liger-kernel 0.5.9__py3-none-any.whl → 0.5.10__py3-none-any.whl

liger_kernel/ops/sparsemax.py

@@ -0,0 +1,167 @@
+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")
+
+
+class LigerSparsemaxFunction(torch.autograd.Function):
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, x: torch.Tensor, dim: int):
+        if dim < 0:
+            dim += x.dim()
+        ctx.dim = 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)
+
+        BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+        out_flat = torch.empty_like(x_flat)
+        grid = (n_rows,)
+
+        x_sorted_flat = torch.sort(x_flat.float(), dim=-1, descending=True).values
+
+        _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,
+        )
+
+        ctx.save_for_backward(out_flat)
+        return out_flat.view_as(x_sw).transpose(dim, -1)
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_out: torch.Tensor):
+        (out_flat,) = ctx.saved_tensors
+        dim = ctx.dim
+
+        go_sw = grad_out.transpose(dim, -1).contiguous()
+        n_cols = go_sw.size(-1)
+        n_rows = go_sw.numel() // n_cols
+        go_flat = go_sw.view(n_rows, n_cols)
+
+        BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+        gi_flat = torch.empty_like(go_flat)
+        grid = (n_rows,)
+
+        _sparsemax_backward_kernel[grid](
+            out_flat,
+            go_flat,
+            gi_flat,
+            out_flat.stride(0),
+            n_cols,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+        )
+
+        return gi_flat.view_as(go_sw).transpose(dim, -1), None

liger_kernel/transformers/__init__.py

@@ -14,6 +14,7 @@ from liger_kernel.transformers.rms_norm import LigerRMSNorm  # noqa: F401
 from liger_kernel.transformers.rope import liger_rotary_pos_emb  # noqa: F401
 from liger_kernel.transformers.swiglu import LigerBlockSparseTop2MLP  # noqa: F401
 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.tvd import LigerTVDLoss  # noqa: F401
 
@@ -40,6 +41,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_5_vl  # noqa: F401
     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
 
 
 # Check if 'transformers' is installed
@@ -95,6 +97,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen2_5_vl",
         "apply_liger_kernel_to_qwen2_vl",
         "apply_liger_kernel_to_qwen3",
+        "apply_liger_kernel_to_qwen3_moe",
     }
 
     if name in monkey_patch_symbols:
@@ -118,6 +121,7 @@ __all__ = [
     "liger_rotary_pos_emb",
     "LigerBlockSparseTop2MLP",
     "LigerPhi3SwiGLUMLP",
+    "LigerQwen3MoeSwiGLUMLP",
     "LigerSwiGLUMLP",
     "LigerTVDLoss",
 ]
@@ -147,5 +151,6 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_qwen2_5_vl",
             "apply_liger_kernel_to_qwen2_vl",
             "apply_liger_kernel_to_qwen3",
+            "apply_liger_kernel_to_qwen3_moe",
         ]
     )

liger_kernel/transformers/dyt.py

@@ -5,16 +5,18 @@ from liger_kernel.ops.dyt import LigerDyTFunction
 
 
 class LigerDyT(nn.Module):
-    def __init__(self, hidden_size, init_alpha=0.5):
+    def __init__(self, hidden_size, beta=True, init_alpha=0.5):
         super().__init__()
         self.hidden_size = hidden_size
         self.init_alpha = init_alpha
         self.alpha = nn.Parameter(torch.ones(1) * init_alpha)
         self.gamma = nn.Parameter(torch.ones(hidden_size))
-        self.beta = nn.Parameter(torch.zeros(hidden_size))
+        self.beta = None
+        if beta:
+            self.beta = nn.Parameter(torch.zeros(hidden_size))
 
     def forward(self, x):
         return LigerDyTFunction.apply(x, self.alpha, self.gamma, self.beta)
 
     def extra_repr(self):
-        return f"{self.hidden_size}, init_alpha={self.init_alpha}"
+        return f"{self.hidden_size}, init_alpha={self.init_alpha}, beta={self.beta}"

liger_kernel/transformers/fsdp.py

@@ -0,0 +1,55 @@
+from typing import Any
+from typing import Callable
+
+from torch.distributed.fsdp import FullyShardedDataParallel
+
+
+class _FSDPForwardRedirection:
+    """
+    Modified based on
+    https://github.com/Lightning-AI/pytorch-lightning/blob/d3f9c83d6efa4f1def36aa6c199600946cdb9117/src/lightning/pytorch/strategies/strategy.py#L601-L648
+    Redirect a method call through FullyShardedDataParallel.forward so that the FSDP module's root pre-forward and
+    post-forward can be properly executed around the method call.
+    This is needed in cases where we call a submodule of a FSDP module. For instance, when we want to call only
+    the `LlamaModel` part out of a FSDP-wrapped `LlamaForCausalLM` to get the hidden states without involving
+    GPU-memory-heavy `lm_head` and cross entropy computation, doing this directly (i.e. `model.model.forward()`)
+    will not work because the first `nn.Embedding` layer is not independently wrapped as a FSDP module (because of
+    the transformer-based wrapping policy), and not calling it through FSDP root module forward will not all-gather
+    its parameter, thus resulting in "RuntimeError: 'weight' must be 2-D" error. Similarly, if we want to call just
+    the `lm_head` part of a model, we need this trick too to properly get its params all-gathered.
+    """
+
+    def __call__(
+        self,
+        wrapper_module: FullyShardedDataParallel,
+        method: Callable,
+        *args: Any,
+        **kwargs: Any,
+    ):
+        """Reroutes a method call through the `wrapper_module`'s `forward` method.
+        Args:
+            wrapper_module: The module that has `original_module` wrapped.
+            original_module: The module that was wrapped inside `wrapper_module`.
+            method_name: The name of the method that should be called on the `original_module` after inputs get
+                redirected through the `wrapper_module`'s `forward` method.
+            *args: The positional arguments to the method `method_name`. They will get passed to a patched
+                `forward` method instead.
+            **kwargs: The keyword arguments to the method `method_name`. They will get passed to a patched
+                `forward` method instead.
+        """
+        assert isinstance(wrapper_module, FullyShardedDataParallel)
+        original_module = wrapper_module._fsdp_wrapped_module
+        original_forward = original_module.forward
+
+        def wrapped_forward(*_args: Any, **_kwargs: Any) -> Any:
+            # Unpatch ourselves immediately before calling the method `method_name`
+            # because itself may want to call the real `forward`
+            original_module.forward = original_forward  # type: ignore[method-assign]
+            # Call the actual method e.g. `.training_step(...)`
+            out = method(*_args, **_kwargs)
+            return out
+
+        # Patch the original_module's forward so we can redirect the arguments back to the real method
+        original_module.forward = wrapped_forward  # type: ignore[method-assign]
+        wrapper_output = wrapper_module(*args, **kwargs)
+        return wrapper_output

liger_kernel/transformers/functional.py

@@ -12,6 +12,7 @@ from liger_kernel.ops.layer_norm import LigerLayerNormFunction
 from liger_kernel.ops.qwen2vl_mrope import LigerQwen2VLMRopeFunction
 from liger_kernel.ops.rms_norm import LigerRMSNormFunction
 from liger_kernel.ops.rope import LigerRopeFunction
+from liger_kernel.ops.sparsemax import LigerSparsemaxFunction
 from liger_kernel.ops.swiglu import LigerSiLUMulFunction
 from liger_kernel.ops.tvd import LigerTVDLossFunction
 
@@ -159,6 +160,13 @@ def liger_kl_div(
     )
 
 
+def liger_sparsemax(
+    input,
+    dim: int = -1,
+):
+    return LigerSparsemaxFunction.apply(input, dim)
+
+
 def liger_tvd(
     input,
     target,

liger_kernel/transformers/grpo_loss.py

@@ -0,0 +1,98 @@
+from liger_kernel.ops.grpo_loss import GrpoLossFunction
+
+
+def triton_grpo_loss(
+    logits,
+    old_logp,
+    ref_logp,
+    completion_ids,
+    advantages,
+    completion_mask=None,
+    temperature=0.9,
+    beta=0.04,
+    eps_low=0.2,
+    eps_high=0.4,
+    inplace=True,
+):
+    assert logits is not None and completion_ids is not None and advantages is not None, (
+        "must provide logits、completion_ids and advantages"
+    )
+
+    return GrpoLossFunction.apply(
+        logits,
+        old_logp,
+        ref_logp,
+        completion_ids,
+        advantages,
+        completion_mask,
+        temperature,
+        beta,
+        eps_low,
+        eps_high,
+        inplace,
+    )
+
+
+# This is a demo how to use grpo_loss in GRPOTrainer. The Trl version must be 0.16
+"""
+import torch
+import trl
+assert trl.__version__.startswith("0.16"), "please pip install trl==0.16"
+from trl.extras.profiling import profiling_decorator
+
+@profiling_decorator
+def _get_per_token_logps(self, model, input_ids, attention_mask, logits_to_keep):
+    # We add 1 to `logits_to_keep` because the last logits of the sequence is later excluded
+    logits = model(input_ids=input_ids, attention_mask=attention_mask, logits_to_keep=logits_to_keep + 1).logits
+    return fused_selective_log_softmax(logits, input_ids, self.temperature, mask=attention_mask)
+
+@profiling_decorator
+def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
+    if return_outputs:
+        raise ValueError("The GRPOTrainer does not support returning outputs")
+    # Compute the per-token log probabilities for the model
+
+    prompt_ids, prompt_mask = inputs["prompt_ids"], inputs["prompt_mask"]
+    completion_ids, completion_mask = inputs["completion_ids"], inputs["completion_mask"]
+    input_ids = torch.cat([prompt_ids, completion_ids], dim=1)
+    attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)
+    logits_to_keep = completion_ids.size(1)  # we only need to compute the logits for the completion tokens
+    logits = model(input_ids=input_ids, attention_mask=attention_mask, logits_to_keep=logits_to_keep + 1).logits
+
+    ref_per_token_logps = inputs["ref_per_token_logps"]
+    advantages = inputs["advantages"]
+    old_per_token_logps = inputs["old_per_token_logps"]
+    
+
+    per_token_loss, per_token_kl, is_clipped = triton_grpo_loss(logits, 
+                                                                old_per_token_logps,
+                                                                ref_per_token_logps,
+                                                                completion_ids,
+                                                                advantages,
+                                                                completion_mask,
+                                                                self.temperature,
+                                                                self.beta,
+                                                                self.epsilon_low,
+                                                                self.epsilon_high,)
+    loss = (per_token_loss * completion_mask).sum() / completion_mask.sum()
+
+    # Log the metrics
+    mode = "eval" if self.control.should_evaluate else "train"
+
+    if self.beta != 0.0:
+        mean_kl = (per_token_kl * completion_mask).sum() / completion_mask.sum()
+        self._metrics[mode]["kl"].append(self.accelerator.gather_for_metrics(mean_kl).mean().item())
+
+    clip_ratio = (is_clipped * completion_mask).sum() / completion_mask.sum()
+    self._metrics[mode]["clip_ratio"].append(self.accelerator.gather_for_metrics(clip_ratio).mean().item())
+    return loss
+
+trl.GRPOTrainer._get_per_token_logps = _get_per_token_logps
+trl.GRPOTrainer.compute_loss = compute_loss
+trigger = None
+"""
+
+# add this line at the first line of grpo.py in open-r1
+"""
+from liger_kernel.transformers.grpo_loss import trigger
+"""

liger_kernel/transformers/model/gemma.py

@@ -8,18 +8,12 @@ import torch
 from torch.nn import CrossEntropyLoss
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.gemma.modeling_gemma import _CONFIG_FOR_DOC
-from transformers.models.gemma.modeling_gemma import GEMMA_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-@add_start_docstrings_to_model_forward(GEMMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -129,8 +123,6 @@ def lce_forward_deprecated(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(GEMMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/gemma2.py

@@ -9,10 +9,6 @@ import torch
 from torch.nn import CrossEntropyLoss
 from transformers.cache_utils import HybridCache
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.gemma2.modeling_gemma2 import _CONFIG_FOR_DOC
-from transformers.models.gemma2.modeling_gemma2 import GEMMA2_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
@@ -136,8 +132,6 @@ def lce_forward_deprecated(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(GEMMA2_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/gemma3.py

@@ -9,13 +9,9 @@ import torch.nn as nn
 from transformers.cache_utils import Cache
 from transformers.cache_utils import HybridCache
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.gemma3.modeling_gemma3 import _CONFIG_FOR_DOC
-from transformers.models.gemma3.modeling_gemma3 import GEMMA3_INPUTS_DOCSTRING
 from transformers.models.gemma3.modeling_gemma3 import Gemma3CausalLMOutputWithPast
-from transformers.utils import add_start_docstrings_to_model_forward
 from transformers.utils import is_torchdynamo_compiling
 from transformers.utils import logging
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
@@ -25,8 +21,6 @@ logger = logging.get_logger(__name__)
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(GEMMA3_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def causal_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -141,8 +135,6 @@ def causal_forward(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(GEMMA3_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=Gemma3CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def multimodal_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/glm4.py

@@ -6,18 +6,12 @@ from typing import Union
 import torch
 
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.glm4.modeling_glm4 import _CONFIG_FOR_DOC
-from transformers.models.glm4.modeling_glm4 import GLM4_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(GLM4_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/llama.py

@@ -7,23 +7,23 @@ from typing import Union
 import torch
 import torch.nn.functional as F
 
+from torch.distributed.fsdp import FullyShardedDataParallel
 from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.llama.modeling_llama import _CONFIG_FOR_DOC
-from transformers.models.llama.modeling_llama import LLAMA_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
+from liger_kernel.transformers.fsdp import _FSDPForwardRedirection
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 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
+
 
-@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -137,8 +137,6 @@ def lce_forward_deprecated(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -221,12 +219,12 @@ def lce_forward(
     loss = None
     # if in training mode, don't materialize logits
     if self.training and (labels is not None or shift_labels is not None):
-        loss = LigerForCausalLMLoss(
+        loss = lce_maybe_trainable_lm_head(
+            self,
             hidden_states=kept_hidden_states,
-            lm_head_weight=self.lm_head.weight,
+            hidden_size=self.config.hidden_size,
             labels=labels,
             shift_labels=shift_labels,
-            hidden_size=self.config.hidden_size,
             **loss_kwargs,
         )
 
@@ -251,3 +249,50 @@ def lce_forward(
         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/model/llava.py

@@ -5,19 +5,13 @@ from typing import Union
 
 import torch
 
-from transformers.models.llava.modeling_llava import _CONFIG_FOR_DOC
-from transformers.models.llava.modeling_llava import LLAVA_INPUTS_DOCSTRING
 from transformers.models.llava.modeling_llava import LlavaCausalLMOutputWithPast
-from transformers.utils import add_start_docstrings_to_model_forward
 from transformers.utils import is_torchdynamo_compiling
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 
 
-@add_start_docstrings_to_model_forward(LLAVA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=LlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -210,9 +204,7 @@ def lce_forward_deprecated(
     )
 
 
-@add_start_docstrings_to_model_forward(LLAVA_INPUTS_DOCSTRING)
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@replace_return_docstrings(output_type=LlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

liger_kernel/transformers/model/mistral.py

@@ -7,18 +7,12 @@ import torch
 
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.mistral.modeling_mistral import _CONFIG_FOR_DOC
-from transformers.models.mistral.modeling_mistral import MISTRAL_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,