Upload modeling_openmoe.py

modeling_openmoe.py

@@ -0,0 +1,1113 @@
+# coding=utf-8
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch OpenMoE model."""
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.models.llama.modeling_llama import LlamaConfig, LlamaRMSNorm
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+
+from colossalai.kernel.cuda_native.mha.flash_attn_2 import HAS_FLASH_ATTN
+from colossalai.kernel.triton.llama_act_combine_kernel import HAS_TRITON
+from colossalai.moe.layers import SparseMLP
+from colossalai.moe.manager import MOE_MANAGER
+from colossalai.moe.utils import get_activation, set_moe_args
+
+if HAS_TRITON:
+    from colossalai.kernel.triton.llama_act_combine_kernel import LlamaActCombine
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LlamaConfig"
+
+class LlamaRotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+        # Build here to make `torch.jit.trace` work.
+        self._set_cos_sin_cache(
+            seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
+        )
+
+    def _set_cos_sin_cache(self, seq_len, device, dtype):
+        self.max_seq_len_cached = seq_len
+        t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
+
+        freqs = torch.outer(t, self.inv_freq)  # (seq_len, dim//2)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)  # (seq_len, dim)
+        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if seq_len > self.max_seq_len_cached:
+            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
+
+        return (
+            self.cos_cached[:seq_len].to(dtype=x.dtype),
+            self.sin_cached[:seq_len].to(dtype=x.dtype),
+        )
+
+def set_openmoe_args(
+    config: LlamaConfig,
+    num_experts: int,
+    moe_layer_interval: int,
+    router_topk: int = 2,
+    router_capacity_factor_train: float = 1.25,
+    router_capacity_factor_eval: float = 2.0,
+    router_min_capacity: int = 4,
+    router_noisy_policy: str = None,
+    router_drop_tks: bool = True,
+    router_aux_loss_factor: float = 0.01,
+    router_z_loss_factor: float = 0.0001,
+    mlp_gated: bool = True,
+    label_smoothing: float = 0.001,
+    z_loss_factor: float = 0.01,
+    enable_load_balance: bool = False,
+    load_balance_tolerance: float = 0.1,
+    load_balance_beam_width: int = 8,
+    load_balance_group_swap_factor: float = 0.4,
+    enable_kernel: bool = False,
+    enable_comm_overlap: bool = False,
+    enable_hierarchical_alltoall: bool = False,
+) -> None:
+    """
+    MoE related arguments.
+    It inserts the MoE arguments into the Llama config.
+
+    Args:
+        config (LlamaConfig): Transformers Llama config.
+        num_experts (int, optional): Number of experts.
+        moe_layer_interval (int, optional): The interval moe layer.
+        router_topk (int, optional): Moe router top k. Defaults to 2.
+        router_capacity_factor_train (float, optional): Moe router max capacity for train. Defaults to 1.25.
+        router_capacity_factor_eval (float, optional): Moe router max capacity for eval. Defaults to 2.0.
+        router_min_capacity (int, optional): Moe router min capacity. Defaults to 4.
+        router_noisy_policy (str, optional): Moe router noisy policy. You can choose [Jitter, Gaussian, None]. Defaults to None.
+        router_drop_tks (bool, optional): Whether moe router drop tokens which exceed max capacity. Defaults to True.
+        router_aux_loss_factor (float, optional): Moe router aux loss. You can refer to STMoE for details. Defaults to 0.01.
+        router_z_loss_factor (float, optional): Moe router z loss. You can refer to STMoE for details. Defaults to 0.01.
+        mlp_gated (bool, optional): Use gate in mlp. Defaults to True.
+        label_smoothing (float, optional): Label smoothing. Defaults to 0.001.
+        z_loss_factor (float, optional): The final outputs' classification z loss factor. Defaults to 0.01.
+        enable_load_balance (bool, optional): Expert load balance. Defaults to False.
+        load_balance_tolerance (float, optional): Expert load balance search's difference tolerance. Defaults to 0.1.
+        load_balance_beam_width (int, optional): Expert load balance search's beam width. Defaults to 8.
+        load_balance_group_swap_factor (float, optional): Expert load balance group swap factor. Longer value encourages less swap. Defaults to 0.4.
+        enable_kernel (bool, optional): Use kernel optimization. Defaults to False.
+        enable_comm_overlap (bool, optional): Use communication overlap for MoE. Recommended to enable for muiti-node training. Defaults to False.
+        enable_hierarchical_alltoall (bool, optional): Use hierarchical alltoall for MoE. Defaults to False.
+    """
+    moe_args = dict(
+        num_experts=num_experts,
+        moe_layer_interval=moe_layer_interval,
+        router_topk=router_topk,
+        router_capacity_factor_train=router_capacity_factor_train,
+        router_capacity_factor_eval=router_capacity_factor_eval,
+        router_min_capacity=router_min_capacity,
+        router_noisy_policy=router_noisy_policy,
+        router_drop_tks=router_drop_tks,
+        router_aux_loss_factor=router_aux_loss_factor,
+        router_z_loss_factor=router_z_loss_factor,
+        mlp_gated=mlp_gated,
+        label_smoothing=label_smoothing,
+        z_loss_factor=z_loss_factor,
+        enable_load_balance=enable_load_balance,
+        load_balance_tolerance=load_balance_tolerance,
+        load_balance_beam_width=load_balance_beam_width,
+        load_balance_group_swap_factor=load_balance_group_swap_factor,
+        enable_kernel=enable_kernel,
+        enable_comm_overlap=enable_comm_overlap,
+        enable_hierarchical_alltoall=enable_hierarchical_alltoall,
+    )
+    set_moe_args(config, moe_args)
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos[position_ids].unsqueeze(unsqueeze_dim)
+    sin = sin[position_ids].unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+def SwiGLU(x):
+    """Gated linear unit activation function.
+    Args:
+        x : input array
+        axis: the axis along which the split should be computed (default: -1)
+    """
+    size = x.shape[-1]
+    assert size % 2 == 0, "axis size must be divisible by 2"
+    x1, x2 = torch.split(x, size // 2, -1)
+    return x1 * (x2 * torch.sigmoid(x2))
+
+
+class OpenMoeMLP(nn.Module):
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.pretraining_tp = config.pretraining_tp
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size * 2, 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)
+        self.hidden_act = config.hidden_act
+        self.act_fn = get_activation(self.hidden_act)
+        self.use_kernel = config.enable_kernel
+
+    def forward(self, x):
+        if self.pretraining_tp > 1:
+            slice = self.intermediate_size // self.pretraining_tp
+            gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
+            up_proj_slices = self.up_proj.weight.split(slice, dim=0)
+            down_proj_slices = self.down_proj.weight.split(slice, dim=1)
+
+            gate_proj = torch.cat([F.linear(x, gate_proj_slices[i]) for i in range(self.pretraining_tp)], dim=-1)
+            up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.pretraining_tp)], dim=-1)
+
+            intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
+            down_proj = [F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.pretraining_tp)]
+            down_proj = sum(down_proj)
+        else:
+            if HAS_TRITON and self.use_kernel and self.hidden_act == "swiglu":
+                down_proj = self.down_proj(LlamaActCombine.apply(self.gate_proj(x), self.up_proj(x)))
+            else:
+                down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+        return down_proj
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class OpenMoeAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = config.head_dim
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.pretraining_tp = config.pretraining_tp
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self._init_rope()
+    
+    def _init_rope(self):
+        if self.config.rope_scaling is None:
+            self.rotary_emb = LlamaRotaryEmbedding(
+                self.head_dim,
+                max_position_embeddings=self.max_position_embeddings,
+                base=self.rope_theta,
+            )
+        else:
+            raise ValueError(f"Only Original RotaryEmbedding is supported yet")
+            
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        use_kernel: bool = True,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        if self.pretraining_tp > 1:
+            key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.pretraining_tp
+            query_slices = self.q_proj.weight.split((self.num_heads * self.head_dim) // self.pretraining_tp, dim=0)
+            key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+            value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+            query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.pretraining_tp)]
+            query_states = torch.cat(query_states, dim=-1)
+
+            key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.pretraining_tp)]
+            key_states = torch.cat(key_states, dim=-1)
+
+            value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.pretraining_tp)]
+            value_states = torch.cat(value_states, dim=-1)
+
+        else:
+            query_states = self.q_proj(hidden_states)
+            key_states = self.k_proj(hidden_states)
+            value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)  # (bsz, num_heads, q_len, head_dim)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)  # (bsz, num_heads, q_len, head_dim)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)  # (bsz, num_heads, q_len, head_dim)
+
+        kv_seq_len = key_states.shape[-2]  
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]  
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)  # (bsz, num_heads, q_len+past_kv_len, head_dim)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)  # (bsz, num_heads, q_len+past_kv_len, head_dim)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        if HAS_FLASH_ATTN and use_kernel:
+            from flash_attn import flash_attn_func
+
+            query_states = query_states.transpose(1, 2)
+            key_states = key_states.transpose(1, 2)
+            value_states = value_states.transpose(1, 2)
+            attn_output = flash_attn_func(query_states, key_states, value_states, softmax_scale=1.0, causal=True)
+            attn_output = attn_output.transpose(1, 2).contiguous()
+        else:
+            attn_weights = torch.matmul(query_states, key_states.transpose(2, 3))
+
+            if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                    f" {attn_weights.size()}"
+                )
+
+            if attention_mask is not None:
+                if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                    raise ValueError(
+                        f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                    )
+                if self.training:
+                    attention_mask = attention_mask.clone().detach()
+                attention_mask[:, :, :, 0] = 0
+                attn_weights = attn_weights + attention_mask
+
+            # upcast attention to fp32
+            attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+            attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim)
+
+        if self.pretraining_tp > 1:
+            attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
+            o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.pretraining_tp, dim=1)
+            attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.pretraining_tp)])
+        else:
+            attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class OpenMoeDecoderLayer(nn.Module):
+    def __init__(self, config: LlamaConfig, moe: bool):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.moe = moe
+        self.self_attn = OpenMoeAttention(config=config)
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        if self.moe:
+            self.mlp = SparseMLP(
+                num_experts=config.num_experts,
+                hidden_size=config.hidden_size,
+                intermediate_size=config.intermediate_size,
+                router_top_k=config.router_topk,
+                router_capacity_factor_train=config.router_capacity_factor_train,
+                router_capacity_factor_eval=config.router_capacity_factor_eval,
+                router_min_capacity=config.router_min_capacity,
+                router_noisy_policy=config.router_noisy_policy,
+                router_drop_tks=config.router_drop_tks,
+                mlp_activation=config.hidden_act,
+                mlp_gated=config.mlp_gated,
+                enable_load_balance=config.enable_load_balance,
+                load_balance_tolerance=config.load_balance_tolerance,
+                load_balance_beam_width=config.load_balance_beam_width,
+                load_balance_group_swap_factor=config.load_balance_group_swap_factor,
+                enable_kernel=config.enable_kernel,
+                enable_comm_overlap=config.enable_comm_overlap,
+            )
+            self.pre_extra_mlp_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+            self.extra_mlp = OpenMoeMLP(config)
+        else:
+            self.mlp = OpenMoeMLP(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        if self.moe:
+            residual = hidden_states
+            hidden_states = self.pre_extra_mlp_layernorm(hidden_states)
+            hidden_states = self.extra_mlp(hidden_states)
+            hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+LLAMA_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`LlamaConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class OpenMoePreTrainedModel(PreTrainedModel):
+    config_class = LlamaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlamaDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, OpenMoeModel):
+            module.gradient_checkpointing = value
+
+
+LLAMA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class OpenMoeModel(OpenMoePreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
+
+    Args:
+        config: LlamaConfig
+    """
+
+    def __init__(self, config: LlamaConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [
+                OpenMoeDecoderLayer(config, moe=True if (i + 1) % config.moe_layer_interval == 0 else False)
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
+                inputs_embeds.device
+            )
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    def forward(
+        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,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        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
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
+            )
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+        )
+
+        hidden_states = inputs_embeds
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    None,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class OpenMoeForCausalLM(OpenMoePreTrainedModel):
+    # _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = OpenMoeModel(config)
+        self.pretraining_tp = config.pretraining_tp
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        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,
+        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,
+        chunk_head: Optional[bool] = True,
+    ) -> 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]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, LlamaForCausalLM
+
+        >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> 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."
+        ```"""
+        # reset moe loss
+        MOE_MANAGER.reset_loss()
+
+        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,
+        )
+
+        hidden_states = outputs[0]
+        if self.pretraining_tp > 1:
+            lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.pretraining_tp, dim=0)
+            logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.pretraining_tp)]
+            logits = torch.cat(logits, dim=-1)
+
+        loss = None
+        # if no training, just do forward
+        if labels is None:
+            logits = self.lm_head(hidden_states)
+            logits = logits.float()
+        # the vocab size for openmoe is 30w+
+        # which causes great activation memory in training, up to 20G for one sequence
+        # so we use chunk and checkpoint to reduce memory
+        else:
+            if chunk_head == True:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        logits = module(inputs[0])
+                        logits = logits.float()
+                        # Shift so that tokens < n predict n
+                        shift_logits = logits[..., :-1, :].contiguous().float()
+                        shift_labels = inputs[1][..., 1:].contiguous()
+                        # Flatten the tokens
+                        loss = self._calculate_loss(shift_logits, shift_labels)
+                        return loss
+
+                    return custom_forward
+
+                aux_loss, z_loss = self._calculate_router_loss()
+                loss = aux_loss + z_loss
+                for batch_idx in range(hidden_states.shape[0]):
+                    loss = loss + torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(self.lm_head),
+                        hidden_states[batch_idx : batch_idx + 1, :],
+                        labels[batch_idx : batch_idx + 1, :],
+                    )
+                logits = None
+            else:
+                logits = self.lm_head(hidden_states)
+                logits = logits.float()
+                # Shift so that tokens < n predict n
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+                # Flatten the tokens
+                aux_loss, z_loss = self._calculate_router_loss()
+                loss = aux_loss + z_loss
+                loss = loss + self._calculate_loss(shift_logits, shift_labels)
+
+        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 prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+    def _calculate_router_loss(self, aux_loss: list = None, z_loss: list = None):
+        if aux_loss is None or z_loss is None:
+            aux_loss, z_loss = MOE_MANAGER.get_loss()
+        assert len(aux_loss) == len(z_loss) == self.config.num_hidden_layers // self.config.moe_layer_interval
+        aux_loss = self.config.router_aux_loss_factor * sum(aux_loss) / len(aux_loss)
+        z_loss = self.config.router_z_loss_factor * sum(z_loss) / len(z_loss)
+        return aux_loss, z_loss
+
+    def _calculate_loss(self, logits: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
+        """Compute cross entropy and entropy for log probs and targets.
+
+        Args:
+            logits: [batch, length, num_classes] float array.
+            targets: categorical targets [batch, length] int array.
+
+        Returns:
+            Tuple of scalar loss.
+        """
+        if len(logits.shape) != len(targets.shape) + 1:
+            raise ValueError(
+                "Incorrect shapes. Got shape %s logits and %s targets" % (str(logits.shape), str(targets.shape))
+            )
+        vocab_size = logits.shape[-1]
+        confidence = 1.0 - self.config.label_smoothing
+        low_confidence = (1.0 - confidence) / (vocab_size - 1)
+        normalizing_constant = -(
+            confidence * math.log(confidence) + (vocab_size - 1) * low_confidence * math.log(low_confidence + 1e-20)
+        )
+
+        # one hot
+        soft_targets = targets[..., None] == torch.arange(vocab_size, device=targets.device).reshape(
+            (1,) * len(targets.shape) + (-1,)
+        )
+        soft_targets = torch.where(
+            soft_targets, torch.full_like(soft_targets, confidence), torch.full_like(soft_targets, low_confidence)
+        )
+        soft_targets = soft_targets.to(torch.float32)
+
+        # cross entropy
+        total_loss = ZLossCrossEntropy.apply(logits, soft_targets, self.config.z_loss_factor)
+        total_loss = total_loss - normalizing_constant
+        total_loss = torch.mean(torch.sum(total_loss, dim=-1), dim=0)
+        return total_loss
+
+
+class ZLossCrossEntropy(torch.autograd.Function):
+    """Computes cross entropy loss with stable custom gradient.
+
+    Computes a stabilized-gradient version of:
+        -jnp.sum(targets * nn.log_softmax(logits), axis=-1)
+
+    If z_loss > 0, then an auxiliary loss equal to z_loss*log(z)^2
+    will be added to the cross entropy loss (z = softmax normalization constant).
+    The two uses of z_loss are:
+    1. To keep the logits from drifting too far from zero, which can cause
+        unacceptable roundoff errors in bfloat16.
+    2. To encourage the logits to be normalized log-probabilities.
+
+    Args:
+        logits: [batch, length, num_classes] float array.
+        targets: categorical one-hot targets [batch, length, num_classes] float
+        array.
+        z_loss: coefficient for auxilliary z-loss loss term.
+
+    Returns:
+        tuple with the total loss and the z_loss, both
+        float arrays with shape [batch, length].
+    """
+
+    @staticmethod
+    def forward(ctx, logits, targets, z_loss):
+        max_logit = torch.max(logits, dim=-1, keepdim=True)[0]
+        shifted = logits - max_logit
+        exp_shifted = torch.exp(shifted)
+        sum_exp = torch.sum(exp_shifted, axis=-1, keepdims=True)
+        sum_exp_log = torch.log(sum_exp)
+        log_softmax = shifted - sum_exp_log
+        loss = -torch.sum(targets * log_softmax, axis=-1)
+        # Add auxilliary z-loss term.
+        log_z = torch.squeeze(sum_exp_log + max_logit, axis=-1)
+        total_z_loss = z_loss * torch.square(log_z)
+        loss += total_z_loss
+        ctx.z_loss = z_loss
+        ctx.save_for_backward(logits, targets, exp_shifted, sum_exp, log_softmax, log_z)
+        return loss
+
+    @staticmethod
+    def backward(ctx, *grad_outputs):
+        assert len(grad_outputs) == 1
+        g = grad_outputs[0]
+        z_loss = ctx.z_loss
+        logits, targets, exp_shifted, sum_exp, log_softmax, log_z = ctx.saved_tensors
+        # z-loss term adds the (2 * z_loss * log_z) factor.
+        deriv = (1 + 2 * z_loss * log_z).unsqueeze(-1) * exp_shifted / sum_exp - targets
+        g_logits = g.unsqueeze(-1) * deriv
+        g_targets = -g.unsqueeze(-1) * log_softmax
+
+        return (
+            g_logits.to(logits.dtype),
+            g_targets.to(targets.dtype),
+            None,
+        )