liger-kernel-nightly 0.6.2.dev20250919191028__py3-none-any.whl → 0.6.4.dev20251202054858__py3-none-any.whl

liger_kernel/ops/layer_norm.py

@@ -1,3 +1,4 @@
+import math
 import operator
 
 import torch
@@ -7,8 +8,9 @@ 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.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 rsqrt
@@ -63,12 +65,11 @@ def _layer_norm_forward_kernel(
     X_f32 = X_row.to(tl.float32)
 
     # Compute statistics in fp32 for numerical stability
-    n_cols_f32 = n_cols.to(tl.float32)
-    mean = tl.sum(X_f32, axis=0) / n_cols_f32
+    mean = tl.sum(X_f32, axis=0) / n_cols
     X_centered = X_f32 - mean
     # Apply mask to variance calculation to exclude contributions from masked elements
     X_centered_masked = tl.where(mask, X_centered, 0.0)
-    var = tl.sum(X_centered_masked * X_centered_masked, axis=0) / n_cols_f32
+    var = tl.sum(X_centered_masked * X_centered_masked, axis=0) / n_cols
     rstd = rsqrt(var + eps)
 
     # Store statistics (convert back to original dtype only once)
@@ -86,69 +87,87 @@ def _layer_norm_forward_kernel(
 @triton.jit
 def _layer_norm_backward_kernel(
     X_ptr,  # pointer to input, shape (n_rows, n_cols)
+    stride_x,  # stride of each row in input
     W_ptr,  # pointer to weights, shape (n_cols,)
     Mean_ptr,  # pointer to mean, shape (n_rows,)
+    stride_mean,  # stride of each row in mean
     RSTD_ptr,  # pointer to rstd, shape (n_rows,)
+    stride_rstd,  # stride of each row in rstd
     DX_ptr,  # pointer to input grad, shape (n_rows, n_cols)
+    stride_dx,  # stride of each row in input grad
     DW_ptr,  # pointer to weights grad, shape (n_cols,)
+    stride_dw,  # stride of each row in weights grad
     DB_ptr,  # pointer to bias grad, shape (n_cols,)
+    stride_db,  # stride of each row in bias grad
     DY_ptr,  # pointer to output grad, shape (n_rows, n_cols)
-    stride_x,  # stride of each row in input
-    stride_dx,  # stride of each row in input grad
     stride_dy,  # stride of each row in output grad
+    n_rows,
     n_cols,
+    rows_per_program: tl.constexpr,
     BLOCK_SIZE: tl.constexpr,
-    dtype: tl.constexpr,
-    atomic_dtype: tl.constexpr,
 ):
     """
     References:
     https://arxiv.org/abs/1607.06450
     https://github.com/karpathy/llm.c/blob/master/doc/layernorm/layernorm.md
     """
-    row_idx = tl.program_id(0).to(tl.int64)
+    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)
     cols = tl.arange(0, BLOCK_SIZE)
     mask = cols < n_cols
 
+    dW_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+    db_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+
     # Pre-load weights once (same optimization as forward pass)
     w = tl.load(W_ptr + cols, mask=mask, other=0.0)
     w_f32 = w.to(tl.float32)
-    n_cols_f32 = n_cols.to(tl.float32)
 
     # Calculate pointers for this specific row
-    row_X_ptr = X_ptr + row_idx * stride_x
-    row_DX_ptr = DX_ptr + row_idx * stride_dx
-    row_DY_ptr = DY_ptr + row_idx * stride_dy
-    row_Mean_ptr = Mean_ptr + row_idx
-    row_RSTD_ptr = RSTD_ptr + row_idx
-
-    # Load data for this row
-    x = tl.load(row_X_ptr + cols, mask=mask, other=0.0)
-    dy = tl.load(row_DY_ptr + cols, mask=mask, other=0.0)
-    mean = tl.load(row_Mean_ptr)
-    rstd = tl.load(row_RSTD_ptr)
-
-    # Convert to fp32 for numerical stability
-    x_f32 = x.to(tl.float32)
-    dy_f32 = dy.to(tl.float32)
-    mean_f32 = mean.to(tl.float32)
-    rstd_f32 = rstd.to(tl.float32)
-
-    # Compute backward pass for this row
-    x_hat = (x_f32 - mean_f32) * rstd_f32
-    wdy = w_f32 * dy_f32
-    c1 = tl.sum(x_hat * wdy, axis=0) / n_cols_f32
-    c2 = tl.sum(wdy, axis=0) / n_cols_f32
-    dx = (wdy - (x_hat * c1 + c2)) * rstd_f32
-
-    # Store input gradient
-    tl.store(row_DX_ptr + cols, dx.to(dtype), mask=mask)
-
-    # Accumulate weight and bias gradients using atomic operations
-    dw = dy_f32 * x_hat
-    db = dy_f32
-    tl.atomic_add(DW_ptr + cols, dw.to(atomic_dtype), mask=mask)
-    tl.atomic_add(DB_ptr + cols, db.to(atomic_dtype), mask=mask)
+    row_X_ptr = X_ptr + row_start * stride_x
+    row_DX_ptr = DX_ptr + row_start * stride_dx
+    row_DY_ptr = DY_ptr + row_start * stride_dy
+    row_Mean_ptr = Mean_ptr + row_start
+    row_RSTD_ptr = RSTD_ptr + row_start
+
+    for _ in range(row_start, row_end):
+        # Load data for this row
+        x = tl.load(row_X_ptr + cols, mask=mask, other=0.0)
+        dy = tl.load(row_DY_ptr + cols, mask=mask, other=0.0)
+        mean = tl.load(row_Mean_ptr)
+        rstd = tl.load(row_RSTD_ptr)
+
+        # Convert to fp32 for numerical stability
+        x_f32 = x.to(tl.float32)
+        dy_f32 = dy.to(tl.float32)
+        mean_f32 = mean.to(tl.float32)
+        rstd_f32 = rstd.to(tl.float32)
+
+        # Compute backward pass for this row
+        x_hat = (x_f32 - mean_f32) * rstd_f32
+        wdy = w_f32 * dy_f32
+        c1 = tl.sum(x_hat * wdy, axis=0) / n_cols
+        c2 = tl.sum(wdy, axis=0) / n_cols
+        dx = (wdy - (x_hat * c1 + c2)) * rstd_f32
+
+        # Store input gradient
+        tl.store(row_DX_ptr + cols, dx, mask=mask)
+
+        # Accumulate weight and bias gradients for this thread block's assigned rows
+        dw = dy_f32 * x_hat
+        db = dy_f32
+        dW_row += dw
+        db_row += db
+
+        row_X_ptr += stride_x
+        row_DX_ptr += stride_dx
+        row_DY_ptr += stride_dy
+        row_Mean_ptr += stride_mean
+        row_RSTD_ptr += stride_rstd
+
+    tl.store(DW_ptr + row_block_id * stride_dw + cols, dW_row, mask=mask)
+    tl.store(DB_ptr + row_block_id * stride_db + cols, db_row, mask=mask)
 
 
 def layer_norm_forward(X, W, B, eps):
@@ -230,31 +249,25 @@ def layer_norm_backward(dY, X, W, B, Mean, RSTD):
     dY = dY.view(-1, dim)
     n_rows, n_cols = dY.shape
 
-    # Allocate gradient tensors
-    DX = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
-    # Use float32 for weight/bias gradients if bfloat16 (due to atomic_add limitation)
-    grad_dtype = torch.float32 if W.dtype == torch.bfloat16 else W.dtype
-    DW = torch.zeros(n_cols, dtype=grad_dtype, device=W.device)
-    DB = torch.zeros(n_cols, dtype=grad_dtype, device=W.device)
+    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
+
+    # fp32 for numerical stability especially.
+    _DW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
+    _DB = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
 
     # Calculate optimal block size and warp configuration
     BLOCK_SIZE, num_warps = calculate_settings(n_cols)
     if n_cols > BLOCK_SIZE:
         raise RuntimeError(f"Feature dimension {n_cols} exceeds maximum supported size of {BLOCK_SIZE}.")
+    rows_per_program = math.ceil(n_rows / sm_count)
+    grid = (sm_count,)
 
-    # Determine dtype for triton operations
-    triton_dtype = (
-        tl.float32
-        if X.dtype == torch.float32
-        else tl.bfloat16
-        if X.dtype == torch.bfloat16
-        else tl.float16
-        if X.dtype == torch.float16
-        else tl.float32  # fallback
-    )
-
-    # Use float32 for atomic operations if bfloat16 is not supported
-    atomic_dtype = tl.float32 if triton_dtype == tl.bfloat16 else triton_dtype
+    # Allocate gradient tensors
+    DX = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
 
     kernel_args = {"num_warps": num_warps}
     # XPU-specific optimization
@@ -262,28 +275,33 @@ def layer_norm_backward(dY, X, W, B, Mean, RSTD):
         kernel_args.update({"grf_mode": "large", "num_warps": 32, "num_stages": 4})
 
     # Launch kernel with one thread block per row for optimal performance
-    grid = (n_rows,)
     _layer_norm_backward_kernel[grid](
         X,
+        X.stride(0),
         W,
         Mean,
+        Mean.stride(0),
         RSTD,
+        RSTD.stride(0),
         DX,
-        DW,
-        DB,
-        dY,
-        X.stride(0),
         DX.stride(0),
+        _DW,
+        _DW.stride(0),
+        _DB,
+        _DB.stride(0),
+        dY,
         dY.stride(0),
+        n_rows,
         n_cols,
+        rows_per_program=rows_per_program,
         BLOCK_SIZE=BLOCK_SIZE,
-        dtype=triton_dtype,
-        atomic_dtype=atomic_dtype,
         **kernel_args,
     )
 
     DX = DX.view(*shape)
-    return DX, DW.to(W.dtype), DB.to(W.dtype)
+    DW = _DW.sum(dim=0).to(W.dtype)
+    DB = _DB.sum(dim=0).to(B.dtype)
+    return DX, DW, DB
 
 
 class LigerLayerNormFunction(torch.autograd.Function):

liger_kernel/ops/poly_norm.py

@@ -0,0 +1,390 @@
+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.utils import get_npu_multi_processor_count
+from liger_kernel.utils import is_npu_available
+
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
+    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
+    elif X.device.type == "npu":
+        sm_count = get_npu_multi_processor_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/ops/rms_norm.py

@@ -21,8 +21,10 @@ 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 torch_to_triton_dtype
+from liger_kernel.utils import get_npu_multi_processor_count
+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 rsqrt
@@ -349,7 +351,8 @@ def _block_rms_norm_backward_kernel(
 
         # calculate the gradient of W
         if casting_mode == _CASTING_MODE_LLAMA:
-            dW_row += tl.sum(dY_row * (X_row * rstd_row[:, None]).to(X_dtype), 0)
+			# TODO(tcc): use tl.sum(..., dtype=tl.float32) once we upgrade to triton>=3.3.0
+            dW_row += tl.sum((dY_row * (X_row * rstd_row[:, None]).to(X_dtype)).to(tl.float32), 0)
         else:
             # here X_row is already in fp32 (see previous if block)
             dW_row += tl.sum(dY_row * (X_row * rstd_row[:, None]), 0)
@@ -449,6 +452,8 @@ def rms_norm_backward(dY, X, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warp
         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
+    elif X.device.type == "npu":
+        sm_count = get_npu_multi_processor_count()
 
     # fp32 for numerical stability especially.
     _dW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)