liger-kernel 0.3.1__py3-none-any.whl → 0.4.0__py3-none-any.whl

liger_kernel/ops/utils.py

@@ -12,13 +12,19 @@ Modifications made by Yanning Chen, 2024.
 
 import functools
 import importlib
+import operator
 from typing import Callable
 
 import torch
 import triton
+import triton.language as tl
 from packaging.version import Version
 
 
+def is_hip() -> bool:
+    return torch.version.hip is not None
+
+
 def ensure_contiguous(fn):
     @functools.wraps(fn)
     def wrapper(ctx, *args, **kwargs):
@@ -45,7 +51,7 @@ def calculate_settings(n):
 
     num_warps = 4
     if BLOCK_SIZE >= 32768:
-        num_warps = 32
+        num_warps = 32 if not is_hip() else 16
     elif BLOCK_SIZE >= 8192:
         num_warps = 16
     elif BLOCK_SIZE >= 2048:
@@ -60,3 +66,58 @@ def compare_version(package: str, operator: Callable, target: str):
         return False
     pkg_version = Version(pkg.__version__)
     return operator(pkg_version, Version(target))
+
+
+def get_amp_custom_fwd_bwd() -> Callable:
+    if compare_version("torch", operator.ge, "2.4.0"):
+        return (
+            functools.partial(torch.amp.custom_fwd, device_type="cuda"),
+            functools.partial(torch.amp.custom_bwd, device_type="cuda"),
+        )
+    return torch.cuda.amp.custom_fwd, torch.cuda.amp.custom_bwd
+
+
+amp_custom_fwd, amp_custom_bwd = get_amp_custom_fwd_bwd()
+
+
+torch_to_triton_dtype = {
+    torch.float32: tl.float32,
+    torch.float16: tl.float16,
+    torch.bfloat16: tl.bfloat16,
+}
+
+
+@triton.jit
+def element_mul_kernel(
+    X_ptr,
+    X_stride,
+    grad_output_ptr,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    This function multiplies each element of the tensor pointed by X_ptr with the value pointed by grad_output_ptr.
+    The multiplication is performed in-place on the tensor pointed by X_ptr.
+
+    Parameters:
+    X_ptr: Pointer to the input tensor.
+    X_stride (int): The stride of the input tensor.
+    grad_output_ptr: Pointer to the gradient output value.
+    n_cols (int): The number of columns in the input tensor.
+    BLOCK_SIZE (int): The block size for Triton operations.
+    """
+
+    # Get the program ID and convert it to int64 to avoid overflow
+    program_id = tl.program_id(0).to(tl.int64)
+
+    # Locate the start index
+    X_ptr += program_id * X_stride
+
+    # Load the gradient output value
+    grad_output = tl.load(grad_output_ptr)
+
+    # Perform the element-wise multiplication
+    for i in range(0, n_cols, BLOCK_SIZE):
+        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)

liger_kernel/transformers/__init__.py

@@ -5,7 +5,9 @@ from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss  # noq
 from liger_kernel.transformers.fused_linear_cross_entropy import (  # noqa: F401
     LigerFusedLinearCrossEntropyLoss,
 )
+from liger_kernel.transformers.fused_linear_jsd import LigerFusedLinearJSD  # noqa: F401
 from liger_kernel.transformers.geglu import LigerGEGLUMLP  # noqa: F401
+from liger_kernel.transformers.jsd import LigerJSD  # noqa: F401
 from liger_kernel.transformers.layer_norm import LigerLayerNorm  # noqa: F401
 from liger_kernel.transformers.monkey_patch import (  # noqa: F401
     _apply_liger_kernel,
@@ -15,6 +17,7 @@ from liger_kernel.transformers.monkey_patch import (  # noqa: F401
     apply_liger_kernel_to_llama,
     apply_liger_kernel_to_mistral,
     apply_liger_kernel_to_mixtral,
+    apply_liger_kernel_to_mllama,
     apply_liger_kernel_to_phi3,
     apply_liger_kernel_to_qwen2,
     apply_liger_kernel_to_qwen2_vl,

liger_kernel/transformers/functional.py

@@ -2,7 +2,9 @@ from liger_kernel.ops.cross_entropy import LigerCrossEntropyFunction
 from liger_kernel.ops.fused_linear_cross_entropy import (
     LigerFusedLinearCrossEntropyFunction,
 )
+from liger_kernel.ops.fused_linear_jsd import LigerFusedLinearJSDFunction
 from liger_kernel.ops.geglu import LigerGELUMulFunction
+from liger_kernel.ops.jsd import LigerJSDFunction
 from liger_kernel.ops.kl_div import LigerKLDivLossFunction
 from liger_kernel.ops.layer_norm import LigerLayerNormFunction
 from liger_kernel.ops.rms_norm import LigerRMSNormFunction
@@ -17,3 +19,5 @@ liger_rms_norm = LigerRMSNormFunction.apply
 liger_rope = LigerRopeFunction.apply
 liger_layer_norm = LigerLayerNormFunction.apply
 liger_kl_div = LigerKLDivLossFunction.apply
+liger_jsd = LigerJSDFunction.apply
+liger_fused_linear_jsd = LigerFusedLinearJSDFunction.apply

liger_kernel/transformers/fused_linear_jsd.py

@@ -0,0 +1,98 @@
+from typing import Optional
+
+import torch
+
+from liger_kernel.ops.fused_linear_jsd import LigerFusedLinearJSDFunction
+
+
+class LigerFusedLinearJSD(torch.nn.Module):
+    r"""Fusing the last linear layer with generalized JSD
+
+    Handle the forward and backward pass of the final linear layer via JSD by avoiding
+    the materialization of the large logits tensor.
+
+    Args:
+        jsd_beta (float): coefficient beta of generalized JSD in the open interval (0, 1). Default: `0.5`
+        ignore_index (int): The index to ignore in the target. Default: `-100`
+        temperature (float): temperature in softmax function to control the output probability distribution. Default: `1.0`
+
+    Shape:
+        - student_input: :math:`(BT, H)`, where B is batch size, T is sequence length, H is hidden dimension.
+        - student_weight: :math:`(V, H)`, where V is vocab size.
+        - teacher_input: :math:`(BT, H')`, where H' is hidden dimension of the teacher model.
+        - teacher_weight: :math:`(V, H')`, where hidden size H and H' can be different.
+        - shift_labels: :math:`(BT,)`
+        - Output: a scalar.
+
+    Examples:
+    ```python
+    >>> (B, T, H_s, H_t, V) = (2, 2, 3, 5, 10)
+    >>> fused_jsd = LigerFusedLinearJSD(jsd_beta=0.1, temperature=2.0)
+    >>> # generate inputs and weights
+    >>> student_input = torch.rand(B * T, H_s, device="cuda", requires_grad=True)
+    >>> student_lin = torch.nn.Linear(H_s, V, bias=False, device="cuda")
+    >>> # teacher input doesn't require grad, hidden_dim can be different from student's
+    >>> teacher_input = torch.rand(B * T, H_t, device="cuda")
+    >>> teacher_lin = torch.nn.Linear(H_t, V, bias=False, device="cuda")
+    >>> output = fused_jsd(student_input, student_lin.weight, teacher_input, teacher_lin.weight)
+    >>> output.backward()
+    >>>
+    >>> # Example with labels for supervised fine-tuning (SFT) context:
+    >>>
+    >>> # Assume hidden_states, lm_heads and corresponding labels are given
+    >>> student_lm_head = torch.nn.Linear(H_s, V, bias=False)
+    >>> student_hidden_states = torch.randn(B * T, H_s, requires_grad=True).log_softmax(dim=-1)
+    >>> teacher_lm_head = torch.nn.Linear(H_t, V, bias=False)
+    >>> teacher_hidden_states = torch.randn(B * T, H_t).log_softmax(dim=-1)
+    >>> labels = torch.randint(0, V, (B * T,), torch.long)
+    >>>
+    >>> # Shift so that tokens < n predict n
+    >>> shift_student_hidden_states = student_hidden_states[..., :-1, :].contiguous()
+    >>> shift_teacher_hidden_states = teacher_hidden_states[..., :-1, :].contiguous()
+    >>> shift_labels = labels[..., 1:].contiguous()
+    >>>
+    >>> # Flatten tokens
+    >>> shift_student_hidden_states = shift_student_hidden_states.view(-1, V)
+    >>> shift_teacher_hidden_states = shift_teacher_hidden_states.view(-1, V)
+    >>> shift_labels = shift_labels.view(-1)
+    >>>
+    >>> # Calculate loss
+    >>> loss_fct = LigerJSD(beta=0.1)
+    >>> loss = loss_fct(
+    >>>     shift_studetn_hidden_states,
+    >>>     student_lm_head.weight,
+    >>>     shift_teacher_hidden_states,
+    >>>     teacher_lm_head.weight,
+    >>>     shift_labels
+    >>> )
+    ```
+    """
+
+    def __init__(self, jsd_beta=0.5, ignore_index=-100, temperature=1.0):
+        super().__init__()
+        assert (
+            jsd_beta > 0 and jsd_beta < 1
+        ), f"beta must be greater than 0 and less than 1. Got: {jsd_beta}"
+        assert temperature != 0, "temperature cannot be 0."
+        self.jsd_beta = jsd_beta
+        self.temperature = temperature
+        self.ignore_index = ignore_index
+
+    def forward(
+        self,
+        student_input: torch.Tensor,
+        student_weight: torch.Tensor,
+        teacher_input: torch.Tensor,
+        teacher_weight: torch.Tensor,
+        shift_labels: Optional[torch.LongTensor],
+    ):
+        return LigerFusedLinearJSDFunction.apply(
+            student_input,
+            student_weight,
+            teacher_input,
+            teacher_weight,
+            shift_labels,
+            self.jsd_beta,
+            self.ignore_index,
+            self.temperature,
+        )

liger_kernel/transformers/jsd.py

@@ -0,0 +1,75 @@
+from typing import Optional
+
+import torch
+
+from liger_kernel.ops.jsd import LigerJSDFunction
+
+
+class LigerJSD(torch.nn.Module):
+    r"""The generalized Jensen-Shannon Divergence.
+    .. math::
+    JSD(\beta)(P || Q)
+        = \beta * KLDiv(P || (\beta * P + (1 - \beta) * Q)) + (1 - \beta) * KLDiv(Q || (\beta * P + (1 - \beta) * Q))
+    .. note::
+    As all the other losses in PyTorch, this function expects the first argument,
+    :attr:`log_q`, to be the predictions, the output of the student model in log-space,
+    and the second, :attr:`log_p`, to be the observations, the output of the teacher model in log-space.
+    This differs from the standard mathematical notation :math:`JSD(P || Q)` where
+    :math:`P` denotes the teacher model and :math:`Q` denotes the student model.
+
+    Args:
+        beta (float): coefficient beta of generalized JSD in the open interval (0, 1). Default: `0.5`
+        ignore_index (int): The index to ignore in the target. Default: `-100`
+
+    Shape:
+        - Input: :math:`(BT, V)`, where B is batch size, T is sequence length, V is vocab size.
+        - Target: :math:`(BT, V)`, same shape as the input.
+        - shift_labels (Optional): :math:`(BT,)`
+        - Output: a scalar.
+
+    Examples:
+    ```python
+    >>> (B, T, V) = (2, 2, 5)
+    >>> jsd = LigerJSD(beta=0.1)
+    >>> # input should be a distribution in the log space
+    >>> input = torch.randn(B * T, V, requires_grad=True).log_softmax(dim=-1)
+    >>> target = torch.randn(B * T, V).log_softmax(dim=-1)
+    >>> output = jsd(input, target)
+    >>>
+    >>> # Example with labels for supervised fine-tuning (SFT) context
+    >>> # Assume logits and corresponding labels are given
+    >>> student_logits = torch.randn(B * T, V, requires_grad=True).log_softmax(dim=-1)
+    >>> teacher_logits = torch.randn(B * T, V).log_softmax(dim=-1)
+    >>> labels = torch.randint(0, V, (B * T,), torch.long)
+    >>> # Shift so that tokens < n predict n
+    >>> shift_student_logits = student_logits[..., :-1, :].contiguous()
+    >>> shift_teacher_logits = teacher_logits[..., :-1, :].contiguous()
+    >>> shift_labels = labels[..., 1:].contiguous()
+    >>> # Flatten tokens
+    >>> shift_student_logits = shift_student_logits.view(-1, V)
+    >>> shift_teacher_logits = shift_teacher_logits.view(-1, V)
+    >>> shift_labels = shift_labels.view(-1)
+    >>> # Calculate loss
+    >>> loss_fct = LigerJSD(beta=0.1)
+    >>> loss = loss_fct(shift_studetn_logits, shift_teacher_logits, shift_labels)
+
+    ```
+    """
+
+    def __init__(self, beta: float = 0.5, ignore_index: int = -100):
+        super().__init__()
+        assert (
+            beta > 0 and beta < 1
+        ), f"beta must be greater than 0 and less than 1. Got: {beta}"
+        self.beta = beta
+        self.ignore_index = ignore_index
+
+    def forward(
+        self,
+        log_q: torch.Tensor,
+        log_p: torch.Tensor,
+        shift_labels: Optional[torch.LongTensor] = None,
+    ):
+        return LigerJSDFunction.apply(
+            log_q, log_p, shift_labels, self.beta, self.ignore_index
+        )

liger_kernel/transformers/model/gemma.py

@@ -22,7 +22,7 @@ from liger_kernel.transformers.fused_linear_cross_entropy import (
 @replace_return_docstrings(
     output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
 )
-def lce_forward(
+def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
     attention_mask: Optional[torch.Tensor] = None,
@@ -136,3 +136,126 @@ def lce_forward(
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
     )
+
+
+@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,
+    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,
+    num_logits_to_keep: int = 0,
+    **loss_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]`.
+
+        num_logits_to_keep (`int`, *optional*):
+            Calculate logits for the last `num_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.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, GemmaForCausalLM
+
+    >>> model = GemmaForCausalLM.from_pretrained("google/gemma-7b")
+    >>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
+
+    >>> prompt = "What is your favorite condiment?"
+    >>> 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]
+    "What is your favorite condiment?"
+    ```"""
+    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,
+    )
+
+    hidden_states = outputs[0]
+
+    logits = None
+    loss = None
+    # if in training mode, don't materialize logits
+    if self.training and (labels is not None):
+        # We do the same thing as ForCausalLMLoss but using Liger FLCE
+
+        shift_hidden_states = hidden_states[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+
+        # flatten tokens
+        shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
+        shift_labels = shift_labels.view(-1)
+
+        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
+        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
+
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        if reduction == "sum":
+            loss /= loss_kwargs["num_items_in_batch"]
+
+    else:  # if in inference mode materialize logits
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **loss_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,
+    )

liger_kernel/transformers/model/llama.py

@@ -1,4 +1,4 @@
-from typing import List, Optional, Tuple, Union
+from typing import TYPE_CHECKING, List, Optional, Tuple, Union
 
 import torch
 import torch.nn.functional as F
@@ -17,17 +17,20 @@ from liger_kernel.transformers.fused_linear_cross_entropy import (
     LigerFusedLinearCrossEntropyLoss,
 )
 
+if TYPE_CHECKING:
+    from transformers.cache_utils import Cache
+
 
 @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
 @replace_return_docstrings(
     output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
 )
-def lce_forward(
+def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
     attention_mask: Optional[torch.Tensor] = None,
     position_ids: Optional[torch.LongTensor] = None,
-    past_key_values: Optional[List[torch.FloatTensor]] = 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,
@@ -120,8 +123,9 @@ def lce_forward(
             logits = torch.cat(logits, dim=-1)
         else:
             logits = self.lm_head(hidden_states)
-        logits = logits.float()
         if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
             # Shift so that tokens < n predict n
             shift_logits = logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()
@@ -144,3 +148,130 @@ def lce_forward(
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
     )
+
+
+@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,
+    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,
+    num_logits_to_keep: int = 0,
+    **loss_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]`.
+
+        num_logits_to_keep (`int`, *optional*):
+            Calculate logits for the last `num_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.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, LlamaForCausalLM
+
+    >>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
+
+    >>> 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,
+    )
+
+    hidden_states = outputs[0]
+
+    if self.config.pretraining_tp > 1:
+        raise Exception("Liger Kernel does not support pretraining_tp!!")
+
+    logits = None
+    loss = None
+    # if in training mode, don't materialize logits
+    if self.training and (labels is not None):
+        # We do the same thing as ForCausalLMLoss but using Liger FLCE
+
+        shift_hidden_states = hidden_states[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+
+        # flatten tokens
+        shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
+        shift_labels = shift_labels.view(-1)
+
+        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
+        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
+
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        if reduction == "sum":
+            loss /= loss_kwargs["num_items_in_batch"]
+
+    else:  # if in inference mode materialize logits
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **loss_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,
+    )

liger_kernel/transformers/model/mistral.py

@@ -136,3 +136,6 @@ def lce_forward(
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
     )
+
+
+# Note: Grad Acc is not fixed in mistral at transformer 4.46.1