liger-kernel-nightly 0.5.5.dev20250402185702__py3-none-any.whl → 0.6.4.dev20260112233432__py3-none-any.whl

liger_kernel/ops/softmax.py

@@ -0,0 +1,201 @@
+from typing import Tuple
+
+import torch
+import triton
+import triton.language as tl
+
+from liger_kernel.ops.utils import calculate_settings
+from liger_kernel.ops.utils import ensure_contiguous
+
+
+@triton.jit
+def _softmax_single_block_forward_kernel(
+    Y_ptr,
+    Y_row_stride,
+    X_ptr,
+    X_row_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    row_id = tl.program_id(0)
+    offs = tl.arange(0, BLOCK_SIZE)
+    mask = offs < n_cols
+
+    x = tl.load(X_ptr + row_id * X_row_stride + offs, mask=mask, other=-float("inf"), cache_modifier=".ca")
+    m = tl.max(x, axis=0)
+    e = tl.exp(x - m)
+    d = tl.sum(e, axis=0)
+    y = e / d
+    tl.store(Y_ptr + row_id * Y_row_stride + offs, y, mask=mask, cache_modifier=".cs")
+
+
+@triton.jit
+def _softmax_multi_block_forward_kernel(
+    Y_ptr,
+    Y_row_stride,
+    X_ptr,
+    X_row_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    row_id = tl.program_id(0)
+    offs = tl.arange(0, BLOCK_SIZE)
+
+    m = tl.float32(-float("inf"))
+    d = tl.float32(0.0)
+    for start in tl.range(0, n_cols, BLOCK_SIZE):
+        idx = start + offs
+        mask = idx < n_cols
+        xblk = tl.load(X_ptr + row_id * X_row_stride + idx, mask=mask, other=-float("inf"), cache_modifier=".ca")
+        blk_max = tl.max(xblk, axis=0)
+        new_m = tl.max(m, blk_max)
+        d = d * tl.exp(m - new_m) + tl.sum(tl.exp(xblk - new_m), axis=0)
+        m = new_m
+
+    for start in tl.range(0, n_cols, BLOCK_SIZE):
+        idx = start + offs
+        mask = idx < n_cols
+        xblk = tl.load(X_ptr + row_id * X_row_stride + idx, mask=mask, other=-float("inf"), cache_modifier=".ca")
+        yblk = tl.exp(xblk - m) / d
+        tl.store(Y_ptr + row_id * Y_row_stride + idx, yblk, mask=mask, cache_modifier=".cs")
+
+
+@triton.jit
+def _softmax_single_block_backward_kernel(
+    dy_ptr,
+    dy_stride,
+    y_ptr,
+    y_stride,
+    dx_ptr,
+    dx_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    row_id = tl.program_id(0)
+    offs = tl.arange(0, BLOCK_SIZE)
+    mask = offs < n_cols
+
+    dy = tl.load(dy_ptr + row_id * dy_stride + offs, mask=mask, other=0.0)
+    y = tl.load(y_ptr + row_id * y_stride + offs, mask=mask, other=0.0, cache_modifier=".ca")
+    dot = tl.sum(dy * y, axis=0)
+    dx = y * (dy - dot)
+    tl.store(dx_ptr + row_id * dx_stride + offs, dx, mask=mask, cache_modifier=".wb")
+
+
+@triton.jit
+def _softmax_multi_block_backward_kernel(
+    dy_ptr,
+    dy_stride,
+    y_ptr,
+    y_stride,
+    dx_ptr,
+    dx_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    row_id = tl.program_id(0)
+    offs = tl.arange(0, BLOCK_SIZE)
+    acc = tl.float32(0.0)
+
+    for start in tl.range(0, n_cols, BLOCK_SIZE):
+        idx = start + offs
+        mask = idx < n_cols
+        dy_blk = tl.load(dy_ptr + row_id * dy_stride + idx, mask=mask, other=0.0)
+        y_blk = tl.load(y_ptr + row_id * y_stride + idx, mask=mask, other=0.0, cache_modifier=".ca")
+        acc += tl.sum(dy_blk * y_blk, axis=0)
+
+    for start in tl.range(0, n_cols, BLOCK_SIZE):
+        idx = start + offs
+        mask = idx < n_cols
+        dy_blk = tl.load(dy_ptr + row_id * dy_stride + idx, mask=mask, other=0.0)
+        y_blk = tl.load(y_ptr + row_id * y_stride + idx, mask=mask, other=0.0, cache_modifier=".ca")
+        dx_blk = y_blk * (dy_blk - acc)
+        tl.store(dx_ptr + row_id * dx_stride + idx, dx_blk, mask=mask, cache_modifier=".wb")
+
+
+def _softmax_forward(x: torch.Tensor) -> Tuple[torch.Tensor, int, int, bool]:
+    *batch, n_cols = x.shape
+    x2d = x.contiguous().view(-1, n_cols)
+    n_rows = x2d.shape[0]
+
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    y2d = torch.empty_like(x2d)
+
+    if n_cols <= BLOCK_SIZE:
+        _softmax_single_block_forward_kernel[(n_rows,)](
+            y2d, y2d.stride(0), x2d, x2d.stride(0), n_cols, BLOCK_SIZE=BLOCK_SIZE, num_warps=num_warps
+        )
+        multi_block_launch = False
+    else:
+        _softmax_multi_block_forward_kernel[(n_rows,)](
+            y2d, y2d.stride(0), x2d, x2d.stride(0), n_cols, BLOCK_SIZE=BLOCK_SIZE, num_warps=num_warps
+        )
+        multi_block_launch = True
+
+    return y2d.view(*batch, n_cols), BLOCK_SIZE, num_warps, multi_block_launch
+
+
+def _softmax_backward(
+    dy: torch.Tensor,
+    y: torch.Tensor,
+    BLOCK_SIZE: int,
+    num_warps: int,
+    multi_block_launch: bool,
+) -> torch.Tensor:
+    *batch, n_cols = dy.shape
+    dy2d = dy.contiguous().view(-1, n_cols)
+    y2d = y.contiguous().view(-1, n_cols)
+    n_rows = dy2d.shape[0]
+    dx2d = torch.empty_like(dy2d)
+
+    if not multi_block_launch and n_cols <= BLOCK_SIZE:
+        _softmax_single_block_backward_kernel[(n_rows,)](
+            dy2d,
+            dy2d.stride(0),
+            y2d,
+            y2d.stride(0),
+            dx2d,
+            dx2d.stride(0),
+            n_cols,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+        )
+    else:
+        _softmax_multi_block_backward_kernel[(n_rows,)](
+            dy2d,
+            dy2d.stride(0),
+            y2d,
+            y2d.stride(0),
+            dx2d,
+            dx2d.stride(0),
+            n_cols,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+        )
+
+    return dx2d.view(*batch, n_cols)
+
+
+class LigerSoftmaxFunction(torch.autograd.Function):
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, input_: torch.Tensor):
+        y, BLOCK_SIZE, num_warps, multi_block_launch = _softmax_forward(input_)
+        ctx.save_for_backward(y)
+        ctx.BLOCK_SIZE = BLOCK_SIZE
+        ctx.num_warps = num_warps
+        ctx.multi_block_launch = multi_block_launch
+        return y
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_output):
+        (y,) = ctx.saved_tensors
+        dx = _softmax_backward(
+            grad_output,
+            y,
+            ctx.BLOCK_SIZE,
+            ctx.num_warps,
+            ctx.multi_block_launch,
+        )
+        return dx

liger_kernel/ops/sparsemax.py

@@ -0,0 +1,179 @@
+from typing import Tuple
+
+import torch
+import triton
+import triton.language as tl
+
+from liger_kernel.ops.utils import calculate_settings
+from liger_kernel.ops.utils import ensure_contiguous
+
+
+@triton.jit
+def _sparsemax_forward_kernel(
+    x_ptr,
+    x_stride_row,
+    sorted_x_ptr,
+    sorted_x_stride_row,
+    o_ptr,
+    o_stride_row,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+    num_warps: tl.constexpr,
+):
+    pid_row = tl.program_id(0)
+    ptr_x_data_row = x_ptr + pid_row * x_stride_row
+    ptr_sorted_x_data_row = sorted_x_ptr + pid_row * sorted_x_stride_row
+    ptr_output_row = o_ptr + pid_row * o_stride_row
+
+    offs = tl.arange(0, BLOCK_SIZE)
+    mask = offs < n_cols
+
+    z_sorted_block = tl.load(
+        ptr_sorted_x_data_row + offs,
+        mask=mask,
+        other=-float("inf"),
+        cache_modifier=".ca",
+    ).to(tl.float32)
+
+    z_valid = tl.where(mask, z_sorted_block, 0.0)
+    cssv = tl.cumsum(z_valid, 0)
+
+    r = (offs + 1).to(tl.float32)
+    safe_r = tl.where(mask, r, 1.0)
+
+    t_vec = (cssv - 1.0) / safe_r
+
+    support = (z_sorted_block > t_vec) & mask
+
+    k_int = tl.sum(support.to(tl.int32), 0)
+    k_clamped_int = tl.maximum(k_int, 1)
+    k = k_clamped_int.to(tl.float32)
+
+    s = tl.sum(tl.where(support, z_sorted_block, 0.0), 0)
+
+    tau = (s - 1.0) / k
+
+    x_block = tl.load(
+        ptr_x_data_row + offs,
+        mask=mask,
+        other=0.0,
+        cache_modifier=".ca",
+    ).to(tl.float32)
+
+    y = tl.maximum(x_block - tau, 0.0)
+
+    tl.store(
+        ptr_output_row + offs,
+        y.to(ptr_output_row.dtype.element_ty),
+        mask=mask,
+        cache_modifier=".cs",
+    )
+
+
+@triton.jit
+def _sparsemax_backward_kernel(
+    o_ptr, go_ptr, gi_ptr, stride, n_cols, BLOCK_SIZE: tl.constexpr, num_warps: tl.constexpr
+):
+    row = tl.program_id(0)
+    o_row = o_ptr + row * stride
+    go_row = go_ptr + row * stride
+    gi_row = gi_ptr + row * stride
+
+    offs = tl.arange(0, BLOCK_SIZE)
+
+    supp_cnt = tl.zeros((), tl.float32)
+    go_sum = tl.zeros((), tl.float32)
+
+    for i in tl.range(0, tl.cdiv(n_cols, BLOCK_SIZE)):
+        offs_iter = i * BLOCK_SIZE + offs
+        mask_iter = offs_iter < n_cols
+        o_val = tl.load(o_row + offs_iter, mask=mask_iter, other=0.0, cache_modifier=".ca").to(tl.float32)
+        go_val = tl.load(go_row + offs_iter, mask=mask_iter, other=0.0).to(tl.float32)
+        supp = o_val > 0.0
+        go_sum += tl.sum(tl.where(supp, go_val, 0.0))
+        supp_cnt += tl.sum(supp.to(tl.float32))
+
+    for i in tl.range(0, tl.cdiv(n_cols, BLOCK_SIZE)):
+        offs_iter = i * BLOCK_SIZE + offs
+        mask_iter = offs_iter < n_cols
+        o_val = tl.load(o_row + offs_iter, mask=mask_iter, other=0.0, cache_modifier=".ca").to(tl.float32)
+        go_val = tl.load(go_row + offs_iter, mask=mask_iter, other=0.0).to(tl.float32)
+        supp = o_val > 0.0
+        gi_val = tl.where(
+            supp,
+            go_val - tl.cast(go_sum / tl.maximum(supp_cnt, 1e-6), gi_row.dtype.element_ty).to(tl.float32),
+            0.0,
+        )
+        tl.store(gi_row + offs_iter, gi_val.to(gi_row.dtype.element_ty), mask=mask_iter, cache_modifier=".wb")
+
+
+def _sparsemax_forward(x: torch.Tensor, dim: int) -> Tuple[torch.Tensor, torch.Tensor]:
+    if dim < 0:
+        dim += x.dim()
+    x_sw = x.transpose(dim, -1).contiguous()
+    n_cols = x_sw.size(-1)
+    n_rows = x_sw.numel() // n_cols
+    x_flat = x_sw.view(n_rows, n_cols)
+    x_sorted_flat = torch.sort(x_flat.float(), dim=-1, descending=True).values
+
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    out_flat = torch.empty_like(x_flat)
+    grid = (n_rows,)
+    _sparsemax_forward_kernel[grid](
+        x_flat,
+        x_flat.stride(0),
+        x_sorted_flat,
+        x_sorted_flat.stride(0),
+        out_flat,
+        out_flat.stride(0),
+        n_cols,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+    )
+
+    y = out_flat.view_as(x_sw).transpose(dim, -1)
+    return y, out_flat
+
+
+def _sparsemax_backward(
+    grad_out: torch.Tensor,
+    out_flat: torch.Tensor,
+    dim: int,
+) -> torch.Tensor:
+    grad_sw = grad_out.transpose(dim, -1).contiguous()
+    n_cols = grad_sw.size(-1)
+    n_rows = grad_sw.numel() // n_cols
+    go_flat = grad_sw.view(n_rows, n_cols)
+
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    dx_flat = torch.empty_like(go_flat)
+    grid = (n_rows,)
+    _sparsemax_backward_kernel[grid](
+        out_flat,
+        go_flat,
+        dx_flat,
+        out_flat.stride(0),
+        n_cols,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+    )
+
+    dx = dx_flat.view_as(grad_sw).transpose(dim, -1)
+    return dx
+
+
+class LigerSparsemaxFunction(torch.autograd.Function):
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, x: torch.Tensor, dim: int):
+        y, out_flat = _sparsemax_forward(x, dim)
+        ctx.save_for_backward(out_flat)
+        ctx.dim = dim
+        return y
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_out: torch.Tensor):
+        (out_flat,) = ctx.saved_tensors
+        dx = _sparsemax_backward(grad_out, out_flat, ctx.dim)
+        return dx, None

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,
+    )

liger_kernel/ops/utils.py

@@ -78,6 +78,8 @@ def get_amp_custom_fwd_bwd() -> Callable:
             functools.partial(torch.amp.custom_fwd, device_type=device),
             functools.partial(torch.amp.custom_bwd, device_type=device),
         )
+    if hasattr(torch, "npu") and getattr(torch.npu, "amp", None) is not None:
+        return torch.npu.amp.custom_fwd, torch.npu.amp.custom_bwd
     return torch.cuda.amp.custom_fwd, torch.cuda.amp.custom_bwd
 
 
@@ -125,3 +127,15 @@ def element_mul_kernel(
         X_offsets = i + tl.arange(0, BLOCK_SIZE)
         X_block = tl.load(X_ptr + X_offsets, mask=X_offsets < n_cols)
         tl.store(X_ptr + X_offsets, X_block * grad_output, mask=X_offsets < n_cols)
+
+
+def get_npu_core_count(default: int = 20) -> int:
+    """Return NPU vector core count.
+    Fallback to `default` if Triton runtime or NPU device is unavailable.
+    """
+    try:
+        utils = triton.runtime.driver.active.utils
+        props = utils.get_device_properties(0)
+        return int(props.get("num_vectorcore", default))
+    except Exception:
+        return default