liger-kernel-nightly 0.5.10.dev20250624183504__py3-none-any.whl → 0.6.4.dev20251121224847__py3-none-any.whl

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/ops/rms_norm.py

@@ -63,7 +63,7 @@ def _rms_norm_forward_kernel(
     3. https://arxiv.org/pdf/1910.07467
     """
 
-    row_idx = tl.program_id(0)
+    row_idx = tl.program_id(0).to(tl.int64)
     col_offsets = tl.arange(0, BLOCK_SIZE)
     mask = col_offsets < n_cols
 
@@ -137,7 +137,7 @@ def _rms_norm_backward_kernel(
     dw = sum(dy * (x / RMS)). summation over BxT dimension
     """
 
-    row_block_id = tl.program_id(0)
+    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)

liger_kernel/ops/rope.py

@@ -32,7 +32,7 @@ def _triton_rope(
 
     # cos size: (1, seq_len, head_dim) or (bsz, seq_len, head_dim)
     # stride: (seq_len * head_dim, head_dim, 1)
-    pid = tl.program_id(0)
+    pid = tl.program_id(0).to(tl.int64)
 
     # locate start address
     q_ptr = q_ptr + pid * q_row_stride

liger_kernel/ops/swiglu.py

@@ -26,7 +26,7 @@ def _swiglu_forward_kernel(a_ptr, b_ptr, c_ptr, stride, n_cols: tl.constexpr, BL
     # sigmoid requires type float32
     a_row = tl.load(a_ptr + col_offsets, mask=mask, other=0).to(tl.float32)
     b_row = tl.load(b_ptr + col_offsets, mask=mask, other=0)
-    c_row = silu(a_row) * b_row
+    c_row = silu(a_row).cast(b_row.dtype) * b_row
     tl.store(c_ptr + col_offsets, c_row, mask=mask)
 
 

liger_kernel/ops/tiled_mlp.py

@@ -0,0 +1,136 @@
+import math
+
+from typing import Callable
+from typing import List
+from typing import Optional
+
+import torch
+
+from liger_kernel.ops.utils import ensure_contiguous
+
+
+class LigerTiledMLPFunction(torch.autograd.Function):
+    """
+    Based on DeepSpeed's TiledMLP:
+    https://github.com/deepspeedai/DeepSpeed/blob/v0.18.2/deepspeed/runtime/sequence_parallel/ulysses_sp.py#L838
+
+    Perform a tiled MLP computation to massively reduce memory usage needed to compute MLP
+    when using very long sequence lengths.
+
+    This module re-computes `forward` in the `backward`. So the `forward` occurs twice each iteration.
+    And if you're using activation checkpointing it then occurs thrice.
+
+    Args:
+        fn: the function to call on sharded inputs (e.g., mlp.forward)
+        mlp_module: the MLP nn.Module object
+        x: the input to MLP.forward (hidden_states)
+        shards: how many shards to use
+        compute_params: a list of weights engaged in the compute
+
+    Returns:
+        the computed hidden_states
+    """
+
+    @staticmethod
+    @ensure_contiguous
+    def forward(
+        ctx,
+        fn: Callable,
+        mlp_module: torch.nn.Module,
+        x: torch.Tensor,
+        shards: int,
+        compute_params: Optional[List[torch.nn.Parameter]] = None,
+    ) -> torch.Tensor:
+        ctx.fn = fn
+        ctx.mlp_module = mlp_module
+        ctx.shards = shards
+        ctx.save_for_backward(x)
+
+        # x.shape could be [bs, seqlen, hidden_size] or [seqlen, hidden_size] (moe experts)
+        x_shards = list(torch.chunk(x, chunks=shards, dim=-2))
+        with torch.no_grad():
+            output_shards = [fn(mlp_module, x_shard) for x_shard in x_shards]
+        output_unsharded = torch.cat(output_shards, dim=-2)
+
+        return output_unsharded
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, *grads) -> tuple:
+        fn = ctx.fn
+        (x,) = ctx.saved_tensors
+        mlp_module = ctx.mlp_module
+        shards = ctx.shards
+
+        x_requires_grad = x.requires_grad
+        x = x.detach()
+        # detach() unsets x.requires_grad, so restore it
+        x.requires_grad_(x_requires_grad)
+
+        # x.shape could be [bs, seqlen, hidden_size] or [seqlen, hidden_size] (moe experts)
+        hidden_size = x.shape[-1]
+        x_shape_orig = x.shape
+
+        # flatten bs+seqlen to avoid having stride issues when narrowing into seqlen w/ bs>1
+        x = x.view(-1, hidden_size)
+        incoming_grad = grads[0].view(-1, hidden_size)
+        x_grad = torch.zeros_like(x)
+
+        x_shards = list(torch.chunk(x, chunks=shards, dim=0))
+
+        for i, x_shard in enumerate(x_shards):
+            x_shard.requires_grad_(x_requires_grad)
+
+            # if seqlen is not exactly divisible by shards the last step will be shorter than shard_step
+            shard_step = x_shards[i].shape[0]
+            shard_offset = i * x_shards[0].shape[0]
+
+            x_shard.grad = x_grad.narrow(0, shard_offset, shard_step).view_as(x_shard)
+            incoming_grad_shard = incoming_grad.narrow(0, shard_offset, shard_step).view_as(x_shard)
+
+            with torch.enable_grad():
+                output = fn(mlp_module, x_shard)
+            torch.autograd.backward(output, incoming_grad_shard)
+
+        # unflatten
+        x_grad = x_grad.view(x_shape_orig)
+
+        return (None, None, x_grad, None, None)
+
+
+def apply_tiled_mlp(
+    fn: Callable,
+    mlp_module: torch.nn.Module,
+    x: torch.Tensor,
+    num_shards: Optional[int] = None,
+    compute_params: Optional[List[torch.nn.Parameter]] = None,
+) -> torch.Tensor:
+    """
+    Apply tiled MLP computation for memory efficiency.
+
+    Args:
+        fn: the function to call on sharded inputs (e.g., lambda module, x: module(x))
+        mlp_module: the MLP nn.Module object
+        x: the input tensor with shape [bs, seqlen, hidden_size] or [seqlen, hidden_size]
+        num_shards: number of shards to use. If None, automatically calculated as ceil(seqlen / hidden_size)
+        compute_params: list of parameters for DeepSpeed ZeRO optimization
+
+    Returns:
+        output tensor with the same shape as input
+    """
+    if num_shards is None:
+        # x.shape could be [bs, seqlen, hidden_size] or [seqlen, hidden_size]
+        hidden_size = x.shape[-1]
+        seqlen = x.shape[-2]
+        num_shards = math.ceil(seqlen / hidden_size)
+
+    # Ensure num_shards is at least 1
+    num_shards = max(1, num_shards)
+
+    return LigerTiledMLPFunction.apply(
+        fn,
+        mlp_module,
+        x,
+        num_shards,
+        compute_params,
+    )