liger-kernel 0.5.4__py3-none-any.whl → 0.5.6__py3-none-any.whl

liger_kernel/chunked_loss/kto_loss.py

@@ -7,10 +7,10 @@ from liger_kernel.chunked_loss.fused_linear_unpaired_preference import LigerFuse
 class LigerFusedLinearKTOFunction(LigerFusedLinearUnpairedPreferenceBase):
     @staticmethod
     def preference_loss_fn(
-        average_log_prob_chunk,
+        log_prob_chunk,
         preference_labels_chunk,
         full_target,
-        ref_average_log_prob_chunk=None,
+        ref_log_prob_chunk=None,
         beta=0.1,
         kl=None,
     ):
@@ -43,30 +43,34 @@ class LigerFusedLinearKTOFunction(LigerFusedLinearUnpairedPreferenceBase):
         3. Maintain reasonable distance from the reference model
 
         Args:
-            average_log_prob_chunk: Log probabilities for the chunk (batch_size,)
+            log_prob_chunk: Log probabilities for the chunk (batch_size,)
             preference_labels_chunk: Preference labels for the chunk (batch_size,)
             full_target: Non chunked full target tensor
-            ref_average_log_prob_chunk: Reference log probs for the chunk (batch_size,)
+            ref_log_prob_chunk: Reference log probs for the chunk (batch_size,)
             beta: Weight for the KTO loss
             kl: KL divergence between the policy model and the reference model for the chosen responses. Shape: (batch_size,)
         Returns:
             - loss: The KTO loss value
         """
-        if ref_average_log_prob_chunk is not None:
-            logratios_chunk = average_log_prob_chunk - ref_average_log_prob_chunk
+        if ref_log_prob_chunk is not None:
+            logratios_chunk = log_prob_chunk - ref_log_prob_chunk
         else:
-            logratios_chunk = average_log_prob_chunk
-
+            logratios_chunk = log_prob_chunk
         multiplier_chunk = torch.where(preference_labels_chunk, 1, -1)
         if kl is not None:
             losses = 1 - F.sigmoid(beta * (logratios_chunk - kl) * multiplier_chunk)
         else:
             losses = 1 - F.sigmoid(beta * logratios_chunk * multiplier_chunk)
 
-        return losses.sum() / (full_target.shape[0])
+        rewards = beta * logratios_chunk
+        chosen_rewards_sum = (rewards * preference_labels_chunk.unsqueeze(1)).sum()
+        rejected_rewards_sum = (rewards * (~preference_labels_chunk).unsqueeze(1)).sum()
 
-    @staticmethod
+        return losses.sum() / (full_target.shape[0]), chosen_rewards_sum, rejected_rewards_sum
+
+    @classmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
@@ -81,15 +85,38 @@ class LigerFusedLinearKTOFunction(LigerFusedLinearUnpairedPreferenceBase):
         beta=0.1,
         compiled=True,
         use_ref_model=True,
+        average_log_prob=False,
+        chunk_size=1,
     ):
-        return LigerFusedLinearUnpairedPreferenceBase.forward(
+        """
+        Fused linear layer with KTO loss.
+        Args:
+            _input (torch.Tensor): Input tensor. Shape: (batch_size * seq_len, hidden_size)
+            weight (torch.Tensor): Weight tensor. Shape: (vocab_size, hidden_size)
+            target (torch.LongTensor): Target tensor. Shape: (batch_size * seq_len,)
+            preference_labels (torch.Tensor): Preference labels tensor. Shape: (batch_size,)
+            bias (torch.Tensor, optional): Bias tensor. Shape: (vocab_size,)
+            ref_input (torch.Tensor, optional): Reference model input tensor. Shape: (batch_size * seq_len, hidden_size)
+            ref_weight (torch.Tensor, optional): Reference model weight tensor. Shape: (vocab_size, hidden_size)
+            ref_bias (torch.Tensor, optional): Reference model bias tensor. Shape: (vocab_size,)
+            kl (torch.Tensor, optional): KL divergence tensor. Shape: (batch_size,)
+            ignore_index (int): Index to ignore in loss computation
+            beta (float): Temperature parameter for the KTO loss
+            compiled (bool): Whether to use torch compile
+            use_ref_model (bool): Whether to use a reference model
+            average_log_prob (bool): Whether to average the log probability per non-masked token
+            chunk_size (int): Size of chunks for processing
+        Returns:
+            torch.Tensor: Computed loss
+        """
+        return super().forward(
+            cls=cls,
             ctx=ctx,
             _input=_input,
             weight=weight,
             target=target,
             preference_labels=preference_labels,
             bias=bias,
-            loss_fn=LigerFusedLinearKTOFunction.preference_loss_fn,
             ignore_index=ignore_index,
             beta=beta,
             compiled=compiled,
@@ -97,7 +124,9 @@ class LigerFusedLinearKTOFunction(LigerFusedLinearUnpairedPreferenceBase):
             ref_input=ref_input,
             ref_weight=ref_weight,
             ref_bias=ref_bias,
+            average_log_prob=average_log_prob,
             kl=kl,
+            chunk_size=chunk_size,
         )
 
     @staticmethod
@@ -115,6 +144,7 @@ class LigerFusedLinearKTOFunction(LigerFusedLinearUnpairedPreferenceBase):
             None,
             None,
             None,
+            None,
         )
 
 
@@ -129,6 +159,8 @@ class LigerFusedLinearKTOLoss(torch.nn.Module):
         beta: float = 0.1,
         compiled: bool = True,
         use_ref_model: bool = False,
+        average_log_prob: bool = False,
+        chunk_size: int = 1,
     ):
         """
         Args:
@@ -136,12 +168,16 @@ class LigerFusedLinearKTOLoss(torch.nn.Module):
             beta (float): Temperature parameter for the KTO loss
             compiled (bool): Whether to use compiled operations
             use_ref_model (bool): Whether to use a reference model for the DPO loss.
+            average_log_prob (bool): Whether to average the log probability per non-masked token
+            chunk_size (int): Size of chunks for processing
         """
         super().__init__()
         self.ignore_index = ignore_index
         self.beta = beta
         self.compiled = compiled
         self.use_ref_model = use_ref_model
+        self.average_log_prob = average_log_prob
+        self.chunk_size = chunk_size
 
     def forward(
         self,
@@ -169,4 +205,6 @@ class LigerFusedLinearKTOLoss(torch.nn.Module):
             self.beta,
             self.compiled,
             self.use_ref_model,
+            self.average_log_prob,
+            self.chunk_size,
         )

liger_kernel/chunked_loss/orpo_loss.py

@@ -42,8 +42,9 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
 
         return loss, chosen_rewards, rejected_rewards, log_odds_ratio, log_odds_chosen
 
-    @staticmethod
+    @classmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
@@ -54,25 +55,43 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
         compute_nll_loss=True,
         nll_target=None,
         compiled=True,
+        chunk_size=1,
     ):
-        return LigerFusedLinearPreferenceBase.forward(
+        """
+        Fused linear layer with ORPO loss.
+        Args:
+            _input (torch.Tensor): Input tensor. Shape: (batch_size * seq_len, hidden_size)
+            weight (torch.Tensor): Weight tensor. Shape: (vocab_size, hidden_size)
+            target (torch.LongTensor): Target tensor. Shape: (batch_size * seq_len,)
+            bias (torch.Tensor, optional): Bias tensor. Shape: (vocab_size,)
+            ignore_index (int): Index to ignore in loss computation
+            beta (float): Weight for the odds ratio loss
+            compute_nll_loss (bool): Whether to compute the NLL loss
+            nll_target (torch.LongTensor, optional): Target tensor for NLL loss. Shape: (batch_size * seq_len,)
+            compiled (bool): Whether to use torch compile
+            chunk_size (int): Size of chunks for processing
+        Returns:
+            torch.Tensor: Computed loss
+        """
+        return super().forward(
+            cls=cls,
             ctx=ctx,
             _input=_input,
             weight=weight,
             target=target,
             bias=bias,
-            loss_fn=LigerFusedLinearORPOFunction.preference_loss_fn,
             ignore_index=ignore_index,
             beta=beta,
             compute_nll_loss=compute_nll_loss,
             nll_target=nll_target,
             compiled=compiled,
+            chunk_size=chunk_size,
         )
 
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None, None
+        return *grads, None, None, None, None, None, None
 
 
 class LigerFusedLinearORPOLoss(torch.nn.Module):
@@ -86,19 +105,31 @@ class LigerFusedLinearORPOLoss(torch.nn.Module):
         beta: float = 0.1,
         compute_nll_loss: bool = True,
         compiled: bool = True,
+        chunk_size: int = 1,
     ):
         """
         Args:
             ignore_index (int): Index to ignore in the loss.
             beta (float): Weight for the odds ratio loss.
+            compute_nll_loss (bool): Whether to compute the NLL loss.
+            compiled (bool): Whether to use the torch compiled kernel.
+            chunk_size (int): Size of chunks for processing.
         """
         super().__init__()
         self.ignore_index = ignore_index
         self.beta = beta
         self.compute_nll_loss = compute_nll_loss
         self.compiled = compiled
+        self.chunk_size = chunk_size
 
-    def forward(self, lin_weight, _input, target, bias=None, nll_target=None):
+    def forward(
+        self,
+        lin_weight,
+        _input,
+        target,
+        bias=None,
+        nll_target=None,
+    ):
         return LigerFusedLinearORPOFunction.apply(
             _input,
             lin_weight,
@@ -109,4 +140,5 @@ class LigerFusedLinearORPOLoss(torch.nn.Module):
             self.compute_nll_loss,
             nll_target,
             self.compiled,
+            self.chunk_size,
         )

liger_kernel/chunked_loss/simpo_loss.py

@@ -47,8 +47,9 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
 
         return loss, chosen_rewards, rejected_rewards
 
-    @staticmethod
+    @classmethod
     def forward(
+        cls,
         ctx,
         _input,
         weight,
@@ -61,27 +62,47 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
         compute_nll_loss=False,
         compiled=True,
         gamma=0.5,
+        chunk_size=1,
     ):
-        return LigerFusedLinearPreferenceBase.forward(
-            ctx,
-            _input,
-            weight,
-            target,
-            bias,
-            loss_fn=LigerFusedLinearSimPOFunction.preference_loss_fn,
-            compute_nll_loss=compute_nll_loss,
+        """
+        Fused linear layer with SimPO loss.
+        Args:
+            _input (torch.Tensor): Input tensor. Shape: (batch_size * seq_len, hidden_size)
+            weight (torch.Tensor): Weight tensor. Shape: (vocab_size, hidden_size)
+            target (torch.LongTensor): Target tensor. Shape: (batch_size * seq_len,)
+            bias (torch.Tensor, optional): Bias tensor. Shape: (vocab_size,)
+            ignore_index (int): Index to ignore in loss computation
+            beta (float): Weight for the odds ratio loss
+            alpha (float): Weight for the alpha parameter
+            label_smoothing (float): Label smoothing factor
+            compute_nll_loss (bool): Whether to compute the NLL loss
+            compiled (bool): Whether to use torch compile
+            gamma (float): Weight for the gamma parameter
+            chunk_size (int): Size of chunks for processing
+        Returns:
+            torch.Tensor: Computed loss
+        """
+        return super().forward(
+            cls=cls,
+            ctx=ctx,
+            _input=_input,
+            weight=weight,
+            target=target,
+            bias=bias,
             ignore_index=ignore_index,
             alpha=alpha,
             beta=beta,
             label_smoothing=label_smoothing,
+            compute_nll_loss=compute_nll_loss,
             compiled=compiled,
             gamma=gamma,
+            chunk_size=chunk_size,
         )
 
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None, None, None, None
+        return *grads, None, None, None, None, None, None, None, None
 
 
 class LigerFusedLinearSimPOLoss(torch.nn.Module):
@@ -98,11 +119,18 @@ class LigerFusedLinearSimPOLoss(torch.nn.Module):
         compute_nll_loss: bool = True,
         compiled: bool = True,
         gamma: float = 0.5,
+        chunk_size: int = 1,
     ):
         """
         Args:
             ignore_index (int): Index to ignore in the loss.
             beta (float): Weight for the odds ratio loss.
+            alpha (float): Weight for the alpha parameter.
+            label_smoothing (float): Label smoothing factor.
+            compute_nll_loss (bool): Whether to compute the NLL loss.
+            compiled (bool): Whether to use the torch compiled kernel.
+            gamma (float): Weight for the gamma parameter.
+            chunk_size (int): Size of chunks for processing.
         """
         super().__init__()
         self.ignore_index = ignore_index
@@ -112,8 +140,15 @@ class LigerFusedLinearSimPOLoss(torch.nn.Module):
         self.compute_nll_loss = compute_nll_loss
         self.compiled = compiled
         self.gamma = gamma
+        self.chunk_size = chunk_size
 
-    def forward(self, lin_weight, _input, target, bias=None):
+    def forward(
+        self,
+        lin_weight,
+        _input,
+        target,
+        bias=None,
+    ):
         return LigerFusedLinearSimPOFunction.apply(
             _input,
             lin_weight,
@@ -126,4 +161,5 @@ class LigerFusedLinearSimPOLoss(torch.nn.Module):
             self.compute_nll_loss,
             self.compiled,
             self.gamma,
+            self.chunk_size,
         )

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(
@@ -285,6 +286,10 @@ def cross_entropy_forward(
 
     target_mask = target != ignore_index
     n_non_ignore = target_mask.sum().item()
+    assert (target * target_mask).max() < _input.shape[-1], (
+        f"Target {target.max()} is out of bounds. Expected < {_input.shape[-1]}"
+    )
+    assert (target * target_mask).min() >= 0, f"Target {target.min()} is out of bounds. Expected >= 0"
     sum_non_ignore_weight = n_non_ignore
     weight_sum = 0.0
     if weight is not None:

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)