liger-kernel-nightly 0.6.0.dev20250718050347__py3-none-any.whl → 0.6.0.dev20250718080702__py3-none-any.whl

liger_kernel/ops/fused_add_rms_norm.py

@@ -0,0 +1,412 @@
+import math
+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.ops.utils import torch_to_triton_dtype
+
+if compare_version("triton", operator.ge, "3.0.0"):
+    try:
+        # typical import path with dispatch available
+        from triton.language.extra.libdevice import rsqrt
+    except ModuleNotFoundError:
+        # for working with NGC containers
+        from triton.language.extra.cuda.libdevice import rsqrt
+else:
+    from triton.language.math import rsqrt
+
+
+_CASTING_MODE_NONE: tl.constexpr = tl.constexpr(-1)
+_CASTING_MODE_LLAMA: tl.constexpr = tl.constexpr(0)
+_CASTING_MODE_GEMMA: tl.constexpr = tl.constexpr(1)
+
+
+@triton.jit
+def _fused_add_rms_norm_forward_kernel(
+    Y_ptr,
+    Y_row_stride,
+    S_ptr,  # output residual
+    S_row_stride,
+    X_ptr,
+    X_row_stride,
+    R_ptr,  # input residual
+    R_row_stride,
+    W_ptr,
+    W_row_stride,
+    RSTD_ptr,
+    RSTD_row_stride,
+    n_cols,
+    eps,
+    offset,
+    casting_mode: tl.constexpr,  # constexpr so the `if` blocks can be optimized out
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    This kernel computes the following:
+    1. hidden_states = residual + hidden_states
+    2. residual = hidden_states
+    3. hidden_states = rmsnorm(hidden_states)
+
+    This is a commonly used pattern in the decoder layers of LLMs.
+    Some examples:
+    1. https://github.com/huggingface/transformers/blob/0dc2df5ddafe3cb5824ad24e85beba13e0aa6726/src/transformers/models/qwen3/modeling_qwen3.py#L271
+    2. https://github.com/huggingface/transformers/blob/0dc2df5ddafe3cb5824ad24e85beba13e0aa6726/src/transformers/models/llama4/modeling_llama4.py#L393
+
+    This kernel is inspired by the rms_norm forward kernel, and is adapted to support the residual addition in the forward pass.
+    The backward pass is also adapted to support the residual addition in the backward pass.
+    """
+
+    row_idx = tl.program_id(0).to(tl.int64)
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    mask = col_offsets < n_cols
+
+    Y_ptr += row_idx * Y_row_stride
+    S_ptr += row_idx * S_row_stride
+    X_ptr += row_idx * X_row_stride
+    R_ptr += row_idx * R_row_stride
+    RSTD_ptr += row_idx * RSTD_row_stride
+
+    X_row = tl.load(X_ptr + col_offsets, mask=mask, other=0)
+    R_row = tl.load(R_ptr + col_offsets, mask=mask, other=0)
+    S_row = X_row + R_row
+    tl.store(S_ptr + col_offsets, S_row, mask=mask)
+    S_row_dtype = S_row.dtype
+    W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0)
+
+    # On Llama, only rstd is computed on fp32
+    if casting_mode == _CASTING_MODE_LLAMA:
+        S_row = S_row.to(tl.float32)
+
+    # Gemma computes everything on fp32, and then casts back the output to the original dtype
+    if casting_mode == _CASTING_MODE_GEMMA:
+        W_row = W_row.to(tl.float32)
+        S_row = S_row.to(tl.float32)
+
+    if casting_mode == _CASTING_MODE_NONE:
+        eps = eps.to(S_row_dtype)
+        offset = offset.to(S_row_dtype)
+
+    mean_square = tl.sum(S_row * S_row, axis=0) / n_cols
+    rstd = rsqrt(mean_square + eps)
+
+    # We can save time by caching rms with minimal memory overhead
+    # because rms is much smaller compared to X_row, as rms is for each row.
+    # However, on the computation side, it can save 4 operations (*, sum, /, sqrt).
+    tl.store(RSTD_ptr, rstd)
+
+    S_row = S_row * rstd
+
+    # On Llama, the multiplication with the weight is done on the original dtype
+    if casting_mode == _CASTING_MODE_LLAMA:
+        S_row = S_row.to(S_row_dtype)
+
+    Y_row = S_row * (offset + W_row)
+
+    if casting_mode == _CASTING_MODE_GEMMA:
+        Y_row = Y_row.to(S_row_dtype)
+
+    tl.store(Y_ptr + col_offsets, Y_row, mask=mask)
+
+
+@triton.jit
+def _fused_add_rms_norm_backward_kernel(
+    dY_ptr,
+    dY_row_stride,
+    dS_out_ptr,
+    dS_out_row_stride,
+    dX_ptr,
+    dX_row_stride,
+    X_ptr,
+    X_row_stride,
+    X_dtype: tl.constexpr,
+    W_ptr,
+    W_row_stride,
+    RSTD_ptr,
+    RSTD_row_stride,
+    dW_ptr,
+    dW_row_stride,
+    n_rows,
+    n_cols,
+    offset,
+    rows_per_program: tl.constexpr,
+    casting_mode: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    has_dS_out: tl.constexpr,
+):
+    """
+    This kernel is adapted from the rms_norm backward kernel, and is adapted to support the residual
+    addition in the backward pass. For the following code pattern:
+    1. hidden_states = residual + hidden_states
+    2. residual = hidden_states
+    3. hidden_states = rmsnorm(hidden_states)
+
+    The gradient of hidden_states and residual comes out be exactly same. The value of this gradient is
+    the sum of the gradient of the hidden_states in step 3 and the gradient of the residual in step 2.
+
+    The backward pass computation logic is same as the rms_norm backward kernel, except that the gradient
+    of the hidden_states in step 3 and the gradient of the residual in step 2 are summed up.
+    """
+
+    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
+
+    dW_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+
+    dY_ptr += row_start * dY_row_stride
+    dX_ptr += row_start * dX_row_stride
+    if has_dS_out:
+        dS_out_ptr += row_start * dS_out_row_stride
+
+    X_ptr += row_start * X_row_stride
+    RSTD_ptr += row_start
+
+    W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0.0)
+    W_row = W_row + offset
+
+    for _ in range(row_start, row_end):
+        dY_row = tl.load(dY_ptr + col_offsets, mask=mask, other=0.0)
+        X_row = tl.load(X_ptr + col_offsets, mask=mask, other=0.0)
+
+        # Get cached rms
+        rstd_row = tl.load(RSTD_ptr)
+
+        X_row = X_row.to(tl.float32)
+
+        # Different bacward graphs for different casting modes
+        if casting_mode == _CASTING_MODE_LLAMA:
+            m = (dY_row * W_row).to(tl.float32)
+
+        elif casting_mode == _CASTING_MODE_GEMMA:
+            dY_row = dY_row.to(tl.float32)
+            m = dY_row * W_row
+        else:
+            m = dY_row * W_row
+
+        dX_row = rstd_row * m
+
+        if has_dS_out:
+            dS_out_row = tl.load(dS_out_ptr + col_offsets, mask=mask, other=0.0)
+            dX_row += (rstd_row) * (
+                -(1 / n_cols) * rstd_row * rstd_row * tl.sum(m * X_row, axis=0) * X_row
+            ) + dS_out_row
+            dS_out_ptr += dS_out_row_stride
+        else:
+            dX_row += (rstd_row) * (-(1 / n_cols) * rstd_row * rstd_row * tl.sum(m * X_row, axis=0) * X_row)
+
+        # calculate the gradient of W
+        if casting_mode == _CASTING_MODE_LLAMA:
+            dW_row += dY_row * (X_row * rstd_row).to(X_dtype)
+        else:
+            # here X_row is already in fp32 (see previous if block)
+            dW_row += dY_row * (X_row * rstd_row)
+
+        tl.store(dX_ptr + col_offsets, dX_row.to(X_dtype), mask=mask)
+
+        dY_ptr += dY_row_stride
+        dX_ptr += dX_row_stride
+        X_ptr += X_row_stride
+        RSTD_ptr += RSTD_row_stride
+
+    tl.store(dW_ptr + row_block_id * dW_row_stride + col_offsets, dW_row, mask=mask)
+
+
+_str_to_casting_mode = {
+    "llama": _CASTING_MODE_LLAMA.value,
+    "gemma": _CASTING_MODE_GEMMA.value,
+    "none": _CASTING_MODE_NONE.value,
+}
+
+
+def fused_add_rms_norm_forward(X, R, W, eps, offset, casting_mode):
+    if not isinstance(casting_mode, int):
+        assert casting_mode in _str_to_casting_mode, f"Invalid casting mode: {casting_mode}"
+        casting_mode = _str_to_casting_mode[casting_mode]
+    else:
+        assert casting_mode in _str_to_casting_mode.values(), f"Invalid casting mode: {casting_mode}"
+
+    shape = X.shape
+    dim = shape[-1]
+    X = X.view(-1, dim)
+    R = R.view(-1, dim)
+    n_rows, n_cols = X.shape
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+
+    Y = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
+    S = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
+    # RSTD is to cache rstd for each row
+    # RSTD is always computed/stored in fp32 if we are using Llama or Gemma casting mode
+    rstd_dtype = torch.float32 if casting_mode in (_CASTING_MODE_LLAMA.value, _CASTING_MODE_GEMMA.value) else X.dtype
+    RSTD = torch.empty(n_rows, dtype=rstd_dtype, device=X.device)
+
+    # Check constraints.
+    assert X.shape[1] == W.shape[0], "Incompatible hidden size dimension between tensor1.shape[1] and tensor2.shape[0]"
+
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        kernel_args["grf_mode"] = "large"
+
+    # TODO: add _block_fused_add_rms_norm_forward_kernel
+    _fused_add_rms_norm_forward_kernel[(n_rows,)](
+        Y,
+        Y.stride(0),
+        S,
+        S.stride(0),
+        X,
+        X.stride(0),
+        R,
+        R.stride(0),
+        W,
+        W.stride(0),
+        RSTD,
+        RSTD.stride(0),
+        n_cols,
+        eps,
+        offset,
+        casting_mode,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+        **kernel_args,  # XPU-specific optimization
+    )
+
+    return Y.view(*shape), S.view(*shape), RSTD, BLOCK_SIZE, num_warps, casting_mode
+
+
+def fused_add_rms_norm_backward(dY, dS_out, S, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warps, in_place):
+    shape = dY.shape
+    dim = shape[-1]
+    dY = dY.view(-1, dim)
+    dS_out = dS_out.view(-1, dim)
+    S = S.view(-1, dim)
+    n_rows, n_cols = dY.shape
+
+    sm_count = 1
+    if S.device.type == "cuda":
+        sm_count = torch.cuda.get_device_properties(S.device).multi_processor_count
+    elif S.device.type == "xpu":
+        sm_count = torch.xpu.get_device_properties(S.device).gpu_eu_count
+
+    # fp32 for numerical stability especially.
+    _dW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
+
+    if n_cols > BLOCK_SIZE:
+        raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
+    rows_per_program = math.ceil(n_rows / sm_count)
+    grid = (sm_count,)
+
+    if in_place is True:
+        dX = dY
+    else:
+        dX = torch.empty_like(dY)
+
+    # XPU-specific optimization
+    kernel_args = {}
+    if S.device.type == "xpu":
+        kernel_args["grf_mode"] = "large"
+
+    # TODO: add _block_fused_add_rms_norm_backward_kernel
+    _fused_add_rms_norm_backward_kernel[grid](
+        dY,
+        dY.stride(0),
+        dS_out,
+        dS_out.stride(0),
+        dX,
+        dX.stride(0),
+        S,
+        S.stride(0),
+        torch_to_triton_dtype[S.dtype],
+        W,
+        W.stride(0),
+        RSTD,
+        RSTD.stride(0),
+        _dW,
+        _dW.stride(0),
+        n_rows,
+        n_cols,
+        offset,
+        rows_per_program,
+        casting_mode,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+        has_dS_out=dS_out is not None,
+        **kernel_args,  # XPU-specific optimization
+    )
+
+    dX = dX.view(*shape)
+    dW = _dW.sum(dim=0).to(W.dtype)
+
+    return dX, dX, dW  # dR is equal to dX
+
+
+class LigerFusedAddRMSNormFunction(torch.autograd.Function):
+    """
+    Performs a fused operation that first adds a residual tensor to the hidden_states tensor (`X`), then applies RMSNorm (Root Mean Square Normalization) to the result using the weight tensor `W`, with optional offset and casting mode.
+
+    This class implements the following sequence, commonly used in transformer decoder layers:
+        1. hidden_states = residual + hidden_states
+        2. residual = hidden_states (after addition)
+        3. hidden_states = rmsnorm(hidden_states)
+
+    Both the normalized hidden_states and the updated residual are returned as outputs.
+
+    Some models use an 'offset' to shift the weight tensor `W` by a constant value. For example, Gemma
+    uses an offset of 1.0, so the computation becomes `(X / RMS(X)) * (W + 1.0)` instead of the usual
+    `(X / RMS(X)) * W`. You can pass the offset value as an argument to the forward function.
+
+    In addition, different models cast their inputs at different places during RMSNorm computation. For
+    example, Gemma casts everything to fp32 before starting the computation, while Llama casts only the
+    inverse RMS to fp32. You can specify the casting mode using the `casting_mode` argument. We currently
+    support the following casting modes (they match HuggingFace Transformers' implementations):
+    - 'llama': matches the Llama implementation, where only the inverse RMS is computed on fp32.
+    - 'gemma': matches the Gemma implementation, where everything is cast to fp32, then computed, then cast back to the original dtype.
+    - 'none': no casting is done. The computation is done in the original dtype. This saves memory and is slightly faster, but has more error w.r.t. the original implementation.
+
+    The `in_place` option determines whether to modify dY in-place to store dX. This defaults to `True` to save memory.
+    """
+
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, X, R, W, eps, offset=0.0, casting_mode="llama", in_place=False):
+        """
+        X: (B, T, H) or (BxT, H)
+        W: (H,)
+        """
+        # TODO: add row_mode
+        Y, S, RSTD, BLOCK_SIZE, num_warps, casting_mode = fused_add_rms_norm_forward(X, R, W, eps, offset, casting_mode)
+        ctx.offset = offset
+        ctx.casting_mode = casting_mode
+        ctx.in_place = in_place
+        ctx.BLOCK_SIZE = BLOCK_SIZE
+        ctx.num_warps = num_warps
+        ctx.save_for_backward(S, W, RSTD)
+        return Y, S
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, dY, dS_out):
+        """
+        Y: (B, T, H) or (BxT, H)
+        """
+        S, W, RSTD = ctx.saved_tensors
+        dX, dR, dW = fused_add_rms_norm_backward(
+            dY,
+            dS_out,
+            S,
+            W,
+            RSTD,
+            ctx.offset,
+            ctx.casting_mode,
+            ctx.BLOCK_SIZE,
+            ctx.num_warps,
+            ctx.in_place,
+        )
+
+        return dX, dR, dW, None, None, None, None, None

liger_kernel/transformers/__init__.py

@@ -5,6 +5,7 @@ from typing import TYPE_CHECKING
 # Always-safe imports (independent of 'transformers')
 from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss  # noqa: F401
 from liger_kernel.transformers.dyt import LigerDyT  # noqa: F401
+from liger_kernel.transformers.fused_add_rms_norm import LigerFusedAddRMSNorm  # noqa: F401
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss  # noqa: F401
 from liger_kernel.transformers.fused_linear_jsd import LigerFusedLinearJSD  # noqa: F401
 from liger_kernel.transformers.geglu import LigerGEGLUMLP  # noqa: F401
@@ -43,6 +44,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_vl  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_moe  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_smollm3  # noqa: F401
 
 
 # Check if 'transformers' is installed
@@ -100,6 +102,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen2_vl",
         "apply_liger_kernel_to_qwen3",
         "apply_liger_kernel_to_qwen3_moe",
+        "apply_liger_kernel_to_smollm3",
     }
 
     if name in monkey_patch_symbols:
@@ -119,6 +122,7 @@ __all__ = [
     "LigerGEGLUMLP",
     "LigerJSD",
     "LigerLayerNorm",
+    "LigerFusedAddRMSNorm",
     "LigerRMSNorm",
     "liger_rotary_pos_emb",
     "LigerBlockSparseTop2MLP",
@@ -155,5 +159,6 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_qwen2_vl",
             "apply_liger_kernel_to_qwen3",
             "apply_liger_kernel_to_qwen3_moe",
+            "apply_liger_kernel_to_smollm3",
         ]
     )

liger_kernel/transformers/functional.py

@@ -2,6 +2,7 @@ from typing import Optional
 
 from liger_kernel.ops.cross_entropy import LigerCrossEntropyFunction
 from liger_kernel.ops.dyt import LigerDyTFunction
+from liger_kernel.ops.fused_add_rms_norm import LigerFusedAddRMSNormFunction
 from liger_kernel.ops.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyFunction
 from liger_kernel.ops.fused_linear_jsd import LigerFusedLinearJSDFunction
 from liger_kernel.ops.fused_neighborhood_attention import LigerFusedNeighborhoodAttentionFunction
@@ -253,6 +254,10 @@ def liger_rms_norm(X, W, eps, offset: float = 0.0, casting_mode: str = "llama",
     return LigerRMSNormFunction.apply(X, W, eps, offset, casting_mode, in_place)
 
 
+def liger_fused_add_rms_norm(X, R, W, eps, offset: float = 0.0, casting_mode: str = "llama", in_place: bool = True):
+    return LigerFusedAddRMSNormFunction.apply(X, R, W, eps, offset, casting_mode, in_place)
+
+
 def liger_rope(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     return LigerRopeFunction.apply(q, k, cos, sin, position_ids, unsqueeze_dim)
 

liger_kernel/transformers/fused_add_rms_norm.py

@@ -0,0 +1,39 @@
+import torch
+import torch.nn as nn
+
+from liger_kernel.ops.fused_add_rms_norm import LigerFusedAddRMSNormFunction
+
+
+class LigerFusedAddRMSNorm(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        eps=1e-6,
+        offset=0.0,
+        casting_mode="llama",
+        init_fn="ones",
+        in_place=False,
+    ):
+        super().__init__()
+        assert init_fn in [
+            "ones",
+            "zeros",
+        ], f"init_fn must be either 'ones' or 'zeros', got {init_fn}"
+        self.weight = nn.Parameter(torch.ones(hidden_size) if init_fn == "ones" else torch.zeros(hidden_size))
+        self.variance_epsilon, self.offset, self.casting_mode, self.in_place = (eps, offset, casting_mode, in_place)
+
+    def forward(self, hidden_states, residual):
+        return LigerFusedAddRMSNormFunction.apply(
+            hidden_states,
+            residual,
+            self.weight,
+            self.variance_epsilon,
+            self.offset,
+            self.casting_mode,
+            self.in_place,
+        )
+
+    def extra_repr(self):
+        return (
+            f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}, offset={self.offset}, in_place={self.in_place}"
+        )

liger_kernel/transformers/model/smollm3.py

@@ -0,0 +1,189 @@
+from typing import TYPE_CHECKING
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from torch.distributed.fsdp import FullyShardedDataParallel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.utils.deprecation import deprecate_kwarg
+
+from liger_kernel.transformers.fsdp import _FSDPForwardRedirection
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.utils import PEFT_AVAILABLE
+
+if TYPE_CHECKING:
+    from transformers.cache_utils import Cache
+
+if PEFT_AVAILABLE:
+    from peft.utils.other import ModulesToSaveWrapper
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Union["Cache", List[torch.FloatTensor]]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+    Args:
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Smollm3ForCausalLM
+
+    >>> model = Smollm3ForCausalLM.from_pretrained("HuggingFaceTB/SmolLM3-3B")
+    >>> tokenizer = AutoTokenizer.from_pretrained("HuggingFaceTB/SmolLM3-3B")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+    kept_hidden_states = hidden_states[:, slice_indices, :]
+
+    shift_labels = kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+    # if in training mode, don't materialize logits
+    if skip_logits and labels is None and shift_labels is None:
+        raise ValueError("skip_logits is True, but labels and shift_labels are None")
+
+    if skip_logits is None:
+        # By default, if in training mode, don't materialize logits
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    if skip_logits:
+        loss = lce_maybe_trainable_lm_head(
+            self,
+            hidden_states=kept_hidden_states,
+            hidden_size=self.config.hidden_size,
+            labels=labels,
+            shift_labels=shift_labels,
+            **kwargs,
+        )
+
+    else:
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+def lce_maybe_trainable_lm_head(self, hidden_states, hidden_size, labels, shift_labels, **loss_kwargs):
+    lm_head = self.lm_head
+
+    # Unwrap the module if lm_head has been added as trainable module in PEFT LoRA configuration,
+    # i.e. listed in the modules_to_save field of LoraConfig, so the lm_head weights are read
+    # from the unwrapped module.
+    # See https://huggingface.co/docs/peft/package_reference/lora for reference.
+    if PEFT_AVAILABLE and isinstance(lm_head, ModulesToSaveWrapper):
+        lm_head = lm_head.modules_to_save.default
+
+    # If FSDP is used and lm_head is trainable, e.g., during full fine-tuning or with LoRA,
+    # reading the lm_head module weights and calling the kernel must be done within FSDP forward pass
+    # so the module entire parameters are summoned and kept in memory during the kernel execution.
+    if isinstance(lm_head, FullyShardedDataParallel):
+        return _FSDPForwardRedirection()(
+            lm_head,
+            _liger_for_causal_lm_loss,
+            lm_head.module,
+            hidden_states,
+            hidden_size,
+            labels,
+            shift_labels,
+            **loss_kwargs,
+        )
+
+    # FSDP is not used so we can read the lm_head weights and call the kernel directly
+    return _liger_for_causal_lm_loss(
+        lm_head=self.lm_head,
+        hidden_states=hidden_states,
+        hidden_size=hidden_size,
+        labels=labels,
+        shift_labels=shift_labels,
+        **loss_kwargs,
+    )
+
+
+def _liger_for_causal_lm_loss(lm_head, hidden_states, hidden_size, labels, shift_labels, **loss_kwargs):
+    return LigerForCausalLMLoss(
+        hidden_states=hidden_states,
+        lm_head_weight=lm_head.weight,
+        labels=labels,
+        hidden_size=hidden_size,
+        shift_labels=shift_labels,
+        **loss_kwargs,
+    )

liger_kernel/transformers/monkey_patch.py

@@ -29,6 +29,7 @@ from liger_kernel.transformers.model.phi3 import lce_forward as phi3_lce_forward
 from liger_kernel.transformers.model.phi3 import lce_forward_deprecated as phi3_lce_forward_deprecated
 from liger_kernel.transformers.model.qwen2 import lce_forward as qwen2_lce_forward
 from liger_kernel.transformers.model.qwen2 import lce_forward_deprecated as qwen2_lce_forward_deprecated
+from liger_kernel.transformers.model.smollm3 import lce_forward as smollm3_lce_forward
 from liger_kernel.transformers.qwen2vl_mrope import liger_multimodal_rotary_pos_emb
 from liger_kernel.transformers.rms_norm import LigerRMSNorm
 from liger_kernel.transformers.rope import liger_rotary_pos_emb
@@ -290,6 +291,77 @@ def apply_liger_kernel_to_llama(
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_smollm3(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace SmolLM3 model
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is True.
+        cross_entropy (bool): Whether to apply Liger's cross entropy loss. Default is False.
+        fused_linear_cross_entropy (bool):
+            Whether to apply Liger's fused linear cross entropy loss. Default is True.
+            `cross_entropy` and `fused_linear_cross_entropy` cannot both be True.
+            If `fused_linear_cross_entropy` is True, the logits will not be materialized but more memory efficient.
+        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is True.
+        swiglu (bool): Whether to apply Liger's SwiGLU MLP. Default is True.
+        model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
+        loaded. Default is None.
+    """
+
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.smollm3 import modeling_smollm3
+    from transformers.models.smollm3.modeling_smollm3 import SmolLM3Model
+
+    if rope:
+        modeling_smollm3.apply_rotary_pos_emb = liger_rotary_pos_emb
+    if rms_norm:
+        modeling_smollm3.SmolLM3RMSNorm = LigerRMSNorm
+    if swiglu:
+        modeling_smollm3.SmolLM3MLP = LigerSwiGLUMLP
+
+    if cross_entropy:
+        if transformer_version >= version.parse(SUPPORTED_TRANSFORMER_VERSION):
+            from transformers.loss.loss_utils import nn
+
+            nn.functional.cross_entropy = liger_cross_entropy
+        else:
+            logger.warning(TRANSFORMER_DEPRECATION_WARNING)
+            modeling_smollm3.CrossEntropyLoss = LigerCrossEntropyLoss
+
+    if fused_linear_cross_entropy:
+        if model is not None:
+            model.forward = MethodType(smollm3_lce_forward, model)
+        else:
+            modeling_smollm3.SmolLM3ForCausalLM.forward = smollm3_lce_forward
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules (e.g. SmolLM3RMSNorm or SmolLM3MLP)
+
+        # get the base model from the model instance
+        base_model: SmolLM3Model = getattr(model, model.base_model_prefix, model)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.input_layernorm)
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+
 def apply_liger_kernel_to_llava(
     cross_entropy: bool = False,
     fused_linear_cross_entropy: bool = True,
@@ -1801,6 +1873,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "qwen2_vl_text": apply_liger_kernel_to_qwen2_vl,
     "qwen2_5_vl": apply_liger_kernel_to_qwen2_5_vl,
     "qwen2_5_vl_text": apply_liger_kernel_to_qwen2_5_vl,
+    "smollm3": apply_liger_kernel_to_smollm3,
     "phi3": apply_liger_kernel_to_phi3,
     "paligemma": apply_liger_kernel_to_paligemma,
 }

{liger_kernel_nightly-0.6.0.dev20250718050347.dist-info → liger_kernel_nightly-0.6.0.dev20250718080702.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.0.dev20250718050347
+Version: 0.6.0.dev20250718080702
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.0.dev20250718050347.dist-info → liger_kernel_nightly-0.6.0.dev20250718080702.dist-info}/RECORD

@@ -19,6 +19,7 @@ liger_kernel/chunked_loss/simpo_loss.py,sha256=fy2w8KbhMrBv7b1jdIeH3bBFxY52bPQPZ
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 liger_kernel/ops/cross_entropy.py,sha256=e8THGnhOcy_0SbOLABx67HEM7-B8a8pG7nDKbCRpQKM,19123
 liger_kernel/ops/dyt.py,sha256=gCLz4S8aul8SY9nvIGaoK67aGb7U9MJRQdo3ONqmQYs,5417
+liger_kernel/ops/fused_add_rms_norm.py,sha256=UBqmlqFCmhSAIpkNKd8rrfXatX7Z4J9bp2dX9A0lrJQ,14017
 liger_kernel/ops/fused_linear_cross_entropy.py,sha256=5fbGhN85n3zf0uIdJ7PYHWIRzTf0VTFiS0ARtOmqIP0,11020
 liger_kernel/ops/fused_linear_jsd.py,sha256=CSoprxb-YcJy-YUKiTcYkxN8sb9h2kdk_iHuncvSV5c,9683
 liger_kernel/ops/fused_neighborhood_attention.py,sha256=vPi5xbnh6wxyZehaqo6Tuilqo2fN5SGDiONjnNmIKqs,35556
@@ -39,12 +40,13 @@ liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
 liger_kernel/ops/utils.py,sha256=uoFKQqo-34N2TWQNvXMFywqGiOMMXNEVBxVojzlUAa0,3836
 liger_kernel/ops/experimental/embedding.py,sha256=tolj3tItkzpSb30zWqDN2_yX4ectflaQ8HMyKyFIQc8,4172
 liger_kernel/ops/experimental/mm_int8int2.py,sha256=TrS9lpwekrik_w5qE7AhMJD1bcq-OidjtbsW80oZ6IM,13314
-liger_kernel/transformers/__init__.py,sha256=mWMEhOabqUkPimMOmkg9DawnO-vL9u_u-N4iIqfNZeg,7259
+liger_kernel/transformers/__init__.py,sha256=VoHQp5emsAJAouql37RuvtGFeZCoMIHgoIxfsyYMTc8,7564
 liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
 liger_kernel/transformers/cross_entropy.py,sha256=z3KTWQnFxr_IZaVjtYt0ZNEWQdDdYThN35xWkHlDGH0,1683
 liger_kernel/transformers/dyt.py,sha256=i-4GPaMrl-jab9TVI5qN0-H9qycn_mCbV82ozU4nbmU,723
 liger_kernel/transformers/fsdp.py,sha256=CUiyjTmjkjY7pLXQv8ly9rnzgXw6529csd9pvtJNMYc,3096
-liger_kernel/transformers/functional.py,sha256=7Emw7D6VPMg8hfasC33NiolvKmQVF1gV6VayKQCEWJM,7446
+liger_kernel/transformers/functional.py,sha256=PXnACWD7kzgge50RdOUuvtmOTS7DVkkrL7mm0cX5bOc,7734
+liger_kernel/transformers/fused_add_rms_norm.py,sha256=7_Bzg-x6lLe6W1qG2DtjDALhEpNZlC6N5GppEs9cTYY,1199
 liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=O8Sg5BT81nTaY9fSGoOY9dOD9ekibwwiuXhdUHaxntQ,1742
 liger_kernel/transformers/fused_linear_jsd.py,sha256=bZ4otCvWBuOnA5XdQL-FzZVItJlDt-ht9e_pG7PG93E,3999
 liger_kernel/transformers/fused_neighborhood_attention.py,sha256=TxYDUAt9B6WSP14aJP66C_2Mbds2sSIPGnamhUSTrC8,7957
@@ -54,7 +56,7 @@ liger_kernel/transformers/grpo_loss.py,sha256=uAkUNKSnUGEOqa82L9w2e6AI1kcmG8K45-
 liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
 liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
 liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
-liger_kernel/transformers/monkey_patch.py,sha256=W7KgJN-rrLZS3pRZ5debO_dSN7zddPegKjqOIP39wR0,85856
+liger_kernel/transformers/monkey_patch.py,sha256=VsN839y5QVEC6BD_-hCiShWLerQM2QDLDoKf2rq02I4,88990
 liger_kernel/transformers/multi_token_attention.py,sha256=l9VDICK0dfmifUDW668hGscP8AHq2rYcM2oGUa3baRQ,1751
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=vkekcvTeWY8vL4H6hg3t0XeY0Ew_3OFMPHuzqlxPPVw,2719
@@ -85,13 +87,14 @@ liger_kernel/transformers/model/qwen2_5_vl.py,sha256=zEVVwotCXnAm3RRc8-1Nc8uitSW
 liger_kernel/transformers/model/qwen2_vl.py,sha256=5vK-vtCDpKZ2w33xYp2BS8kQYWUbKMqaiKvQcI27Mss,5884
 liger_kernel/transformers/model/qwen3.py,sha256=w2jBHuK9kK9EmOr5dnEIXNQXUgUSV_sJUkXSEwxLPHs,4885
 liger_kernel/transformers/model/qwen3_moe.py,sha256=BkpfFH3fOH0yRfA7LF-AoHTLut2GV0Y4MOlkiIYewfU,5511
+liger_kernel/transformers/model/smollm3.py,sha256=mqayvpwpMbp2yd_Ue7IPzy-dA4KHSDi_ROZW5vHCHfQ,7596
 liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7HHWHwku25A-GYL0WU,193
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
 liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
 liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
-liger_kernel_nightly-0.6.0.dev20250718050347.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.6.0.dev20250718050347.dist-info/METADATA,sha256=piNBLykxYNNlxW-gKCkLrzHXuBvhbYMV03RZgFImk2Q,24672
-liger_kernel_nightly-0.6.0.dev20250718050347.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.6.0.dev20250718050347.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.6.0.dev20250718050347.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.6.0.dev20250718050347.dist-info/RECORD,,
+liger_kernel_nightly-0.6.0.dev20250718080702.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.6.0.dev20250718080702.dist-info/METADATA,sha256=mNkIMGPTMPdmmjDsW54kDe0WhPi8Ep0Cpt4koWQQuaE,24672
+liger_kernel_nightly-0.6.0.dev20250718080702.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.6.0.dev20250718080702.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.6.0.dev20250718080702.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.6.0.dev20250718080702.dist-info/RECORD,,