liger-kernel-nightly 0.6.2.dev20251013144132__tar.gz → 0.6.2.dev20251014205028__tar.gz

liger_kernel_nightly-0.6.2.dev20251014205028/.github/workflows/docs.yml

@@ -0,0 +1,64 @@
+name: Publish documentation
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - 'docs/**'
+      - 'mkdocs.yml'
+
+permissions:
+  contents: write
+jobs:
+  deploy:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Configure Git Credentials
+        run: |
+          git config user.name github-actions[bot]
+          git config user.email 41898282+github-actions[bot]@users.noreply.github.com
+      - uses: actions/setup-python@v5
+        with:
+          python-version: 3.x
+      - run: echo "cache_id=$(date --utc '+%V')" >> $GITHUB_ENV 
+      - uses: actions/cache@v4
+        with:
+          key: mkdocs-material-${{ env.cache_id }}
+          path: .cache
+          restore-keys: |
+            mkdocs-material-
+      - run: pip install mkdocs-material mkdocstrings[python]
+      # ====== Backup the benchmarks from gh-pages ======
+      # This is necessary because the benchmarks are not included in the documentation build process.
+      # So we need to backup the benchmarks from gh-pages and restore them after the documentation is built.
+      - name: Backup benchmarks from gh-pages
+        run: |
+          git fetch origin gh-pages
+          # create worktree bound to local gh-pages, tracking origin/gh-pages
+          git branch -f gh-pages origin/gh-pages || true
+          mkdir -p ghp && git worktree add ghp gh-pages || true
+          if [ -d ghp/benchmarks ]; then
+            tar -C ghp -czf /tmp/benchmarks.tgz benchmarks
+          fi
+          # IMPORTANT: remove worktree so gh-pages isn't checked out anywhere
+          git worktree remove ghp --force || true
+          echo "Backed up benchmarks from gh-pages"
+      # ====== Deploy the documentation ======
+      - name: Deploy documentation
+        run: mkdocs gh-deploy --force
+      # ====== Restore the benchmarks onto gh-pages ======
+      # This is necessary because the benchmarks are not included in the documentation build process.
+      # So we need to restore the benchmarks onto gh-pages after the documentation is built.
+      - name: Restore benchmarks onto gh-pages
+        run: |
+          # Refresh remote tracking and recreate a clean worktree
+          git fetch origin gh-pages
+          git worktree add -B gh-pages ghp origin/gh-pages
+          if [ -f /tmp/benchmarks.tgz ]; then
+            tar -C ghp -xzf /tmp/benchmarks.tgz
+            git -C ghp add -A
+            git -C ghp commit -m "Restore benchmarks after gh-deploy" || echo "No changes"
+            git -C ghp push origin gh-pages
+          fi
+          git worktree remove ghp --force || true

{liger_kernel_nightly-0.6.2.dev20251013144132 → liger_kernel_nightly-0.6.2.dev20251014205028}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20251013144132
+Version: 0.6.2.dev20251014205028
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

liger_kernel_nightly-0.6.2.dev20251014205028/benchmark/scripts/benchmark_poly_norm.py

@@ -0,0 +1,197 @@
+import torch
+import torch.nn as nn
+import triton
+
+from utils import QUANTILES
+from utils import SingleBenchmarkRunInput
+from utils import SingleBenchmarkRunOutput
+from utils import _test_memory
+from utils import parse_benchmark_script_args
+from utils import run_benchmarks
+
+from liger_kernel.transformers.poly_norm import LigerPolyNorm
+from liger_kernel.utils import infer_device
+
+device = infer_device()
+
+
+class NaivePolyNorm(nn.Module):
+    """
+    Naive PyTorch implementation of PolyNorm.
+
+    Reference:
+        https://github.com/BryceZhuo/PolyCom/
+
+    PolyNorm formula:
+        y = w₀·norm(x³) + w₁·norm(x²) + w₂·norm(x) + b
+        where norm(u) = u / sqrt(mean(u²) + ε)
+    """
+
+    def __init__(self, eps=1e-6):
+        super().__init__()
+        # Align with PolyCom reference: (1/3, 1/3, 1/3) and bias=1.0
+        self.weight = nn.Parameter(torch.full((3,), 1.0 / 3.0))
+        self.bias = nn.Parameter(torch.tensor(1.0))
+        self.variance_epsilon = eps
+
+    def _norm(self, x):
+        """RMSNorm operation"""
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.variance_epsilon)
+
+    def forward(self, hidden_states):
+        """
+        Forward pass of PolyNorm
+
+        Args:
+            hidden_states: input tensor of shape (..., H)
+
+        Returns:
+            output tensor of same shape as input
+        """
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+
+        # Compute powers
+        x_pow3 = hidden_states**3
+        x_pow2 = hidden_states**2
+        x_pow1 = hidden_states**1
+
+        # Normalize each power
+        norm_x3 = self._norm(x_pow3)
+        norm_x2 = self._norm(x_pow2)
+        norm_x1 = self._norm(x_pow1)
+
+        # Weighted sum with bias
+        output = self.weight[0] * norm_x3 + self.weight[1] * norm_x2 + self.weight[2] * norm_x1 + self.bias
+
+        return output.to(input_dtype)
+
+
+def bench_speed_poly_norm(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOutput:
+    N = input.x
+    provider = input.kernel_provider
+    mode = input.kernel_operation_mode
+
+    extra_benchmark_config = input.extra_benchmark_config
+    M = extra_benchmark_config["M"]
+    eps = extra_benchmark_config["eps"]
+    dtype = extra_benchmark_config["dtype"]
+
+    x_shape = (M, N)
+
+    triton_poly = LigerPolyNorm(eps=eps).to(device)
+    naive_poly = NaivePolyNorm(eps=eps).to(device)
+
+    x = torch.randn(x_shape, dtype=dtype, device=device)
+    dy = torch.randn_like(x)
+    x.requires_grad_(True)
+
+    # utility functions
+
+    def y_fwd():
+        if provider == "liger":
+            return triton_poly(x)
+
+        if provider == "huggingface":
+            return naive_poly(x)
+
+    if mode == "forward":
+        ms_50, ms_20, ms_80 = triton.testing.do_bench(
+            y_fwd,
+            grad_to_none=[x],
+            rep=500,
+            quantiles=QUANTILES,
+        )
+    elif mode == "backward":
+        y = y_fwd()
+        ms_50, ms_20, ms_80 = triton.testing.do_bench(
+            lambda: y.backward(dy, retain_graph=True),
+            grad_to_none=[x],
+            rep=500,
+            quantiles=QUANTILES,
+        )
+    elif mode == "full":
+
+        def full():
+            y = y_fwd()
+            y.backward(dy, retain_graph=True)
+
+        ms_50, ms_20, ms_80 = triton.testing.do_bench(
+            full,
+            grad_to_none=[x],
+            rep=500,
+            quantiles=QUANTILES,
+        )
+
+    return SingleBenchmarkRunOutput(
+        y_20=ms_20,
+        y_50=ms_50,
+        y_80=ms_80,
+    )
+
+
+def bench_memory_poly_norm(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOutput:
+    N = input.x
+    provider = input.kernel_provider
+
+    extra_benchmark_config = input.extra_benchmark_config
+    M = extra_benchmark_config["M"]
+    eps = extra_benchmark_config["eps"]
+    dtype = extra_benchmark_config["dtype"]
+
+    x_shape = (M, N)
+
+    triton_poly = LigerPolyNorm(eps=eps).to(device)
+    naive_poly = NaivePolyNorm(eps=eps).to(device)
+
+    x = torch.randn(x_shape, dtype=dtype, device=device)
+    dy = torch.randn_like(x)
+    x.requires_grad_(True)
+
+    # utility functions
+    def y_fwd():
+        if provider == "liger":
+            return triton_poly(x)
+        if provider == "huggingface":
+            return naive_poly(x)
+
+    def full():
+        y = y_fwd()
+        y.backward(dy, retain_graph=True)
+
+    mem_50, mem_20, mem_80 = _test_memory(full, quantiles=QUANTILES)
+
+    return SingleBenchmarkRunOutput(
+        y_20=mem_20,
+        y_50=mem_50,
+        y_80=mem_80,
+    )
+
+
+if __name__ == "__main__":
+    args = parse_benchmark_script_args()
+
+    common_configs = {
+        "kernel_name": "poly_norm",
+        "x_name": "H",
+        "x_label": "hidden size",
+        "x_values": [2**i for i in range(10, 16)],
+        "kernel_providers": ["liger", "huggingface"],
+        "extra_benchmark_configs": [{"M": 2048, "dtype": torch.bfloat16, "eps": 1e-6}],
+        "overwrite": args.overwrite,
+    }
+
+    run_benchmarks(
+        bench_test_fn=bench_speed_poly_norm,
+        kernel_operation_modes=["forward", "full", "backward"],
+        metric_name="speed",
+        metric_unit="ms",
+        **common_configs,
+    )
+    run_benchmarks(
+        bench_test_fn=bench_memory_poly_norm,
+        kernel_operation_modes=["full"],
+        metric_name="memory",
+        metric_unit="MB",
+        **common_configs,
+    )

{liger_kernel_nightly-0.6.2.dev20251013144132 → liger_kernel_nightly-0.6.2.dev20251014205028}/docs/acknowledgement.md

@@ -9,7 +9,6 @@
 We referenced or used the following projects:
 
 
-
 | # | Project                                                                                      | Description                                                                             | Location                                                                                                                         | License                                                                              |
 |---|----------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------|
 | 1 | [Unsloth](https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/unsloth/kernels/utils.py#L43)                              | `calculate_settings` to determine block size and warp; We reuse it for Norm and MLP     | [Liger Kernel Utils](https://github.com/linkedin/Liger-Kernel/blob/e249eee723978bf8610ff1ea2297d048a2417e20/src/liger_kernel/ops/utils.py#L23) | [Apache](https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/LICENSE) |

{liger_kernel_nightly-0.6.2.dev20251013144132 → liger_kernel_nightly-0.6.2.dev20251014205028}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.2.dev20251013144132"
+version = "0.6.2.dev20251014205028"
 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-0.6.2.dev20251014205028/src/liger_kernel/ops/poly_norm.py

@@ -0,0 +1,386 @@
+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
+
+if compare_version("triton", operator.ge, "3.0.0"):
+    try:
+        from triton.language.extra.libdevice import rsqrt
+    except ModuleNotFoundError:
+        from triton.language.extra.cuda.libdevice import rsqrt
+else:
+    from triton.language.math import rsqrt
+
+
+@triton.jit
+def _poly_norm_forward_kernel(
+    Y_ptr,
+    Y_row_stride,
+    X_ptr,
+    X_row_stride,
+    W_ptr,  # weight: [3] for [w0, w1, w2]
+    B_ptr,  # bias: scalar
+    RSTD_ptr,  # cache rstd for backward: shape (n_rows, 3)
+    RSTD_row_stride,
+    n_cols,
+    eps,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    PolyNorm formula:
+        y = w₀·norm(x³) + w₁·norm(x²) + w₂·norm(x) + b
+        where norm(u) = u / sqrt(mean(u²) + ε)
+
+    Reference:
+    1. https://github.com/BryceZhuo/PolyCom/
+    2. https://arxiv.org/pdf/2411.03884
+
+    Cache rstd values for backward pass
+    """
+    row_idx = tl.program_id(0).to(tl.int64)
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    mask = col_offsets < n_cols
+
+    # Load pointers
+    Y_ptr += row_idx * Y_row_stride
+    X_ptr += row_idx * X_row_stride
+    RSTD_ptr += row_idx * RSTD_row_stride
+
+    # Load input row
+    X_row = tl.load(X_ptr + col_offsets, mask=mask, other=0.0)
+
+    # Load weights and bias
+    w0 = tl.load(W_ptr + 0)
+    w1 = tl.load(W_ptr + 1)
+    w2 = tl.load(W_ptr + 2)
+    b = tl.load(B_ptr)
+
+    # Compute x³, x², x
+    X_pow3 = X_row * X_row * X_row
+    X_pow2 = X_row * X_row
+    X_pow1 = X_row
+
+    # Compute norm(x³): norm(u) = u * rsqrt(mean(u²) + eps)
+    mean_square_3 = tl.sum(X_pow3 * X_pow3, axis=0) / n_cols
+    rstd_3 = rsqrt(mean_square_3 + eps)
+    norm_x3 = X_pow3 * rstd_3
+
+    # Compute norm(x²)
+    mean_square_2 = tl.sum(X_pow2 * X_pow2, axis=0) / n_cols
+    rstd_2 = rsqrt(mean_square_2 + eps)
+    norm_x2 = X_pow2 * rstd_2
+
+    # Compute norm(x)
+    mean_square_1 = tl.sum(X_pow1 * X_pow1, axis=0) / n_cols
+    rstd_1 = rsqrt(mean_square_1 + eps)
+    norm_x1 = X_pow1 * rstd_1
+
+    # Cache rstd values for backward
+    tl.store(RSTD_ptr + 0, rstd_3)
+    tl.store(RSTD_ptr + 1, rstd_2)
+    tl.store(RSTD_ptr + 2, rstd_1)
+
+    # Compute output: y = w₀·norm(x³) + w₁·norm(x²) + w₂·norm(x) + b
+    Y_row = w0 * norm_x3 + w1 * norm_x2 + w2 * norm_x1 + b
+
+    # Store output
+    tl.store(Y_ptr + col_offsets, Y_row, mask=mask)
+
+
+@triton.jit
+def _poly_norm_backward_kernel(
+    dY_ptr,
+    dY_row_stride,
+    dX_ptr,
+    dX_row_stride,
+    X_ptr,
+    X_row_stride,
+    W_ptr,
+    RSTD_ptr,
+    RSTD_row_stride,
+    dW_ptr,  # shape: (n_programs, 3)
+    dW_row_stride,
+    dB_ptr,  # shape: (n_programs,)
+    n_rows,
+    n_cols,
+    rows_per_program: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    PolyNorm Backward Kernel Gradient:
+        ∂L/∂x_i = Σ_p w_p * [p*x_i^(p-1) * grad_i/D_p - (p/d)*x_i^(2p-1) * S_p/(D_p³)]
+
+    where:
+        - D_p = RMS(x^p) = 1/rstd_p
+        - S_p = sum(grad * x^p) over the row
+        - d = n_cols
+        - p ∈ {3, 2, 1}
+    """
+    row_block_id = tl.program_id(0).to(tl.int64)
+    row_start = row_block_id * rows_per_program
+    row_end = min((row_block_id + 1) * rows_per_program, n_rows)
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    mask = col_offsets < n_cols
+
+    # Initialize accumulators for weight and bias gradients (scalars)
+    dW0_acc = 0.0
+    dW1_acc = 0.0
+    dW2_acc = 0.0
+    dB_acc = 0.0
+
+    # Load weights
+    w0 = tl.load(W_ptr + 0).to(tl.float32)
+    w1 = tl.load(W_ptr + 1).to(tl.float32)
+    w2 = tl.load(W_ptr + 2).to(tl.float32)
+
+    dY_ptr += row_start * dY_row_stride
+    dX_ptr += row_start * dX_row_stride
+    X_ptr += row_start * X_row_stride
+    RSTD_ptr += row_start * RSTD_row_stride
+
+    for _ in range(row_start, row_end):
+        # Load input and gradient
+        dY_row = tl.load(dY_ptr + col_offsets, mask=mask, other=0.0).to(tl.float32)
+        X_row = tl.load(X_ptr + col_offsets, mask=mask, other=0.0).to(tl.float32)
+
+        # Load cached rstd values
+        rstd_3 = tl.load(RSTD_ptr + 0).to(tl.float32)
+        rstd_2 = tl.load(RSTD_ptr + 1).to(tl.float32)
+        rstd_1 = tl.load(RSTD_ptr + 2).to(tl.float32)
+
+        # Compute powers
+        X_pow3 = X_row * X_row * X_row
+        X_pow2 = X_row * X_row
+        X_pow1 = X_row
+
+        # Accumulate bias gradient: dB = sum(dY)
+        dB_acc += tl.sum(dY_row, axis=0)
+
+        # Compute gradient w.r.t. input using closed-form formula
+        # For p=3: ∂L/∂x from w0 * norm(x³)
+        S_3 = tl.sum(dY_row * X_pow3, axis=0)  # scalar
+        grad_x_3 = w0 * (
+            3.0 * X_pow2 * rstd_3 * dY_row
+            - (3.0 / n_cols) * X_row * X_row * X_row * X_row * X_row * (rstd_3 * rstd_3 * rstd_3) * S_3
+        )
+
+        # For p=2: ∂L/∂x from w1 * norm(x²)
+        S_2 = tl.sum(dY_row * X_pow2, axis=0)  # scalar
+        grad_x_2 = w1 * (
+            2.0 * X_row * rstd_2 * dY_row - (2.0 / n_cols) * X_row * X_row * X_row * (rstd_2 * rstd_2 * rstd_2) * S_2
+        )
+
+        # For p=1: ∂L/∂x from w2 * norm(x)
+        S_1 = tl.sum(dY_row * X_pow1, axis=0)  # scalar
+        grad_x_1 = w2 * (1.0 * rstd_1 * dY_row - (1.0 / n_cols) * X_row * (rstd_1 * rstd_1 * rstd_1) * S_1)
+
+        # Accumulate weight gradients using closed-form: dW_p = rstd_p * S_p
+        dW0_acc += rstd_3 * S_3
+        dW1_acc += rstd_2 * S_2
+        dW2_acc += rstd_1 * S_1
+
+        # Total gradient
+        dX_row = grad_x_3 + grad_x_2 + grad_x_1
+
+        # Store gradient
+        tl.store(dX_ptr + col_offsets, dX_row, mask=mask)
+
+        # Update pointers
+        dY_ptr += dY_row_stride
+        dX_ptr += dX_row_stride
+        X_ptr += X_row_stride
+        RSTD_ptr += RSTD_row_stride
+
+    # Store accumulated gradients (scalars)
+    tl.store(dW_ptr + row_block_id * dW_row_stride + 0, dW0_acc)
+    tl.store(dW_ptr + row_block_id * dW_row_stride + 1, dW1_acc)
+    tl.store(dW_ptr + row_block_id * dW_row_stride + 2, dW2_acc)
+    tl.store(dB_ptr + row_block_id, dB_acc)
+
+
+def poly_norm_forward(X, W, B, eps=1e-6):
+    """
+    PolyNorm Forward Pass
+
+    Args:
+        X: input tensor of shape (*, H) where H is hidden dimension
+        W: weight tensor of shape (3,) for [w0, w1, w2]
+        B: bias scalar tensor
+        eps: epsilon for numerical stability
+
+    Returns:
+        Y: output tensor of same shape as X
+        X: reshaped input (for backward)
+        RSTD: cached rstd values (for backward)
+        BLOCK_SIZE: block size used
+        num_warps: number of warps used
+    """
+    shape = X.shape
+    dim = shape[-1]
+    X = X.view(-1, dim)
+    n_rows, n_cols = X.shape
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+
+    # RSTD is to cache rstd for each row
+    Y = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
+    RSTD = torch.empty((n_rows, 3), dtype=torch.float32, device=X.device)
+
+    # Check constraints
+    assert W.shape[0] == 3, "Weight tensor must have shape (3,)"
+    assert B.numel() == 1, "Bias must be a scalar"
+
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        kernel_args["grf_mode"] = "large"
+
+    # Launch kernel
+    _poly_norm_forward_kernel[(n_rows,)](
+        Y,
+        Y.stride(0),
+        X,
+        X.stride(0),
+        W,
+        B,
+        RSTD,
+        RSTD.stride(0),
+        n_cols,
+        eps,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+        **kernel_args,
+    )
+
+    return Y.view(*shape), X, RSTD, BLOCK_SIZE, num_warps
+
+
+def poly_norm_backward(dY, X, W, RSTD, BLOCK_SIZE, num_warps, in_place):
+    """
+    PolyNorm Backward Pass
+
+    Args:
+        dY: gradient of output
+        X: input tensor (already reshaped to 2D)
+        W: weight tensor
+        RSTD: cached rstd values from forward
+        BLOCK_SIZE: block size from forward
+        num_warps: number of warps from forward
+        in_place: whether to in-place modify dY to store dX (saves memory)
+
+    Returns:
+        dX: gradient w.r.t. input
+        dW: gradient w.r.t. weight
+        dB: gradient w.r.t. bias
+    """
+    shape = dY.shape
+    dim = shape[-1]
+    dY = dY.view(-1, dim)
+    n_rows, n_cols = dY.shape
+
+    # Get number of SMs for parallelization
+    import math
+
+    sm_count = 1
+    if X.device.type == "cuda":
+        sm_count = torch.cuda.get_device_properties(X.device).multi_processor_count
+    elif X.device.type == "xpu":
+        sm_count = torch.xpu.get_device_properties(X.device).gpu_eu_count
+
+    # Allocate or reuse gradients
+    if in_place is True:
+        dX = dY
+    else:
+        dX = torch.zeros_like(dY)
+
+    _dW = torch.empty((sm_count, 3), dtype=torch.float32, device=W.device)
+    _dB = torch.empty((sm_count,), dtype=torch.float32, device=W.device)
+
+    rows_per_program = math.ceil(n_rows / sm_count)
+    grid = (sm_count,)
+
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        kernel_args["grf_mode"] = "large"
+
+    # Launch backward kernel
+    _poly_norm_backward_kernel[grid](
+        dY,
+        dY.stride(0),
+        dX,
+        dX.stride(0),
+        X,
+        X.stride(0),
+        W,
+        RSTD,
+        RSTD.stride(0),
+        _dW,
+        _dW.stride(0),
+        _dB,
+        n_rows,
+        n_cols,
+        rows_per_program,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+        **kernel_args,
+    )
+
+    # Reduce gradients across SMs
+    dX = dX.view(*shape)
+    dW = _dW.sum(dim=0).to(W.dtype)
+    dB = _dB.sum().to(W.dtype)
+
+    return dX, dW, dB
+
+
+class LigerPolyNormFunction(torch.autograd.Function):
+    """
+    PolyNorm Function with forward and backward pass
+
+    PolyNorm formula:
+        y = w₀·norm(x³) + w₁·norm(x²) + w₂·norm(x) + b
+        where norm(u) = u / sqrt(mean(u²) + ε)
+
+    Backward uses closed-form gradient:
+        ∂L/∂x_i = Σ_p w_p * [p*x_i^(p-1) * grad_i/D_p - (p/d)*x_i^(2p-1) * S_p/(D_p³)]
+    """
+
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, X, W, B, eps=1e-6, in_place=True):
+        """
+        Args:
+            X: input tensor of shape (B, T, H) or (BxT, H)
+            W: weight tensor of shape (3,) for [w0, w1, w2]
+            B: bias scalar
+            eps: epsilon for numerical stability
+            in_place: whether to in-place modify grad_output in backward (saves memory)
+
+        Returns:
+            Y: output tensor of same shape as X
+        """
+        Y, X, RSTD, BLOCK_SIZE, num_warps = poly_norm_forward(X, W, B, eps)
+        ctx.BLOCK_SIZE = BLOCK_SIZE
+        ctx.num_warps = num_warps
+        ctx.in_place = in_place
+        ctx.save_for_backward(X, W, RSTD)
+        return Y
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_output):
+        """
+        Args:
+            grad_output: gradient of output
+
+        Returns:
+            dX, dW, dB: gradients w.r.t. X, W, B
+        """
+        X, W, RSTD = ctx.saved_tensors
+        dX, dW, dB = poly_norm_backward(grad_output, X, W, RSTD, ctx.BLOCK_SIZE, ctx.num_warps, ctx.in_place)
+        return dX, dW, dB, None, None

{liger_kernel_nightly-0.6.2.dev20251013144132 → liger_kernel_nightly-0.6.2.dev20251014205028}/src/liger_kernel/transformers/__init__.py

@@ -15,6 +15,7 @@ from liger_kernel.transformers.layer_norm import LigerLayerNorm  # noqa: F401
 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.softmax import LigerSoftmax  # noqa: F401
@@ -137,6 +138,7 @@ __all__ = [
     "LigerJSD",
     "LigerLayerNorm",
     "LigerFusedAddRMSNorm",
+    "LigerPolyNorm",
     "LigerRMSNorm",
     "liger_rotary_pos_emb",
     "liger_llama4_text_rotary_pos_emb",