liger FLCE
-
-        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)
-
-        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
-        lce = LigerFusedLinearCrossEntropyLoss(
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            hidden_size=self.config.hidden_size,
             softcap=self.config.final_logit_softcapping,
-            reduction=reduction,
+            **loss_kwargs,
         )
 
-        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
-        if reduction == "sum":
-            loss /= loss_kwargs["num_items_in_batch"]
-
     else:  # if in inference mode materialize logits
         logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
         if self.config.final_logit_softcapping is not None:

liger_kernel/transformers/model/llama.py

@@ -15,6 +15,7 @@ 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
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 if TYPE_CHECKING:
     from transformers.cache_utils import Cache
@@ -212,21 +213,13 @@ def lce_forward(
     loss = None
     # if in training mode, don't materialize logits
     if self.training and (labels is not None):
-        # We do the same thing as ForCausalLMLoss but using Liger FLCE
-
-        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)
-
-        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
-        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
-
-        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
-        if reduction == "sum":
-            loss /= loss_kwargs["num_items_in_batch"]
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            hidden_size=self.config.hidden_size,
+            **loss_kwargs,
+        )
 
     else:  # if in inference mode materialize logits
         logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])