Update modeling_Llamoe.py

modeling_Llamoe.py

@@ -4,967 +4,1374 @@
 # Copyright (c) 2022, Tri Dao, [email protected].
 # Licensed under the BSD 3-Clause License.
 
-from __future__ import annotations
-
+import inspect
 import math
-from dataclasses import dataclass, field
-from typing import Any, Dict, Optional, Tuple, Union
+import warnings
+from typing import List, Optional, Tuple, Union
 
 import torch
-import torch.nn as nn
-from einops import rearrange, repeat
-from transformers import PretrainedConfig, PreTrainedModel
-from transformers.activations import ACT2FN
-from transformers.modeling_outputs import CausalLMOutputWithPast
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import (
+    _prepare_4d_causal_attention_mask,
+    _prepare_4d_causal_attention_mask_for_sdpa,
+)
+from transformers.modeling_outputs import (
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+    SequenceClassifierOutputWithPast,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_13
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils.import_utils import is_torch_fx_available
 from .configuration_Llamoe import LlamoeConfig
 
-try:
-    from flash_attn.bert_padding import pad_input, unpad_input
-    from flash_attn.layers.rotary import RotaryEmbedding as FlashRotaryEmbedding
-    from flash_attn.modules.mha import FlashCrossAttention, FlashSelfAttention
-    from flash_attn.ops.fused_dense import FusedDense
-except:
-    pad_input, unpad_input = None, None
-    FlashRotaryEmbedding = None
-    FlashSelfAttention, FlashCrossAttention = None, None
-    FusedDense = None
-
-
-@dataclass
-class InferenceParams:
-    """Inference parameters passed to model to efficiently calculate
-    and store context during inference.
-    Reference:
-        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/utils/generation.py.
-    Args:
-        max_seqlen: Maximum sequence length.
-        max_batch_size: Maximum batch size.
-        seqlen_offset: Sequence length offset.
-        batch_size_offset: Batch size offset.
-        key_value_memory_dict: Key value memory dictionary.
-        lengths_per_sample: Lengths per sample.
-    """
-
-    max_seqlen: int = field(metadata={"help": "Maximum sequence length."})
 
-    max_batch_size: int = field(metadata={"help": "Maximum batch size."})
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
 
-    seqlen_offset: int = field(default=0, metadata={"help": "Sequence length offset."})
+    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
 
-    batch_size_offset: int = field(default=0, metadata={"help": "Batch size offset."})
+# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
+# It means that the function will not be traced through and simply appear as a node in the graph.
+if is_torch_fx_available():
+    if not is_torch_greater_or_equal_than_1_13:
+        import torch.fx
 
-    key_value_memory_dict: Dict[str, Any] = field(
-        default_factory=dict, metadata={"help": "Key value memory dictionary."}
-    )
+    _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
 
-    lengths_per_sample: torch.Tensor = field(default=None, metadata={"help": "Lengths per sample."})
 
+logger = logging.get_logger(__name__)
 
-class Embedding(nn.Module):
-    """Token embedding with dropout."""
+_CONFIG_FOR_DOC = "LlamoeConfig"
 
-    def __init__(self, config: LlamoeConfig) -> None:
-        super().__init__()
 
-        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
-        self.drop = nn.Dropout(config.embd_pdrop)
+def load_balancing_loss_func(
+    gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2, attention_mask: Optional[torch.Tensor] = None
+) -> float:
+    r"""
+    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
 
-    def forward(self, input_ids: torch.LongTensor) -> torch.FloatTensor:
-        input_shape = input_ids.size()
-        input_ids = input_ids.view(-1, input_shape[-1])
+    See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss
+    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
+    experts is too unbalanced.
 
-        hidden_states = self.wte(input_ids)
-        hidden_states = self.drop(hidden_states)
+    Args:
+        gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):
+            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
+            shape [batch_size X sequence_length, num_experts].
+        attention_mask (`torch.Tensor`, None):
+            The attention_mask used in forward function
+            shape [batch_size X sequence_length] if not None.
+        num_experts (`int`, *optional*):
+            Number of experts
+
+    Returns:
+        The auxiliary loss.
+    """
+    if gate_logits is None or not isinstance(gate_logits, tuple):
+        return 0
 
-        return hidden_states
+    if isinstance(gate_logits, tuple):
+        compute_device = gate_logits[0].device
+        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
 
+    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
 
-def _apply_rotary_emb(
-    x: torch.FloatTensor,
-    cos: torch.FloatTensor,
-    sin: torch.FloatTensor,
-) -> torch.FloatTensor:
-    _, seqlen, _, _ = x.shape
-    _, rotary_dim = cos.shape
-    rotary_dim *= 2
+    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
 
-    x_rot = x[:, :, :, :rotary_dim]
-    x_pass = x[:, :, :, rotary_dim:]
+    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
 
-    x1, x2 = x_rot.chunk(2, dim=-1)
-    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
-    x1, x2, c, s = [t.to(dtype=torch.float32) for t in [x1, x2, c, s]]
+    if attention_mask is None:
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
 
-    x_rot = torch.cat([x1 * c - x2 * s, x1 * s + x2 * c], axis=-1).to(x.dtype)
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.mean(routing_weights, dim=0)
+    else:
+        batch_size, sequence_length = attention_mask.shape
+        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
 
-    return torch.cat([x_rot, x_pass], axis=-1)
+        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
+        expert_attention_mask = (
+            attention_mask[None, :, :, None, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
+            .reshape(-1, top_k, num_experts)
+            .to(compute_device)
+        )
 
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
+            expert_attention_mask, dim=0
+        )
 
-def _apply_rotary_emb_kv(
-    kv: torch.FloatTensor,
-    cos: torch.FloatTensor,
-    sin: torch.FloatTensor,
-    cos_k: Optional[torch.FloatTensor] = None,
-    sin_k: Optional[torch.FloatTensor] = None,
-) -> torch.FloatTensor:
-    _, seqlen, _, _, _ = kv.shape
-    _, rotary_dim = cos.shape
-    rotary_dim *= 2
+        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
+        router_per_expert_attention_mask = (
+            attention_mask[None, :, :, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
+            .reshape(-1, num_experts)
+            .to(compute_device)
+        )
 
-    k_rot = kv[:, :, 0, :, :rotary_dim]
-    k_pass = kv[:, :, 0, :, rotary_dim:]
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
+            router_per_expert_attention_mask, dim=0
+        )
 
-    k1, k2 = k_rot.chunk(2, dim=-1)
-    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
-    k1, k2, c, s = [t.to(dtype=torch.float32) for t in [k1, k2, c, s]]
+    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
+    return overall_loss * num_experts
 
-    k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(kv.dtype)
 
-    return torch.cat(
-        [
-            torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
-            kv[:, :, 1:2, :, :],
-        ],
-        axis=2,
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
     )
 
 
-def _apply_rotary_emb_qkv(
-    qkv: torch.FloatTensor,
-    cos: torch.FloatTensor,
-    sin: torch.FloatTensor,
-    cos_k: Optional[torch.FloatTensor] = None,
-    sin_k: Optional[torch.FloatTensor] = None,
-) -> torch.FloatTensor:
-    _, seqlen, _, _, _ = qkv.shape
-    _, rotary_dim = cos.shape
-    rotary_dim *= 2
-
-    q_rot = qkv[:, :, 0, :, :rotary_dim]
-    q_pass = qkv[:, :, 0, :, rotary_dim:]
-
-    k_rot = qkv[:, :, 1, :, :rotary_dim]
-    k_pass = qkv[:, :, 1, :, rotary_dim:]
-
-    q1, q2 = q_rot.chunk(2, dim=-1)
-    k1, k2 = k_rot.chunk(2, dim=-1)
-    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
-    q1, q2, k1, k2, c, s = [t.to(dtype=torch.float32) for t in [q1, q2, k1, k2, c, s]]
-
-    q_rot = torch.cat([q1 * c - q2 * s, q1 * s + q2 * c], axis=-1).to(qkv.dtype)
-    k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(qkv.dtype)
-
-    return torch.cat(
-        [
-            torch.cat([q_rot, q_pass], axis=-1).unsqueeze(2),
-            torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
-            qkv[:, :, 2:3, :, :],
-        ],
-        axis=2,
-    )
+# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Mixtral
+class LlamoeRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        MixtralRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
 
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
 
-class RotaryEmbedding(nn.Module):
-    """Rotary positional embedding (RoPE).
-    Reference:
-        RoFormer: Enhanced Transformer with Rotary Position Embedding.
-        https://arxiv.org/pdf/2104.09864.pdf.
-    """
 
-    def __init__(
-        self,
-        dim: int,
-        base: int = 10000,
-        scale_base: Optional[float] = None,
-        pos_idx_in_fp32: bool = True,
-        max_position_embeddings: int = 2048,
-        device: Optional[str] = None,
-        **kwargs,
-    ) -> None:
+# Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->Mixtral
+class LlamoeRotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()
 
-        if scale_base is not None:
-            raise NotImplementedError
-
         self.dim = dim
-        self.base = float(base)
-        self.scale_base = scale_base
-        self.pos_idx_in_fp32 = pos_idx_in_fp32
         self.max_position_embeddings = max_position_embeddings
-        self.device = device
-
-        # Generate and save the inverse frequency buffer (non-trainable)
-        inv_freq = self._compute_inv_freq(device)
+        self.base = base
+        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
         self.register_buffer("inv_freq", inv_freq, persistent=False)
 
-        # Generate and save the scale buffer (non-trainable)
-        scale = (
-            (torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
-            if scale_base is not None
-            else None
+        # 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()
         )
-        self.register_buffer("scale", scale, persistent=False)
-
-        # Initialize cached attributes since ONNX can't rely on dynamic initialization
-        self._update_cos_sin_cache(max_position_embeddings, device=device, dtype=torch.float32)
-
-    def _compute_inv_freq(self, device: Optional[str] = None) -> torch.FloatTensor:
-        return 1.0 / (self.base ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim))
-
-    def _update_cos_sin_cache(
-        self,
-        seqlen: int,
-        device: Optional[str] = None,
-        dtype: Optional[torch.dtype] = None,
-    ) -> None:
-        self._seq_len_cached = seqlen
-
-        # fp32 is preferred since the output of `torch.arange` can be quite large
-        # and bf16 would lose a lot of precision
-        if self.pos_idx_in_fp32:
-            t = torch.arange(seqlen, device=device, dtype=torch.float32)
-            if self.inv_freq.dtype != torch.float32:
-                inv_freq = self._compute_inv_freq(device=device)
-            else:
-                inv_freq = self.inv_freq
-        else:
-            t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
-            inv_freq = self.inv_freq
-
-        # `torch.outer` is preferred since `torch.einsum` converts from fp32 to fp16 if used with AMP
-        freqs = torch.outer(t, inv_freq)
-        if self.scale is None:
-            self._cos_cached = torch.cos(freqs).to(dtype)
-            self._sin_cached = torch.sin(freqs).to(dtype)
-        else:
-            power = (
-                torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device) - seqlen // 2
-            ) / self.scale_base
-            scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")
 
-            # Force the scale multiplication to happen in fp32
-            self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
-            self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
-            self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
-            self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
-
-    def forward(
-        self,
-        qkv: torch.Tensor,
-        kv: Optional[torch.Tensor] = None,
-        seqlen_offset: int = 0,
-        **kwargs,
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
-        if (
-            self._seq_len_cached < qkv.shape[1] + seqlen_offset
-            or self._cos_cached.device != qkv.device
-            or self._cos_cached.dtype != qkv.dtype
-            or (self.training and self._cos_cached.is_inference())
-        ):
-            self._update_cos_sin_cache(qkv.shape[1] + seqlen_offset, device=qkv.device, dtype=qkv.dtype)
-
-        if kv is None:
-            return _apply_rotary_emb_qkv(
-                qkv,
-                self._cos_cached[seqlen_offset:],
-                self._sin_cached[seqlen_offset:],
-            )
-        else:
-            q = _apply_rotary_emb(
-                qkv,
-                self._cos_cached[seqlen_offset:],
-                self._sin_cached[seqlen_offset:],
-            )
-            kv = _apply_rotary_emb_kv(
-                kv,
-                self._cos_cached[seqlen_offset:],
-                self._sin_cached[seqlen_offset:],
-            )
+    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=torch.int64).type_as(self.inv_freq)
 
-            return q, kv
+        freqs = torch.outer(t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        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)
 
-class MoE(nn.Module):
-    def __init__(
-        self,
-        config: LlamoeConfig,
-    ):
-        super().__init__()
-        self.mlp = nn.ModuleList([MLP(config) for i in range(config.num_local_experts)])
-        self.gate = nn.Linear(config.n_embd, config.num_local_experts, bias=False)
-        self.num_experts_per_tok = config.num_experts_per_tok
+        return (
+            self.cos_cached[:seq_len].to(dtype=x.dtype),
+            self.sin_cached[:seq_len].to(dtype=x.dtype),
+        )
 
-    def forward(self, x):
-        orig_shape = x.shape
-        x = x.view(-1, x.shape[-1])
 
-        scores = self.gate(x)
-        expert_weights, expert_indices = torch.topk(scores, self.num_experts_per_tok, dim=-1)
-        expert_weights = expert_weights.softmax(dim=-1)
-        flat_expert_indices = expert_indices.view(-1)
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+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)
 
-        x = x.repeat_interleave(self.num_experts_per_tok, dim=0)
-        y = torch.empty_like(x)
-        for i, expert in enumerate(self.mlp):
-            y[flat_expert_indices == i] = expert(x[flat_expert_indices == i])
-        y = (y.view(*expert_weights.shape, -1) * expert_weights.unsqueeze(-1)).sum(dim=1)
-        return y.view(*orig_shape)
 
+# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
 
-class MLP(nn.Module):
-    """Multi-Layer Perceptron.
-    Reference:
-        Attention Is All You Need.
-        https://arxiv.org/pdf/1706.03762.pdf.
+    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 __init__(
-        self,
-        config: PretrainedConfig,
-        n_inner: Optional[int] = None,
-        act_fn: Optional[str] = None,
-    ) -> None:
-        super().__init__()
-
-        act_fn = config.activation_function if act_fn is None else act_fn
-
-        n_inner = getattr(config, "n_inner", None) if n_inner is None else n_inner
-        n_inner = n_inner if n_inner is not None else 4 * config.n_embd
-
-        self.fc1 = nn.Linear(config.n_embd, n_inner)
-        self.fc2 = nn.Linear(n_inner, config.n_embd)
-        self.act = ACT2FN[act_fn]
-
-    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
-        hidden_states = self.fc1(hidden_states)
-        hidden_states = self.act(hidden_states)
-        hidden_states = self.fc2(hidden_states)
 
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+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 SelfAttention(nn.Module):
-    """Self-attention layer (compatible with PyTorch).
-    Reference:
-        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.
-    """
+# Copied from transformers.models.mistral.modeling_mistral.MistralAttention with Mistral->Mixtral
+class LlamoeAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
 
-    def __init__(
-        self,
-        causal: bool = True,
-        softmax_scale: Optional[float] = None,
-        attention_dropout: float = 0.0,
-    ) -> None:
+    def __init__(self, config: LlamoeConfig, layer_idx: Optional[int] = None):
         super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
 
-        self.causal = causal
-        self.softmax_scale = softmax_scale
-        self.drop = nn.Dropout(attention_dropout)
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.is_causal = True
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=config.attention_bias)
+        self._init_rope()
+
+    def _init_rope(self):
+        if self.config.rope_scaling is None:
+            self.rotary_emb = LlamoeRotaryEmbedding(
+                self.head_dim,
+                max_position_embeddings=self.max_position_embeddings,
+                base=self.rope_theta,
+            )
+        
+        else:
+            raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
 
-    @torch.autocast("cpu", enabled=False)
-    @torch.autocast("cuda", enabled=False)
     def forward(
         self,
-        qkv: torch.FloatTensor,
-        causal: bool = None,
-        key_padding_mask: Optional[torch.BoolTensor] = None,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs,
-    ) -> torch.FloatTensor:
-        batch_size, seqlen = qkv.shape[0], qkv.shape[1]
-        q, k, v = qkv.unbind(dim=2)
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        if self.config.pretraining_tp > 1:
+            key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
+            query_slices = self.q_proj.weight.split(
+                (self.num_heads * self.head_dim) // self.config.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)
 
-        q = q.to(torch.float32)
-        k = k.to(torch.float32)
+            query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
+            query_states = torch.cat(query_states, dim=-1)
 
-        causal = self.causal if causal is None else causal
-        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+            key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
+            key_states = torch.cat(key_states, dim=-1)
 
-        # Autocast is manually disabled to avoid `torch.einsum` performing the operation
-        # using float16, which might lead to overflow
-        scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
+            value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
+            value_states = torch.cat(value_states, dim=-1)
 
-        if key_padding_mask is not None:
-            padding_mask = torch.full((batch_size, seqlen), -10000.0, dtype=scores.dtype, device=scores.device)
-            padding_mask.masked_fill_(key_padding_mask, 0.0)
+        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)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        past_key_value = getattr(self, "past_key_value", past_key_value)
+        cos, sin = self.rotary_emb(value_states, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask
+            if cache_position is not None:
+                causal_mask = attention_mask[:, :, cache_position, : key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        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()}"
+            )
 
-            scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
+        attn_output = attn_output.transpose(1, 2).contiguous()
 
-        if causal:
-            causal_mask = torch.triu(torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1)
-            scores = scores + causal_mask.to(dtype=scores.dtype)
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
 
-        attention = torch.softmax(scores, dim=-1).to(v.dtype)
-        attention = self.drop(attention)
+        if self.config.pretraining_tp > 1:
+            attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
+            o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
+            attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
+        else:
+            attn_output = self.o_proj(attn_output)
 
-        output = torch.einsum("bhts,bshd->bthd", attention, v)
+        if not output_attentions:
+            attn_weights = None
 
-        return output
+        return attn_output, attn_weights, past_key_value
 
 
-class CrossAttention(nn.Module):
-    """Cross-attention layer (compatible with PyTorch).
-    Reference:
-        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.
+class LlamoeFlashAttention2(LlamoeAttention):
+    """
+    Llama flash attention module. This module inherits from `LlamaAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
     """
 
-    def __init__(
-        self,
-        causal: bool = True,
-        softmax_scale: Optional[float] = None,
-        attention_dropout: float = 0.0,
-    ) -> None:
-        super().__init__()
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
 
-        self.causal = causal
-        self.softmax_scale = softmax_scale
-        self.drop = nn.Dropout(attention_dropout)
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
 
-    @torch.autocast("cpu", enabled=False)
-    @torch.autocast("cuda", enabled=False)
     def forward(
         self,
-        q: torch.FloatTensor,
-        kv: torch.FloatTensor,
-        causal: bool = None,
-        key_padding_mask: Optional[torch.BoolTensor] = None,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs,
-    ) -> torch.FloatTensor:
-        batch_size, seqlen_q = q.shape[0], q.shape[1]
-        seqlen_k = kv.shape[1]
-
-        if kv.shape[3] != q.shape[2]:
-            kv = repeat(kv, "... hkv d -> ... (hkv g) d", g=q.shape[2] // kv.shape[3])
-        k, v = kv.unbind(dim=2)
-
-        q = q.to(torch.float32)
-        k = k.to(torch.float32)
-
-        causal = self.causal if causal is None else causal
-        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
-
-        # Autocast is manually disabled to avoid `torch.einsum` performing the operation
-        # using float16, which might lead to overflow
-        scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
-
-        if key_padding_mask is not None:
-            padding_mask = torch.full(
-                (batch_size, seqlen_k),
-                -10000.0,
-                dtype=scores.dtype,
-                device=scores.device,
-            )
-            padding_mask.masked_fill_(key_padding_mask, 0.0)
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        cos, sin = self.rotary_emb(value_states, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        past_key_value = getattr(self, "past_key_value", past_key_value)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.attention_dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (LlamaRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
 
-            scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
 
-        if causal:
-            rows = rearrange(torch.arange(seqlen_q, device=q.device, dtype=torch.long), "s -> s 1")
-            cols = torch.arange(seqlen_k, device=k.device, dtype=torch.long)
-            causal_mask = cols > rows + seqlen_k - seqlen_q
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
 
-            scores = scores.masked_fill(causal_mask, -10000.0)
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate
+        )
 
-        attention = torch.softmax(scores, dim=-1).to(v.dtype)
-        attention = self.drop(attention)
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
 
-        output = torch.einsum("bhts,bshd->bthd", attention, v)
+        if not output_attentions:
+            attn_weights = None
 
-        return output
+        return attn_output, attn_weights, past_key_value
 
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
 
-def _find_mha_dims(
-    config: PretrainedConfig,
-    n_head: Optional[int] = None,
-    n_head_kv: Optional[int] = None,
-    head_dim: Optional[int] = None,
-) -> Tuple[int, int]:
-    if n_head is None and head_dim is None:
-        head_dim = config.n_embd // config.n_head
-        n_head = config.n_head
-    elif n_head is None or head_dim is None:
-        raise ValueError("`n_head` and `head_dim` must be both specified or `None`.")
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
 
-    if n_head_kv is None:
-        n_head_kv = getattr(config, "n_head_kv", None) or n_head
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
 
-    return n_head, n_head_kv, head_dim
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
 
+        return attn_output
 
-def _update_kv_cache(kv: torch.FloatTensor, inference_params: InferenceParams, layer_idx: int) -> torch.FloatTensor:
-    num_heads, head_dim = kv.shape[-2:]
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
 
-    if layer_idx not in inference_params.key_value_memory_dict:
-        inference_params.key_value_memory_dict[layer_idx] = torch.empty(
-            inference_params.max_batch_size,
-            inference_params.max_seqlen,
-            2,
-            num_heads,
-            head_dim,
-            dtype=kv.dtype,
-            device=kv.device,
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
         )
 
-    batch_start = inference_params.batch_size_offset
-    batch_end = batch_start + kv.shape[0]
-
-    sequence_start = inference_params.seqlen_offset
-    sequence_end = sequence_start + kv.shape[1]
 
-    # When the current sequence length is equal to or larger than the maximum sequence length,
-    # we need to concatenate the current `kv` with the cached `kv` to expand its length
-    if sequence_end >= inference_params.max_seqlen:
-        inference_params.key_value_memory_dict[layer_idx] = torch.concatenate((inference_params.key_value_memory_dict[layer_idx], kv), dim=1)
+class LlamoeSdpaAttention(LlamoeAttention):
+    """
+    Llama attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `LlamaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
 
-    inference_params.key_value_memory_dict[layer_idx][batch_start:batch_end, sequence_start:sequence_end, ...] = kv
-    kv = inference_params.key_value_memory_dict[layer_idx][batch_start:batch_end, :sequence_end, ...]
-        
-    return kv
+    # Adapted from LlamaAttention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "LlamaModel is using LlamaSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                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,
+                cache_position=cache_position,
+            )
 
+        bsz, q_len, _ = hidden_states.size()
 
-class MHA(nn.Module):
-    """Multi-head attention layer."""
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
 
-    def __init__(
-        self,
-        config: PretrainedConfig,
-        dtype: Optional[torch.dtype] = None,
-        device: Optional[str] = None,
-        rotary_dim: Optional[int] = None,
-        rotary_base: float = 10000.0,
-        rotary_scale_base: Optional[float] = None,
-        n_head: Optional[int] = None,
-        n_head_kv: Optional[int] = None,
-        head_dim: Optional[int] = None,
-        bias: bool = True,
-        causal: bool = True,
-        softmax_scale: Optional[float] = None,
-        layer_idx: Optional[int] = None,
-        return_residual: bool = False,
-        checkpointing: bool = False,
-    ) -> None:
-        super().__init__()
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
-        # Rotary embedding
-        self.rotary_dim = rotary_dim if rotary_dim is not None else getattr(config, "rotary_dim", 0)
-        if self.rotary_dim > 0:
-            rotary_cls = FlashRotaryEmbedding if config.flash_rotary else RotaryEmbedding
-            if rotary_cls is None:
-                rotary_cls = RotaryEmbedding
-
-            rotary_kwargs = {}
-            if rotary_cls is RotaryEmbedding:
-                rotary_kwargs["max_position_embeddings"] = config.n_positions
-
-            self.rotary_emb = rotary_cls(
-                self.rotary_dim,
-                base=rotary_base,
-                scale_base=rotary_scale_base,
-                device=device,
-                **rotary_kwargs,
-            )
+        cos, sin = self.rotary_emb(value_states, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
-        # MLP
-        self.n_head, self.n_head_kv, self.head_dim = _find_mha_dims(
-            config, n_head=n_head, n_head_kv=n_head_kv, head_dim=head_dim
-        )
-        op_size = self.head_dim * (self.n_head + 2 * self.n_head_kv)
-        hidden_size = config.n_embd
+        # In case static cache is used, it is an instance attribute.
+        past_key_value = getattr(self, "past_key_value", past_key_value)
 
-        linear_cls = FusedDense if config.fused_dense else nn.Linear
-        if linear_cls is None:
-            linear_cls = nn.Linear
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        self.Wqkv = linear_cls(hidden_size, op_size, bias=bias, device=device, dtype=dtype)
-        self.out_proj = linear_cls(hidden_size, hidden_size, bias=bias, device=device, dtype=dtype)
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
 
-        # Attention
-        attn_cls = FlashSelfAttention if config.flash_attn else SelfAttention
-        if attn_cls is None:
-            attn_cls = SelfAttention
+        causal_mask = attention_mask
+        if attention_mask is not None and cache_position is not None:
+            causal_mask = causal_mask[:, :, cache_position, : key_states.shape[-2]]
 
-        cross_attn_cls = FlashCrossAttention if config.flash_attn else CrossAttention
-        if cross_attn_cls is None:
-            cross_attn_cls = CrossAttention
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and causal_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
 
-        self.inner_attn = attn_cls(
-            causal=causal,
-            softmax_scale=softmax_scale,
-            attention_dropout=config.attn_pdrop,
-        )
-        self.inner_cross_attn = cross_attn_cls(
-            causal=causal,
-            softmax_scale=softmax_scale,
-            attention_dropout=config.attn_pdrop,
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
         )
 
-        self.flash_attn = config.flash_attn and attn_cls is FlashSelfAttention
-        self.layer_idx = layer_idx
-        self.return_residual = return_residual
-        self.checkpointing = checkpointing
-
-    def _forward_self_attn(
-        self, x: torch.FloatTensor, key_padding_mask: Optional[torch.BoolTensor]
-    ) -> torch.FloatTensor:
-        qkv = self.Wqkv(x)
-        qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
-
-        if self.rotary_dim > 0:
-            qkv = self.rotary_emb(qkv)
-
-        if self.flash_attn:
-            batch_size, seqlen = qkv.shape[0], qkv.shape[1]
-
-            cu_seqlens, max_seqlen = None, None
-            if key_padding_mask is not None:
-                # If `key_padding_mask` is supplied, we need to unpad the input and retrieve
-                # the `cu_seqlens` and `max_seqlen` to be used by `flash-attn`
-                qkv, indices, cu_seqlens, max_seqlen = unpad_input(qkv, key_padding_mask)
-
-            if self.checkpointing:
-                attn_output = torch.utils.checkpoint.checkpoint(
-                    self.inner_attn, qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen
-                )
-            else:
-                attn_output = self.inner_attn(qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen).to(qkv.device)
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
 
-            # If `key_padding_mask` is supplied, we need to pad the output back to the original shape
-            return pad_input(attn_output, indices, batch_size, seqlen) if key_padding_mask is not None else attn_output
+        attn_output = self.o_proj(attn_output)
 
-        if self.checkpointing:
-            return torch.utils.checkpoint.checkpoint(self.inner_attn, qkv, key_padding_mask=key_padding_mask)
+        return attn_output, None, past_key_value
 
-        return self.inner_attn(qkv, key_padding_mask=key_padding_mask)
 
-    def _forward_cross_attn(
-        self,
-        x: torch.FloatTensor,
-        past_key_values: Optional[InferenceParams],
-        key_padding_mask: Optional[torch.BoolTensor],
-    ) -> torch.FloatTensor:
-        batch_size = x.shape[0]
+LLAMOE_ATTENTION_CLASSES = {
+    "eager": LlamoeAttention,
+    "flash_attention_2": LlamoeFlashAttention2,
+    "sdpa": LlamoeSdpaAttention,
+}
 
-        qkv = self.Wqkv(x)
 
-        q = qkv[..., : self.n_head * self.head_dim]
-        q = rearrange(q, "... (h d) -> ... h d", d=self.head_dim)
+class LlamoeBlockSparseTop2MLP(nn.Module):
+    def __init__(self, config:LlamoeConfig):
+        super().__init__()
+        self.ffn_dim = config.intermediate_size
+        self.hidden_dim = config.hidden_size
 
-        kv = qkv[..., self.n_head * self.head_dim :]
-        kv = rearrange(kv, "... (two hkv d) -> ... two hkv d", two=2, d=self.head_dim)
+        self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
+        self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
+        self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
 
-        seqlen_offset = past_key_values.seqlen_offset if past_key_values is not None else 0
-        causal = None if seqlen_offset == 0 else False
-        if self.rotary_dim > 0:
-            q, kv = self.rotary_emb(q, kv=kv, seqlen_offset=seqlen_offset)
+        self.act_fn = ACT2FN[config.hidden_act]
 
-        if past_key_values is not None:
-            kv = _update_kv_cache(kv, past_key_values, self.layer_idx)
+    def forward(self, hidden_states):
+        current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
+        current_hidden_states = self.w2(current_hidden_states)
+        return current_hidden_states
 
-        if self.flash_attn:
-            batch_size, seqlen_q = q.shape[0], q.shape[1]
-            seqlen_k = kv.shape[1]
 
-            cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k = (
-                None,
-                None,
-                None,
-                None,
-            )
-            if key_padding_mask is not None:
-                kv, _, cu_seqlens_k, max_seqlen_k = unpad_input(kv, key_padding_mask)
-
-                if seqlen_q == 1:
-                    key_padding_mask = torch.ones(batch_size, 1, device=q.device)
-                elif seqlen_q != seqlen_k:
-                    key_padding_mask = key_padding_mask[:, -seqlen_q:]
-
-                q, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, key_padding_mask)
-
-            if self.checkpointing:
-                attn_output = torch.utils.checkpoint.checkpoint(
-                    self.inner_cross_attn,
-                    q,
-                    kv,
-                    causal=causal,
-                    cu_seqlens=cu_seqlens_q,
-                    max_seqlen=max_seqlen_q,
-                    cu_seqlens_k=cu_seqlens_k,
-                    max_seqlen_k=max_seqlen_k,
-                )
-            else:
-                attn_output = self.inner_cross_attn(
-                    q,
-                    kv,
-                    causal=causal,
-                    cu_seqlens=cu_seqlens_q,
-                    max_seqlen=max_seqlen_q,
-                    cu_seqlens_k=cu_seqlens_k,
-                    max_seqlen_k=max_seqlen_k,
-                )
+class LlamoeBLockSparseTop2MLP(LlamoeBlockSparseTop2MLP):
+    def __init__(self, *args, **kwargs):
+        logger.warning_once(
+            "MixtralBLockSparseTop2MLP is deprecated by MixtralBlockSparseTop2MLP and will be removed in v4.40."
+        )
+        super().__init__(*args, **kwargs)
 
-            return (
-                pad_input(attn_output, indices_q, batch_size, max_seqlen_q)
-                if key_padding_mask is not None
-                else attn_output
-            )
 
-        if self.checkpointing:
-            return torch.utils.checkpoint.checkpoint(
-                self.inner_cross_attn,
-                q,
-                kv,
-                key_padding_mask=key_padding_mask,
-                causal=causal,
-            )
+class LlamoeSparseMoeBlock(nn.Module):
+    """
+    This implementation is
+    strictly equivalent to standard MoE with full capacity (no
+    dropped tokens). It's faster since it formulates MoE operations
+    in terms of block-sparse operations to accomodate imbalanced
+    assignments of tokens to experts, whereas standard MoE either
+    (1) drop tokens at the cost of reduced performance or (2) set
+    capacity factor to number of experts and thus waste computation
+    and memory on padding.
+    """
 
-        return self.inner_cross_attn(q, kv, key_padding_mask=key_padding_mask, causal=causal)
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_dim = config.hidden_size
+        self.ffn_dim = config.intermediate_size
+        self.num_experts = config.num_local_experts
+        self.top_k = config.num_experts_per_tok
+
+        # gating
+        self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
+
+        self.experts = nn.ModuleList([LlamoeBlockSparseTop2MLP(config) for _ in range(self.num_experts)])
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
+        )
 
-    def forward(
-        self,
-        x: torch.FloatTensor,
-        past_key_values: Optional[InferenceParams] = None,
-        attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
-        **kwargs,
-    ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
-        if attention_mask is not None:
-            attention_mask = attention_mask.bool()
-        else:
-            attention_mask = None
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
 
-        # MHA
-        if self.n_head == self.n_head_kv:
-            if past_key_values is None:
-                # If `past_key_values` are not supplied, we run self-attention
-                attn_output = self._forward_self_attn(x, attention_mask)
-            else:
-                # If `past_key_values` are supplied, it means that we might have cached values and
-                # could take advantage of cross-attention
-                attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
-        # MQA / GQA
-        else:
-            # Regardless of `past_key_values` being supplied or not, it always use cross-attention
-            # because `q` and `kv` lengths might be different
-            attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
 
-        output = rearrange(attn_output, "... h d -> ... (h d)")
-        output = self.out_proj(output)
+            if top_x.shape[0] == 0:
+                continue
 
-        return output if not self.return_residual else (output, x)
+            # in torch it is faster to index using lists than torch tensors
+            top_x_list = top_x.tolist()
+            idx_list = idx.tolist()
 
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None]
 
-class ParallelBlock(nn.Module):
-    """Parallel block.
-    This block applies parallel mixer and MLP layers to the input (used in GPT-J and CodeGen).
-    """
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states, router_logits
 
-    def __init__(
-        self,
-        config: PretrainedConfig,
-        block_idx: Optional[int] = None,
-    ) -> None:
+
+class LlamoeDecoderLayer(nn.Module):
+    def __init__(self, config: LlamoeConfig, layer_idx: int):
         super().__init__()
+        self.hidden_size = config.hidden_size
 
-        self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
-        self.resid_dropout = nn.Dropout(config.resid_pdrop)
-        self.block_idx = block_idx
+        self.self_attn = LLAMOE_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
 
-        self.mixer = MHA(config, layer_idx=block_idx)
-        self.moe = MoE(config)
+        self.block_sparse_moe = LlamoeSparseMoeBlock(config)
+        self.input_layernorm = LlamoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
     def forward(
         self,
-        hidden_states: torch.FloatTensor,
-        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
-        attention_mask: Optional[torch.BoolTensor] = None,
+        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,
+        output_router_logits: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
         **kwargs,
-    ) -> torch.FloatTensor:
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+        """
+        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, sequence_length)` where padding elements are indicated by 0.
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            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_router_logits (`bool`, *optional*):
+                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
+                should not be returned during inference.
+            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`).
+        """
+
         residual = hidden_states
-        hidden_states = self.ln(hidden_states)
 
-        attn_outputs = self.mixer(
-            hidden_states,
-            past_key_values=past_key_values,
+        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,
         )
-        if isinstance(attn_outputs, tuple):
-            attn_outputs = attn_outputs[0]
-
-        attn_outputs = self.resid_dropout(attn_outputs)
-        feed_forward_hidden_states = self.resid_dropout(self.moe(hidden_states))
-
-        hidden_states = attn_outputs + feed_forward_hidden_states + residual
-
-        return hidden_states
-
+        hidden_states = residual + hidden_states
 
-class CausalLMHead(nn.Module):
-    """Causal Language Modeling head.
-    Reference:
-        Improving Language Understanding by Generative Pre-Training.
-        https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
-    """
-
-    def __init__(self, config: PretrainedConfig) -> None:
-        super().__init__()
-
-        self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
-        self.linear = nn.Linear(config.n_embd, config.vocab_size)
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states, router_logits = self.block_sparse_moe(hidden_states)
+        hidden_states = residual + hidden_states
 
-    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
-        hidden_states = self.ln(hidden_states)
-        logits = self.linear(hidden_states).to(torch.float32)
+        outputs = (hidden_states,)
 
-        return logits
+        if output_attentions:
+            outputs += (self_attn_weights,)
 
+        if use_cache:
+            outputs += (present_key_value,)
 
-class CausalLMLoss(nn.Module):
-    """Causal Language Modeling loss.
-    Reference:
-        Improving Language Understanding by Generative Pre-Training.
-        https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
-    """
+        if output_router_logits:
+            outputs += (router_logits,)
 
-    def __init__(self, shift_labels: bool = True) -> None:
-        super().__init__()
+        return outputs
 
-        self.shift_labels = shift_labels
-        self.loss_fct = nn.CrossEntropyLoss()
 
-    def forward(self, logits: torch.FloatTensor, labels: torch.LongTensor) -> torch.FloatTensor:
-        if self.shift_labels:
-            logits = logits[..., :-1, :].contiguous()
-            labels = labels[..., 1:].contiguous()
+LLAMOE_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.)
 
-        loss = self.loss_fct(logits.view(-1, logits.size(-1)), labels.view(-1))
+    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.
 
-        return loss
+    Parameters:
+        config ([`MixtralConfig`]):
+            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 Mixtral Model outputting raw hidden-states without any specific head on top.",
+    LLAMOE_START_DOCSTRING,
+)
+# Copied from transformers.models.mistral.modeling_mistral.MistralPreTrainedModel with Mistral->Mixtral
 class LlamoePreTrainedModel(PreTrainedModel):
-    """Phi pre-trained model."""
-
     config_class = LlamoeConfig
-    base_model_prefix = "transformer"
-    supports_gradient_checkpointing = False
-    _no_split_modules = ["ParallelBlock"]
-
-    def __init__(self, *inputs, **kwargs) -> None:
-        super().__init__(*inputs, **kwargs)
-
-    def _init_weights(self, module: nn.Module) -> None:
-        if isinstance(module, (nn.Linear,)):
-            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlamoeDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+
+    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=self.config.initializer_range)
+            module.weight.data.normal_(mean=0.0, std=std)
             if module.padding_idx is not None:
                 module.weight.data[module.padding_idx].zero_()
-        elif isinstance(module, nn.LayerNorm):
-            if module.bias is not None:
-                module.bias.data.zero_()
-            module.weight.data.fill_(1.0)
-
-    def prepare_inputs_for_generation(
-        self,
-        input_ids: torch.LongTensor,
-        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
-        attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
-        **kwargs,
-    ) -> Dict[str, Any]:
-        if past_key_values is None or not (isinstance(past_key_values, InferenceParams)):
-            past_key_values = InferenceParams(
-                max_seqlen=self.config.n_positions,
-                max_batch_size=input_ids.shape[0],
-                seqlen_offset=0,
-                batch_size_offset=0,
-                key_value_memory_dict={},
-                lengths_per_sample=None,
-            )
-        else:
-            # Assume that `past_key_values` has cached all tokens up to the last token in `input_ids`
-            past_key_values.seqlen_offset = input_ids.shape[1] - 1
-            input_ids = input_ids[:, -1].unsqueeze(-1)
-
-        return {
-            "input_ids": input_ids,
-            "past_key_values": past_key_values,
-            "attention_mask": attention_mask,
-        }
 
 
+LLAMOE_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.
+        output_router_logits (`bool`, *optional*):
+            Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
+            should not be returned during inference.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare Mixtral Model outputting raw hidden-states without any specific head on top.",
+    LLAMOE_START_DOCSTRING,
+)
+# Copied from transformers.models.mistral.modeling_mistral.MistralModel with MISTRAL->MIXTRAL,Mistral->Mixtral
 class LlamoeModel(LlamoePreTrainedModel):
-    """Llamoe model."""
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MixtralDecoderLayer`]
 
-    _keys_to_ignore_on_load_missing = [""]
-    _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]
+    Args:
+        config: MixtralConfig
+    """
 
-    def __init__(self, config: LlamoeConfig) -> None:
+    def __init__(self, config: LlamoeConfig):
         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(
+            [LlamoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self._attn_implementation = config._attn_implementation
+        self.norm = LlamoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
-        self.embd = Embedding(config)
-        self.h = nn.ModuleList([ParallelBlock(config, block_idx=i) for i in range(config.n_layer)])
         self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
         self.post_init()
 
-    def get_input_embeddings(self) -> nn.Embedding:
-        return self.embd.wte
+    def get_input_embeddings(self):
+        return self.embed_tokens
 
-    def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
-        self.embd.wte = new_embeddings
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
 
+    # Ignore copy
+    @add_start_docstrings_to_model_forward(LLAMOE_INPUTS_DOCSTRING)
     def forward(
         self,
-        input_ids: torch.LongTensor,
-        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
-        attention_mask: Optional[torch.BoolTensor] = None,
-    ) -> torch.FloatTensor:
-        hidden_states = self.embd(input_ids)
-
-        for layer in self.h:
-            hidden_states = layer(
-                hidden_states,
-                past_key_values=past_key_values,
-                attention_mask=attention_mask,
+        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,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, MoeModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+        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")
+
+        past_key_values_length = 0
+
+        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
+
+        if use_cache:
+            use_legacy_cache = not isinstance(past_key_values, Cache)
+            if use_legacy_cache:
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            past_key_values_length = past_key_values.get_usable_length(seq_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)
+
+        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
+            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
+            if is_padding_right:
+                raise ValueError(
+                    "You are attempting to perform batched generation with padding_side='right'"
+                    " this may lead to unexpected behaviour for Flash Attention version of Mixtral. Make sure to "
+                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+                )
 
-        return hidden_states
+        if self._attn_implementation == "flash_attention_2":
+            # 2d mask is passed through the layers
+            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+        elif self._attn_implementation == "sdpa" and not output_attentions:
+            # output_attentions=True can not be supported when using SDPA, and we fall back on
+            # the manual implementation that requires a 4D causal mask in all cases.
+            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_key_values_length,
+            )
+        else:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_key_values_length,
+                sliding_window=self.config.sliding_window,
+            )
 
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_router_logits = () if output_router_logits else None
+        next_decoder_cache = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    output_router_logits,
+                    use_cache,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                )
 
-class LlamoeForCausalLM(LlamoePreTrainedModel):
-    """Llamoe for Causal Language Modeling."""
+            hidden_states = layer_outputs[0]
 
-    _keys_to_ignore_on_load_missing = [""]
-    _keys_to_ignore_on_load_unexpected = [r"transformer\.h\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
 
-    def __init__(self, config: PhiConfig) -> None:
-        super().__init__(config)
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            if output_router_logits:
+                all_router_logits += (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,)
 
-        self.transformer = LlamoeModel(config)
-        self.lm_head = CausalLMHead(config)
-        self.loss = CausalLMLoss()
+        next_cache = None
+        if use_cache:
+            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
+                if v is not None
+            )
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            router_logits=all_router_logits,
+        )
+
+
+class LlamoeForCausalLM(LlamoePreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
 
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = LlamoeModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.router_aux_loss_coef = config.router_aux_loss_coef
+        self.num_experts = config.num_local_experts
+        self.num_experts_per_tok = config.num_experts_per_tok
+        # Initialize weights and apply final processing
         self.post_init()
 
-    def get_output_embeddings(self) -> nn.Linear:
-        return self.lm_head.linear
+    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: nn.Linear) -> None:
-        self.lm_head.linear = new_embeddings
+    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(LLAMOE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    # Ignore copy
     def forward(
         self,
-        input_ids: torch.LongTensor,
-        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
-        attention_mask: Optional[torch.BoolTensor] = None,
+        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,
-        **kwargs,
-    ) -> CausalLMOutputWithPast:
-        hidden_states = self.transformer(input_ids, past_key_values=past_key_values, attention_mask=attention_mask)
-        lm_logits = self.lm_head(hidden_states)
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+        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, MixtralForCausalLM
+
+        >>> model = MixtralForCausalLM.from_pretrained("mistralai/Mixtral-8x7B-v0.1")
+        >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mixtral-8x7B-v0.1")
+
+        >>> 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_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+
+        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,
+            output_router_logits=output_router_logits,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
 
         loss = None
         if labels is not None:
-            loss = self.loss(lm_logits, labels)
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        aux_loss = None
+        if output_router_logits:
+            aux_loss = load_balancing_loss_func(
+                outputs.router_logits if return_dict else outputs[-1],
+                self.num_experts,
+                self.num_experts_per_tok,
+                attention_mask,
+            )
+            if labels is not None:
+                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            if output_router_logits:
+                output = (aux_loss,) + output
+            return (loss,) + output if loss is not None else output
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
 
-        return CausalLMOutputWithPast(loss=loss, logits=lm_logits, past_key_values=past_key_values)
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        output_router_logits=False,
+        **kwargs,
+    ):
+        # Omit tokens covered by past_key_values
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        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[:, -input_ids.shape[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,
+                "output_router_logits": output_router_logits,
+            }
+        )
+        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