liger-kernel-nightly 0.5.6.dev20250403190551__py3-none-any.whl → 0.6.4.dev20251212103629__py3-none-any.whl

liger_kernel/ops/cross_entropy.py

@@ -10,8 +10,9 @@ 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
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         # typical import path with dispatch available
         from triton.language.extra.libdevice import tanh
@@ -32,6 +33,8 @@ def liger_cross_entropy_kernel(
     loss_ptr,
     z_loss_ptr,
     loss_stride,
+    token_accuracy_ptr,
+    token_accuracy_stride,
     n_cols,
     n_non_ignore,
     sum_non_ignore_weight,
@@ -42,9 +45,11 @@ def liger_cross_entropy_kernel(
     reduction: tl.constexpr,  # set it as constexpr since reduction is always known at compile time
     softcap,
     RETURN_Z_LOSS: tl.constexpr,
+    RETURN_TOKEN_ACCURACY: tl.constexpr,
     BLOCK_SIZE: tl.constexpr,
     HAS_WEIGHT: tl.constexpr,
     HAS_SOFTCAPPING: tl.constexpr,
+    HAS_GRADIENTS: tl.constexpr,
 ):
     """
     This kernel computes both cross entropy loss and the gradient of the input.
@@ -59,6 +64,8 @@ def liger_cross_entropy_kernel(
     loss_ptr: Pointer to tensor to store the loss.
     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.
+    token_accuracy_ptr: Pointer to tensor to store the per-token accuracy. No operation if RETURN_TOKEN_ACCURACY is 0.
+    token_accuracy_stride (int): The stride of the token accuracy tensor.
     n_cols (int): The number of columns in the input tensor.
     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.
@@ -68,10 +75,12 @@ def liger_cross_entropy_kernel(
     lse_square_scale (float): The scaler of (logsumexp(_input)) ^ 2 adding to the loss for the stability of training.
     reduction (str): The string for the reduction to apply
     softcap (float): The upper threshold for scaling logits to the range (-softcap, +softcap).
-    RETURN_Z_LOSS (int): The boolean value to decide whether storing z loss to z_loss_ptr or not. It must be 0 or 1.
+    RETURN_Z_LOSS (int): The boolean value to decide whether to store z loss to z_loss_ptr or not. It must be 0 or 1.
+    RETURN_TOKEN_ACCURACY (int): The boolean value to decide whether to store per-token accuracy to token_accuracy_ptr or not. It must be 0 or 1.
     BLOCK_SIZE (int): The block size for Triton operations.
     HAS_WEIGHT (bool): The boolean value to determine whether assigning weight to each of the classes.
     HAS_SOFTCAPPING (bool): The boolean value to determine whether applying soft-capping or not.
+    HAS_GRADIENTS (bool): The boolean value to determine whether calculating gradients in forward pass.
     """
 
     # https://github.com/triton-lang/triton/issues/1058
@@ -90,11 +99,17 @@ def liger_cross_entropy_kernel(
         for i in range(0, n_cols, BLOCK_SIZE):
             X_offsets = i + tl.arange(0, BLOCK_SIZE)
             tl.store(X_ptr + X_offsets, 0.0, mask=X_offsets < n_cols)
+        # For ignored tokens, set token accuracy to 0
+        if RETURN_TOKEN_ACCURACY:
+            token_accuracy_ptr += program_id * token_accuracy_stride
+            tl.store(token_accuracy_ptr, 0.0)
         return
 
     loss_ptr += program_id * loss_stride
     if RETURN_Z_LOSS:
         z_loss_ptr += program_id * loss_stride
+    if RETURN_TOKEN_ACCURACY:
+        token_accuracy_ptr += program_id * token_accuracy_stride
 
     if HAS_WEIGHT:
         weight_y = tl.load(weight_ptr + y).cast(tl.float32)
@@ -105,6 +120,7 @@ def liger_cross_entropy_kernel(
     # 3. [Online softmax] first pass: find max + sum
     m = float("-inf")  # m is the max value. use the notation from the paper
     d = 0.0  # d is the sum. use the notation from the paper
+    argmax_idx = 0  # Track the index of the maximum value for token accuracy computation
     ori_X_y = tl.load(X_ptr + y).cast(tl.float32)  # we need to store the original value of X_y for the loss calculation
     if HAS_SOFTCAPPING:
         ori_X_y = softcap * tanh(ori_X_y / softcap)
@@ -125,6 +141,16 @@ def liger_cross_entropy_kernel(
         if HAS_SOFTCAPPING:
             X_block = softcap * tanh(X_block / softcap)
         block_max = tl.max(X_block)
+
+        # Track argmax for accuracy computation
+        if RETURN_TOKEN_ACCURACY and block_max > m:
+            # Find the index of the maximum value in this block
+            is_max_mask = X_block == block_max
+            # Mask out invalid indices with a value larger than n_cols
+            masked_offsets = tl.where(is_max_mask, X_offsets, n_cols)
+            # Get the first (smallest) index where max occurs
+            argmax_idx = tl.min(masked_offsets)
+
         if label_smoothing > 0:
             # scale X beforehand to avoid overflow
             if HAS_WEIGHT:
@@ -155,58 +181,58 @@ def liger_cross_entropy_kernel(
     # For 'sum' reduction, no normalization is applied:
     # dx_y = softmax(x_y) - 1
     # dx_i = softmax(x_i), for i ≠ y
-
-    for i in range(0, n_cols, BLOCK_SIZE):
-        X_offsets = i + tl.arange(0, BLOCK_SIZE)
-        X_block = tl.load(
-            X_ptr + X_offsets,
-            mask=X_offsets < n_cols,
-            other=float("-inf"),
-            # Ensure float32 precision for softmax calculation
-        ).cast(tl.float32)
-        if HAS_SOFTCAPPING:
-            intermediate = tanh(X_block / softcap)
-            X_block = softcap * intermediate
-
-        if not HAS_WEIGHT:
-            # softmax(x_i)
-            X_block = tl.exp(X_block - m) / d
-            # derivative of z-loss: 2 * lse_square_scale * lse * softmax(x_i)
-            X_block += 2 * lse_square_scale * lse * X_block
-            # smoothing term
-            X_block += -eps
-            # special handle dx_y
-            X_block = tl.where(X_offsets != y, X_block, X_block - (1 - label_smoothing))
-            # reduction scale
-            if reduction == "mean":
-                X_block = X_block / n_non_ignore
-        else:
-            weight_block = tl.load(weight_ptr + X_offsets, mask=X_offsets < n_cols)
-            softmax_X = tl.exp(X_block - m) / d
-            # derivative of original_loss
-            dloss_ori = (1 - label_smoothing) * softmax_X
-            # specially handle dx_y
-            dloss_ori = tl.where(X_offsets != y, dloss_ori, dloss_ori - (1 - label_smoothing))
-            dloss_ori = dloss_ori * weight_y
-            # derivative of smooth_loss
-            dloss_smooth = eps * (-weight_block + softmax_X * weight_sum)
-            # derivative of z-loss
-            dz_loss = 2 * lse_square_scale * lse * softmax_X
-            # reduction scale
-            if reduction == "mean":
-                dloss_ori = dloss_ori / sum_non_ignore_weight
-                dloss_smooth = dloss_smooth / sum_non_ignore_weight
-                # TODO: Implement weighted z_loss. Currently, z_loss is not scaled by weight.
-                dz_loss = dz_loss / n_non_ignore
-            # derivative of total_loss
-            X_block = dloss_ori + dloss_smooth + dz_loss
-
-        # chain rule softcapping
-        # d(softcap * tanh(x / softcap)) = (1 - tanh^2(x / softcap))
-        if HAS_SOFTCAPPING:
-            X_block = X_block * (1 - intermediate * intermediate)
-
-        tl.store(X_ptr + X_offsets, X_block, mask=X_offsets < n_cols)
+    if HAS_GRADIENTS:
+        for i in range(0, n_cols, BLOCK_SIZE):
+            X_offsets = i + tl.arange(0, BLOCK_SIZE)
+            X_block = tl.load(
+                X_ptr + X_offsets,
+                mask=X_offsets < n_cols,
+                other=float("-inf"),
+                # Ensure float32 precision for softmax calculation
+            ).cast(tl.float32)
+            if HAS_SOFTCAPPING:
+                intermediate = tanh(X_block / softcap)
+                X_block = softcap * intermediate
+
+            if not HAS_WEIGHT:
+                # softmax(x_i)
+                X_block = tl.exp(X_block - m) / d
+                # derivative of z-loss: 2 * lse_square_scale * lse * softmax(x_i)
+                X_block += 2 * lse_square_scale * lse * X_block
+                # smoothing term
+                X_block += -eps
+                # special handle dx_y
+                X_block = tl.where(X_offsets != y, X_block, X_block - (1 - label_smoothing))
+                # reduction scale
+                if reduction == "mean":
+                    X_block = X_block / n_non_ignore
+            else:
+                weight_block = tl.load(weight_ptr + X_offsets, mask=X_offsets < n_cols)
+                softmax_X = tl.exp(X_block - m) / d
+                # derivative of original_loss
+                dloss_ori = (1 - label_smoothing) * softmax_X
+                # specially handle dx_y
+                dloss_ori = tl.where(X_offsets != y, dloss_ori, dloss_ori - (1 - label_smoothing))
+                dloss_ori = dloss_ori * weight_y
+                # derivative of smooth_loss
+                dloss_smooth = eps * (-weight_block + softmax_X * weight_sum)
+                # derivative of z-loss
+                dz_loss = 2 * lse_square_scale * lse * softmax_X
+                # reduction scale
+                if reduction == "mean":
+                    dloss_ori = dloss_ori / sum_non_ignore_weight
+                    dloss_smooth = dloss_smooth / sum_non_ignore_weight
+                    # TODO: Implement weighted z_loss. Currently, z_loss is not scaled by weight.
+                    dz_loss = dz_loss / n_non_ignore
+                # derivative of total_loss
+                X_block = dloss_ori + dloss_smooth + dz_loss
+
+            # chain rule softcapping
+            # d(softcap * tanh(x / softcap)) = (1 - tanh^2(x / softcap))
+            if HAS_SOFTCAPPING:
+                X_block = X_block * (1 - intermediate * intermediate)
+
+            tl.store(X_ptr + X_offsets, X_block, mask=X_offsets < n_cols)
 
     # We need tl.debug_barrier() to ensure the new result of X_ptr is written as mentioned in
     # https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/ops/cross_entropy.py#L34
@@ -254,6 +280,10 @@ def liger_cross_entropy_kernel(
     tl.store(loss_ptr, loss)
     if RETURN_Z_LOSS:
         tl.store(z_loss_ptr, z_loss)
+    if RETURN_TOKEN_ACCURACY:
+        # Store 1.0 if prediction is correct, 0.0 otherwise
+        is_correct = 1.0 if argmax_idx == y else 0.0
+        tl.store(token_accuracy_ptr, is_correct)
 
 
 # The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576 https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
@@ -272,8 +302,12 @@ def cross_entropy_forward(
     reduction,
     softcap,
     return_z_loss,
+    return_token_accuracy=False,
 ):
     assert isinstance(return_z_loss, bool), f"return_z_loss must be True or False. Got: {return_z_loss}"
+    assert isinstance(return_token_accuracy, bool), (
+        f"return_token_accuracy must be True or False. Got: {return_token_accuracy}"
+    )
 
     BT, V = _input.shape
     n_rows = BT
@@ -283,6 +317,9 @@ def cross_entropy_forward(
     # unreduced loss
     loss_1d = torch.zeros(n_rows, dtype=_input.dtype, device=_input.device)
     z_loss_1d = torch.zeros(n_rows, dtype=_input.dtype, device=_input.device) if return_z_loss else None
+    token_accuracy_1d = (
+        torch.zeros(n_rows, dtype=torch.float32, device=_input.device) if return_token_accuracy else None
+    )
 
     target_mask = target != ignore_index
     n_non_ignore = target_mask.sum().item()
@@ -319,6 +356,10 @@ def cross_entropy_forward(
         loss_ptr=loss_1d,
         z_loss_ptr=z_loss_1d,
         loss_stride=loss_1d.stride(-1),  # always 1
+        token_accuracy_ptr=token_accuracy_1d,
+        token_accuracy_stride=token_accuracy_1d.stride(-1)
+        if return_token_accuracy
+        else 0,  # always 1 if accuracy is enabled
         n_cols=V,
         n_non_ignore=n_non_ignore,
         sum_non_ignore_weight=sum_non_ignore_weight,
@@ -329,9 +370,11 @@ def cross_entropy_forward(
         reduction=reduction,
         softcap=softcap,
         RETURN_Z_LOSS=return_z_loss,
+        RETURN_TOKEN_ACCURACY=return_token_accuracy,
         BLOCK_SIZE=BLOCK_SIZE,
         HAS_WEIGHT=True if weight is not None else False,
         HAS_SOFTCAPPING=True if softcap is not None else False,
+        HAS_GRADIENTS=_input.requires_grad,
         # TODO: 32 seems to give the best performance
         # Performance is quite sensitive to num_warps
         num_warps=32 if not is_hip() else 16,
@@ -340,18 +383,24 @@ def cross_entropy_forward(
     if reduction == "none":
         loss = loss_1d
         z_loss = z_loss_1d if return_z_loss else None
+        token_accuracy = token_accuracy_1d if return_token_accuracy else None
     else:
         loss = torch.sum(loss_1d)
         z_loss = torch.sum(z_loss_1d) if return_z_loss else None
+        # For accuracy, we compute the mean across all non-ignored tokens
+        token_accuracy = torch.sum(token_accuracy_1d) / n_non_ignore if return_token_accuracy else None
 
-    return loss, z_loss, _input
+    return loss, z_loss, token_accuracy, _input
 
 
 def cross_entropy_backward(_input, grad_output):
     # If cross entropy is the last layer, grad_output is 1.0. Skip the mul to save time
     if torch.equal(grad_output, torch.tensor(1.0, device=grad_output.device)):
         pass
-
+    # If reduction is 'none'
+    elif grad_output.ndim > 0:
+        _input = _input * grad_output.unsqueeze(dim=1)
+    # If reduction is ['mean', 'sum'], grad_output is just a scalar
     # We use a Triton kernel instead of a PyTorch operation because modifying inputs in-place
     # for gradient storage and backward multiple times causes anomalies with PyTorch but not with Triton.
     else:
@@ -389,6 +438,7 @@ class LigerCrossEntropyFunction(torch.autograd.Function):
         reduction: str = "mean",
         softcap: Optional[float] = None,
         return_z_loss: bool = False,
+        return_token_accuracy: bool = False,
     ):
         """
         The forward pass of the Liger Cross Entropy loss.
@@ -403,12 +453,15 @@ class LigerCrossEntropyFunction(torch.autograd.Function):
         label_smoothing (float): The amount of smoothing when computing the loss, where 0.0 means no smoothing.
         reduction (str): The reduction to apply to the output: "none" | "mean | "sum".
         softcap (Optional[float]): The upper threshold for scaling logits to the range (-softcap, +softcap).
-        return_z_loss (bool): When `return_z_loss` is `True`, returns (loss, z_loss) instead of (loss, None). Default: `False`
+        return_z_loss (bool): When `return_z_loss` is `True`, returns (loss, z_loss, token_accuracy) instead of (loss, None, None). Default: `False`
+        return_token_accuracy (bool): When `return_token_accuracy` is `True`, computes and returns per-token accuracy without materializing logits. Default: `False`
 
         Returns:
-        tuple: A tuple with the compouted losses with respect to loss and z loss. The elements are tensors or None.
+        tuple: A tuple with the computed losses and accuracy: (loss, z_loss, token_accuracy). z_loss and token_accuracy are None if not requested.
         """
-        loss, z_loss, _input = cross_entropy_forward(
+        input_requires_grad = _input.requires_grad
+
+        loss, z_loss, token_accuracy, _input = cross_entropy_forward(
             _input,
             target,
             weight,
@@ -418,29 +471,35 @@ class LigerCrossEntropyFunction(torch.autograd.Function):
             reduction,
             softcap,
             return_z_loss,
+            return_token_accuracy,
         )
         # TODO: investigation
         # If we don't detach the _input tensor, the memory will double
         # Not sure why but seems that there will be a time both grad and value exist but in different location
-        ctx.save_for_backward(_input.detach())
+        if input_requires_grad:
+            ctx.save_for_backward(_input.detach())
         ctx.return_z_loss = return_z_loss
+        ctx.return_token_accuracy = return_token_accuracy
 
-        return loss, z_loss
+        return loss, z_loss, token_accuracy
 
     @staticmethod
-    def backward(ctx, grad_output, grad_ouput2):
+    def backward(ctx, grad_output, grad_output2, grad_output3):
         """
         The backward pass of the Liger Cross Entropy loss.
 
         Parameters:
         ctx : The context object with saved tensors.
         grad_output (tensor): The tensor containing the gradient of the loss with respect to the output.
-        grad_output2 (tenosr): No use.
+        grad_output2 (tensor): No use. Gradient for z_loss (not used as z_loss is only for logging).
+        grad_output3 (tensor): No use. Gradient for token_accuracy (not used as token_accuracy is only for metrics).
         Returns:
         tuple: A tuple with the gradients with respect to the inputs. The elements are tensors or None.
         """
         if ctx.return_z_loss:
-            del grad_ouput2  # z_loss is only for logging
+            del grad_output2  # z_loss is only for logging
+        if ctx.return_token_accuracy:
+            del grad_output3  # token_accuracy is only for metrics
 
         (_input,) = ctx.saved_tensors
         _input = cross_entropy_backward(_input, grad_output)
@@ -454,4 +513,5 @@ class LigerCrossEntropyFunction(torch.autograd.Function):
             None,
             None,
             None,
+            None,
         )