liger-kernel-nightly 0.5.2.dev20250115025212__tar.gz → 0.5.2.dev20250119025401__tar.gz

{liger_kernel_nightly-0.5.2.dev20250115025212 → liger_kernel_nightly-0.5.2.dev20250119025401}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20250115025212
+Version: 0.5.2.dev20250119025401
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20250115025212 → liger_kernel_nightly-0.5.2.dev20250119025401}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.2.dev20250115025212"
+version = "0.5.2.dev20250119025401"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }

{liger_kernel_nightly-0.5.2.dev20250115025212 → liger_kernel_nightly-0.5.2.dev20250119025401}/src/liger_kernel/ops/fused_linear_cross_entropy.py

@@ -24,7 +24,9 @@ def fused_linear_cross_entropy_forward(
     label_smoothing=0.0,
     reduction="mean",
     softcap=None,
+    return_z_loss=False,
 ):
+    assert isinstance(return_z_loss, bool), f"return_z_loss must be True or False. Got: {return_z_loss}"
     device = _input.device
 
     # inputs have shape: BT x H
@@ -47,6 +49,7 @@ def fused_linear_cross_entropy_forward(
     grad_bias = torch.zeros_like(bias, device=device) if bias is not None else None
     # we use fp32 for loss accumulator
     loss_1d = torch.zeros(BT, dtype=torch.float32, device=device)
+    z_loss_1d = torch.zeros(BT, dtype=_input.dtype, device=_input.device) if return_z_loss else None
 
     # TODO: evaluate how CUDA synchronization caused by .item() affects the speed
     target_mask = target != ignore_index
@@ -81,6 +84,7 @@ def fused_linear_cross_entropy_forward(
 
         # unreduced loss
         loss_1d_slice = loss_1d[start_idx:end_idx]  # chunk_size,
+        z_loss_1d_slice = z_loss_1d[start_idx:end_idx] if return_z_loss else None
 
         # ensure _input and target are contiguous
         logits_chunk = logits_chunk.contiguous()
@@ -94,7 +98,7 @@ def fused_linear_cross_entropy_forward(
             Y_stride=target_chunk.stride(-1),  # always 1
             weight_ptr=ce_weight,
             loss_ptr=loss_1d_slice,
-            z_loss_ptr=None,
+            z_loss_ptr=z_loss_1d_slice,
             loss_stride=loss_1d_slice.stride(-1),  # always 1
             n_cols=V,
             n_non_ignore=total_n_non_ignore,
@@ -105,7 +109,7 @@ def fused_linear_cross_entropy_forward(
             label_smoothing=label_smoothing,
             reduction=reduction,
             softcap=softcap,
-            RETURN_Z_LOSS=False,
+            RETURN_Z_LOSS=return_z_loss,
             HAS_WEIGHT=True if ce_weight is not None else False,
             HAS_SOFTCAPPING=True if softcap is not None else False,
             BLOCK_SIZE=BLOCK_SIZE,
@@ -113,6 +117,8 @@ def fused_linear_cross_entropy_forward(
         )
 
         loss_1d[start_idx:end_idx] = loss_1d_slice
+        if return_z_loss:
+            z_loss_1d[start_idx:end_idx] = z_loss_1d_slice
         grad_logits_chunk = logits_chunk  # chunk_size x V
 
         grad_input[start_idx:end_idx] = grad_logits_chunk @ weight
@@ -139,9 +145,12 @@ def fused_linear_cross_entropy_forward(
 
     if reduction == "none":
         loss = loss_1d
+        z_loss = z_loss_1d if return_z_loss else None
+
     else:
         loss = torch.sum(loss_1d)
-    return loss, grad_input, grad_weight, grad_bias
+        z_loss = torch.sum(z_loss_1d) if return_z_loss else None
+    return loss, z_loss, grad_input, grad_weight, grad_bias
 
 
 def fused_linear_cross_entropy_backward(grad_output, grad_input, grad_weight, grad_bias):
@@ -206,6 +215,7 @@ class LigerFusedLinearCrossEntropyFunction(torch.autograd.Function):
         label_smoothing=0.0,
         reduction="mean",
         softcap=None,
+        return_z_loss: bool = False,
     ):
         """
         Fusing the last linear layer with cross-entropy loss
@@ -226,7 +236,7 @@ class LigerFusedLinearCrossEntropyFunction(torch.autograd.Function):
         reduction: reduction to apply
         """
 
-        loss, grad_input, grad_weight, grad_bias = fused_linear_cross_entropy_forward(
+        loss, z_loss, grad_input, grad_weight, grad_bias = fused_linear_cross_entropy_forward(
             _input=_input,
             weight=weight,
             target=target,
@@ -237,6 +247,7 @@ class LigerFusedLinearCrossEntropyFunction(torch.autograd.Function):
             label_smoothing=label_smoothing,
             reduction=reduction,
             softcap=softcap,
+            return_z_loss=return_z_loss,
         )
         # downcast to dtype and store for backward
         ctx.save_for_backward(
@@ -244,11 +255,14 @@ class LigerFusedLinearCrossEntropyFunction(torch.autograd.Function):
             grad_weight.detach() if grad_weight is not None else None,
             grad_bias.detach() if bias is not None else None,
         )
-        return loss
+        ctx.return_z_loss = return_z_loss
+        return loss, z_loss
 
     @staticmethod
     @amp_custom_bwd
-    def backward(ctx, grad_output):
+    def backward(ctx, grad_output, grad_output2):
+        if ctx.return_z_loss:
+            del grad_output2  # z_loss is only for logging
         (grad_input, grad_weight, grad_bias) = ctx.saved_tensors
         grad_input, grad_weight, grad_bias = fused_linear_cross_entropy_backward(
             grad_output, grad_input, grad_weight, grad_bias
@@ -264,4 +278,5 @@ class LigerFusedLinearCrossEntropyFunction(torch.autograd.Function):
             None,
             None,
             None,
+            None,
         )

{liger_kernel_nightly-0.5.2.dev20250115025212 → liger_kernel_nightly-0.5.2.dev20250119025401}/src/liger_kernel/transformers/functional.py

@@ -56,8 +56,9 @@ def liger_fused_linear_cross_entropy(
     label_smoothing: float = 0.0,
     reduction: str = "mean",
     softcap: Optional[float] = None,
+    return_z_loss: bool = False,
 ):
-    return LigerFusedLinearCrossEntropyFunction.apply(
+    loss, z_loss = LigerFusedLinearCrossEntropyFunction.apply(
         input,
         weight,
         target,
@@ -68,7 +69,11 @@ def liger_fused_linear_cross_entropy(
         label_smoothing,
         reduction,
         softcap,
+        return_z_loss,
     )
+    if not return_z_loss:
+        return loss
+    return loss, z_loss
 
 
 def liger_fused_linear_jsd(

{liger_kernel_nightly-0.5.2.dev20250115025212 → liger_kernel_nightly-0.5.2.dev20250119025401}/src/liger_kernel/transformers/fused_linear_cross_entropy.py

@@ -14,6 +14,7 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         label_smoothing: float = 0.0,
         reduction: str = "mean",
         softcap: Optional[float] = None,
+        return_z_loss: bool = False,
     ):
         super().__init__()
         assert (label_smoothing >= 0) and (
@@ -31,9 +32,10 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         self.label_smoothing = label_smoothing
         self.reduction = reduction
         self.softcap = softcap
+        self.return_z_loss = return_z_loss
 
     def forward(self, lin_weight, _input, target, bias=None):
-        return LigerFusedLinearCrossEntropyFunction.apply(
+        loss, z_loss = LigerFusedLinearCrossEntropyFunction.apply(
             _input,
             lin_weight,
             target,
@@ -44,4 +46,8 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
             self.label_smoothing,
             self.reduction,
             self.softcap,
+            self.return_z_loss,
         )
+        if not self.return_z_loss:
+            return loss
+        return loss, z_loss

{liger_kernel_nightly-0.5.2.dev20250115025212 → liger_kernel_nightly-0.5.2.dev20250119025401}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20250115025212
+Version: 0.5.2.dev20250119025401
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20250115025212 → liger_kernel_nightly-0.5.2.dev20250119025401}/test/transformers/test_fused_linear_cross_entropy.py

@@ -45,6 +45,7 @@ class TorchLMHeadCE(torch.nn.Module):
         label_smoothing: float = 0.0,
         reduction: str = "mean",
         softcap: Optional[float] = None,
+        return_z_loss: bool = False,
     ):
         super().__init__()
         self.lin = torch.nn.Linear(in_features=H, out_features=V, bias=bias, dtype=dtype)
@@ -54,6 +55,7 @@ class TorchLMHeadCE(torch.nn.Module):
             lse_square_scale=lse_square_scale,
             label_smoothing=label_smoothing,
             reduction=reduction,
+            return_z_loss=return_z_loss,
         )
         self.softcap = softcap
 
@@ -77,6 +79,7 @@ class LigerLMHeadCE(torch.nn.Module):
         label_smoothing: float = 0.0,
         reduction: str = "mean",
         softcap: Optional[float] = None,
+        return_z_loss: bool = False,
     ):
         super().__init__()
         self.lin = torch.nn.Linear(in_features=H, out_features=V, bias=bias, dtype=dtype)
@@ -87,6 +90,7 @@ class LigerLMHeadCE(torch.nn.Module):
             label_smoothing=label_smoothing,
             reduction=reduction,
             softcap=softcap,
+            return_z_loss=return_z_loss,
         )
 
     def forward(self, x, y):
@@ -118,11 +122,11 @@ class LigerLMHeadCE(torch.nn.Module):
 )
 @pytest.mark.parametrize("bias", [True, False])
 @pytest.mark.parametrize(
-    "has_ce_weight, label_smoothing, ignore_index, lse_square_scale, softcap",
+    "has_ce_weight, label_smoothing, ignore_index, lse_square_scale, softcap, return_z_loss",
     [
-        (False, 0, -100, 0, None),
+        (False, 0, -100, 0, None, False),
         # Pass non-default values once to ensure all params work along
-        (True, 0.1, 42, 1e-4, 30.0),
+        (True, 0.1, 42, 1e-4, 30.0, True),
     ],
 )
 def test_correctness(
@@ -139,6 +143,7 @@ def test_correctness(
     ignore_index,
     reduction,
     softcap,
+    return_z_loss,
     atol,
     rtol,
 ):
@@ -156,6 +161,7 @@ def test_correctness(
         ignore_index=ignore_index,
         reduction=reduction,
         softcap=softcap,
+        return_z_loss=return_z_loss,
         dtype=dtype,
     ).to(device)
     liger_lm_head_ce = LigerLMHeadCE(
@@ -168,6 +174,7 @@ def test_correctness(
         ignore_index=ignore_index,
         reduction=reduction,
         softcap=softcap,
+        return_z_loss=return_z_loss,
         dtype=dtype,
     ).to(device)
 
@@ -189,10 +196,16 @@ def test_correctness(
     indices_to_assign = torch.randperm(B * T)[:num_elements_to_assign]  # Randomly select indices
     target[indices_to_assign] = ignore_index
 
-    output1 = torch_lm_head_ce(_input1, target)
-    output2 = liger_lm_head_ce(_input2, target)
+    if return_z_loss:
+        output1, z_output1 = torch_lm_head_ce(_input1, target)
+        output2, z_output2 = liger_lm_head_ce(_input2, target)
+    else:
+        output1 = torch_lm_head_ce(_input1, target)
+        output2 = liger_lm_head_ce(_input2, target)
 
     assert_verbose_allclose(output1, output2, atol=atol, rtol=rtol)
+    if return_z_loss:
+        assert_verbose_allclose(z_output1, z_output2, atol=atol, rtol=rtol)
 
     output1.backward(gradient=torch.ones_like(output1))
     output2.backward(gradient=torch.ones_like(output2))
@@ -230,8 +243,9 @@ def test_correctness(
         (1.0, torch.float32, 1e-5, 5e-4),
     ],
 )
+@pytest.mark.parametrize("ce_weight", [True, False])
 @pytest.mark.parametrize("bias", [True, False])
-def test_correctness_functional(B, T, H, V, scalar, dtype, bias, atol, rtol):
+def test_correctness_functional(B, T, H, V, scalar, dtype, bias, ce_weight, atol, rtol):
     _input = torch.randn(B * T, H, device=device, dtype=dtype) * scalar
     x1 = _input.detach().clone().requires_grad_(True)
     x2 = _input.detach().clone().requires_grad_(True)
@@ -241,15 +255,26 @@ def test_correctness_functional(B, T, H, V, scalar, dtype, bias, atol, rtol):
     weight = torch.randn(V, H, device=device, dtype=dtype)
     bias = torch.randn(V, device=device, dtype=dtype) if bias else None
 
-    y1 = liger_fused_linear_cross_entropy(
+    ce_weight = torch.randn(V, device=device) if ce_weight else None
+    y1, z1 = liger_fused_linear_cross_entropy(
         input=x1,
         weight=weight,
         target=target,
         bias=bias,
+        ce_weight=ce_weight,
+        ignore_index=-100,
+        lse_square_scale=1e-4,
+        label_smoothing=0.1,
+        reduction="mean",
+        softcap=30.0,
+        return_z_loss=True,
+    )
+    y2, z2 = LigerFusedLinearCrossEntropyFunction.apply(
+        x2, weight, target, bias, ce_weight, -100, 1e-4, 0.1, "mean", 30.0, True
     )
-    y2 = LigerFusedLinearCrossEntropyFunction.apply(x2, weight, target, bias)
 
     assert torch.allclose(y1, y2, atol=atol, rtol=rtol)
+    assert torch.allclose(z1, z2, atol=atol, rtol=rtol)
 
     grad_output = torch.randn_like(y1)