liger-kernel-nightly 0.6.3.dev20251106220336__py3-none-any.whl → 0.6.3.dev20251118154655__py3-none-any.whl

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/transformers/__init__.py

@@ -24,6 +24,8 @@ from liger_kernel.transformers.swiglu import LigerBlockSparseTop2MLP  # noqa: F4
 from liger_kernel.transformers.swiglu import LigerPhi3SwiGLUMLP  # noqa: F401
 from liger_kernel.transformers.swiglu import LigerQwen3MoeSwiGLUMLP  # noqa: F401
 from liger_kernel.transformers.swiglu import LigerSwiGLUMLP  # noqa: F401
+from liger_kernel.transformers.tiled_mlp import LigerTiledGEGLUMLP  # noqa: F401
+from liger_kernel.transformers.tiled_mlp import LigerTiledSwiGLUMLP  # noqa: F401
 from liger_kernel.transformers.tvd import LigerTVDLoss  # noqa: F401
 
 # Static-only imports for IDEs and type checkers
@@ -155,6 +157,8 @@ __all__ = [
     "LigerPhi3SwiGLUMLP",
     "LigerQwen3MoeSwiGLUMLP",
     "LigerSwiGLUMLP",
+    "LigerTiledGEGLUMLP",
+    "LigerTiledSwiGLUMLP",
     "LigerTVDLoss",
     "LigerKLDIVLoss",
     "LigerMultiTokenAttention",

liger_kernel/transformers/tiled_mlp.py

@@ -0,0 +1,133 @@
+from typing import Optional
+
+import torch.nn as nn
+
+from liger_kernel.ops.geglu import LigerGELUMulFunction
+from liger_kernel.ops.swiglu import LigerSiLUMulFunction
+from liger_kernel.ops.tiled_mlp import apply_tiled_mlp
+
+
+class LigerTiledGEGLUMLP(nn.Module):
+    """
+    Memory-efficient GEGLU MLP using tiled computation.
+
+    This module combines GEGLU activation with tiled processing to handle
+    very long sequences efficiently. The forward pass is recomputed during
+    backward to save memory.
+
+    Args:
+        config: Model configuration with hidden_size and intermediate_size attributes
+        num_shards: Number of shards to split the sequence. If None, automatically
+                   calculated as ceil(seqlen / hidden_size)
+    """
+
+    def __init__(self, config, num_shards: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.num_shards = num_shards
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+        # Validate activation function
+        if hasattr(config, "hidden_act") and config.hidden_act not in [
+            "gelu",
+            "gelu_new",
+            "gelu_pytorch_tanh",
+        ]:
+            raise ValueError(f"LigerTiledGEGLUMLP requires GELU activation, got {config.hidden_act}")
+
+    def _mlp_forward(self, module, x):
+        """Internal MLP forward function for tiled computation."""
+        gate = module.gate_proj(x)
+        up = module.up_proj(x)
+        return module.down_proj(LigerGELUMulFunction.apply(gate, up))
+
+    def forward(self, x):
+        """
+        Forward pass with tiled computation.
+
+        Args:
+            x: Input tensor of shape [batch_size, seq_len, hidden_size]
+               or [seq_len, hidden_size]
+
+        Returns:
+            Output tensor of the same shape as input
+        """
+        compute_params = [
+            self.gate_proj.weight,
+            self.up_proj.weight,
+            self.down_proj.weight,
+        ]
+
+        return apply_tiled_mlp(
+            fn=self._mlp_forward,
+            mlp_module=self,
+            x=x,
+            num_shards=self.num_shards,
+            compute_params=compute_params,
+        )
+
+
+class LigerTiledSwiGLUMLP(nn.Module):
+    """
+    Memory-efficient SwiGLU MLP using tiled computation.
+
+    This module combines SwiGLU activation with tiled processing to handle
+    very long sequences efficiently. The forward pass is recomputed during
+    backward to save memory.
+
+    Args:
+        config: Model configuration with hidden_size and intermediate_size attributes
+        num_shards: Number of shards to split the sequence. If None, automatically
+                   calculated as ceil(seqlen / hidden_size)
+    """
+
+    def __init__(self, config, num_shards: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.num_shards = num_shards
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+        # Validate activation function
+        if hasattr(config, "hidden_act") and config.hidden_act not in ["silu", "swish"]:
+            raise ValueError(f"LigerTiledSwiGLUMLP requires SiLU/Swish activation, got {config.hidden_act}")
+
+    def _mlp_forward(self, module, x):
+        """Internal MLP forward function for tiled computation."""
+        gate = module.gate_proj(x)
+        up = module.up_proj(x)
+        return module.down_proj(LigerSiLUMulFunction.apply(gate, up))
+
+    def forward(self, x):
+        """
+        Forward pass with tiled computation.
+
+        Args:
+            x: Input tensor of shape [batch_size, seq_len, hidden_size]
+               or [seq_len, hidden_size]
+
+        Returns:
+            Output tensor of the same shape as input
+        """
+        compute_params = [
+            self.gate_proj.weight,
+            self.up_proj.weight,
+            self.down_proj.weight,
+        ]
+
+        return apply_tiled_mlp(
+            fn=self._mlp_forward,
+            mlp_module=self,
+            x=x,
+            num_shards=self.num_shards,
+            compute_params=compute_params,
+        )

{liger_kernel_nightly-0.6.3.dev20251106220336.dist-info → liger_kernel_nightly-0.6.3.dev20251118154655.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.3.dev20251106220336
+Version: 0.6.3.dev20251118154655
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.3.dev20251106220336.dist-info → liger_kernel_nightly-0.6.3.dev20251118154655.dist-info}/RECORD

@@ -38,11 +38,12 @@ liger_kernel/ops/rope.py,sha256=v-7JHRrv-5ImoROkpKfl30WwWI4qTa2tAl7zQeB4ml4,8956
 liger_kernel/ops/softmax.py,sha256=tgORx6MK1IDDtZKqGarj0IPIVjqAIEUXXYPiinhRdtI,5864
 liger_kernel/ops/sparsemax.py,sha256=AeWe1xgkHJFEKWTj2vu_0hj7LztGvjqXAps-QTpCY0U,5087
 liger_kernel/ops/swiglu.py,sha256=D7nd4u_LInwsIRNCDdY77lqnTz8-W5dJrpEAt8zEO_A,3033
+liger_kernel/ops/tiled_mlp.py,sha256=eyMFsFFgHch8a_6R6IYRG24_jqKg5GF_BQUoQuAG8SY,4529
 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=iV1X0gH1JXwgeb7AeY8Ryv7q3r44MLQvSvn79yIVDzw,9874
+liger_kernel/transformers/__init__.py,sha256=XX1ySRgZXeQe0or-6GNclAsNQG_VkABQlkwqpB1Wn8A,10090
 liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
 liger_kernel/transformers/cross_entropy.py,sha256=DMtHkKrVJDSsels7KgGQJqrXkEAd6Zopcdr-5oRmQgE,2010
 liger_kernel/transformers/dyt.py,sha256=i-4GPaMrl-jab9TVI5qN0-H9qycn_mCbV82ozU4nbmU,723
@@ -68,6 +69,7 @@ liger_kernel/transformers/rope.py,sha256=VMlDZI6zss9mLaLcN5XCE_ktmYRwAi_Eh4TIgO6
 liger_kernel/transformers/softmax.py,sha256=yadlAgE4V2JByMwrDDa2s5SUBp8Jgd57xwnVvAWoBaI,264
 liger_kernel/transformers/sparsemax.py,sha256=0lQA0UEOs4mu8CMruZ3VLhImxQVXJWhPsAKUsYA7vj8,403
 liger_kernel/transformers/swiglu.py,sha256=LZ8YeLIdv2k46JleZMjzubGk98smt6t780kSgcVLsQk,3454
+liger_kernel/transformers/tiled_mlp.py,sha256=J51-kpzwikDMMhT5bX-RZCKMaXBK6zZc1bhgRYTK5F0,4651
 liger_kernel/transformers/trainer_integration.py,sha256=W3ON51O5GkyzNJsItz0y5rKx-uy2f2cFfveZpqbUdhw,123
 liger_kernel/transformers/tvd.py,sha256=XrRfyJIqN6HFxXk8MYyFVZM1OLz3mtSbRZvWfZ_JerQ,450
 liger_kernel/transformers/experimental/__init__.py,sha256=oQqk-f32JYgWEP9DJCj6ty6bbJSGrdXsFDQFwGeX6vI,127
@@ -106,9 +108,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7H
 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-0.6.3.dev20251106220336.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.6.3.dev20251106220336.dist-info/METADATA,sha256=dy_9atp4YioeU8GBh82zuDxFpz-nYGyfStlvUa4RxwY,24777
-liger_kernel_nightly-0.6.3.dev20251106220336.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.6.3.dev20251106220336.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.6.3.dev20251106220336.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.6.3.dev20251106220336.dist-info/RECORD,,
+liger_kernel_nightly-0.6.3.dev20251118154655.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.6.3.dev20251118154655.dist-info/METADATA,sha256=vo51Vsu8KGOEZOqaFBhwOogKYG2kQBPJhMC4-KUpyeY,24777
+liger_kernel_nightly-0.6.3.dev20251118154655.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.6.3.dev20251118154655.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.6.3.dev20251118154655.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.6.3.dev20251118154655.dist-info/RECORD,,