randommoe.py

from typing import Callable, Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F

from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...modeling_attn_mask_utils import AttentionMaskConverter
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import (
    LossKwargs,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    can_return_tuple,
    is_torch_flex_attn_available,
    logging,
    replace_return_docstrings,
)
from .configuration_olmo import OlmoConfig


if is_torch_flex_attn_available():
    from torch.nn.attention.flex_attention import BlockMask

    from ...integrations.flex_attention import make_flex_block_causal_mask


logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "OlmoConfig"


class OlmoLayerNorm(nn.Module):
    """LayerNorm but with no learnable weight or bias."""

    def __init__(self, hidden_size: int) -> None:
        super().__init__()
        self.normalized_shape = (hidden_size,)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        orig_dtype = hidden_states.dtype
        return F.layer_norm(hidden_states.to(dtype=torch.float32), self.normalized_shape, None, None, eps=1e-5).to(
            orig_dtype
        )


class OlmoMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        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.unsqueeze(unsqueeze_dim)
    sin = sin.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 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)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


class OlmoAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: OlmoConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        if self.config.clip_qkv is not None:
            query_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
            key_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)
            value_states.clamp_(min=-self.config.clip_qkv, max=self.config.clip_qkv)

        query_states = query_states.view(hidden_shape).transpose(1, 2)
        key_states = key_states.view(hidden_shape).transpose(1, 2)
        value_states = value_states.view(hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        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)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
                logger.warning_once(
                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
                )
            else:
                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class OlmoDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: OlmoConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = OlmoAttention(config=config, layer_idx=layer_idx)

        self.mlp = OlmoMLP(config)
        self.input_layernorm = OlmoLayerNorm(config.hidden_size)
        self.post_attention_layernorm = OlmoLayerNorm(config.hidden_size)

    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: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights = 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,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)
        if output_attentions:
            outputs += (self_attn_weights,)

        return outputs


class OlmoRotaryEmbedding(nn.Module):
    def __init__(self, config: OlmoConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


OLMO_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`OlmoConfig`]):
            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 Olmo Model outputting raw hidden-states without any specific head on top.",
    OLMO_START_DOCSTRING,
)
class OlmoPreTrainedModel(PreTrainedModel):
    config_class = OlmoConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["OlmoDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_cache_class = True
    _supports_quantized_cache = True
    _supports_static_cache = True
    _supports_attention_backend = 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=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()


OLMO_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) or `BlockMask`, *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**.

            If the model is configured to use flex_attention, it will attempt to convert the mask Tensor into a BlockMask,
            but you can also pass a `BlockMask` object directly here.

            [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 `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 (`Cache`, *optional*):
            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.

            It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
            of shape `(batch_size, sequence_length)`.
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
            the complete sequence length.
"""


@add_start_docstrings(
    "The bare Olmo Model outputting raw hidden-states without any specific head on top.",
    OLMO_START_DOCSTRING,
)
class OlmoModel(OlmoPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`OlmoDecoderLayer`]
    olmo's mapping in https://github.com/huggingface/transformers/blob/main/src/transformers/models/auto/modeling_auto.py

    Args:
        config: OlmoConfig
    """

    def __init__(self, config: OlmoConfig):
        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(
            [OlmoDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = OlmoLayerNorm(config.hidden_size)
        self.rotary_emb = OlmoRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value

    @can_return_tuple
    @add_start_docstrings_to_model_forward(OLMO_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> BaseModelOutputWithPast:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if self.gradient_checkpointing and self.training and use_cache:
            logger.warning_once(
                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
            )
            use_cache = False

        # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache
        if not isinstance(past_key_values, (type(None), Cache)):
            raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache()

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        causal_mask = self._update_causal_mask(
            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
        )

        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                position_ids=position_ids,
                past_key_value=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **flash_attn_kwargs,
            )

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )

    def _update_causal_mask(
        self,
        attention_mask: Union[torch.Tensor, "BlockMask"],
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool = False,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and (attention_mask == 0.0).any():
                return attention_mask
            return None
        if self.config._attn_implementation == "flex_attention":
            if isinstance(attention_mask, torch.Tensor):
                attention_mask = make_flex_block_causal_mask(attention_mask)
            return attention_mask

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_compilable_cache = past_key_values.is_compileable if past_key_values is not None else False

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        if self.config._attn_implementation == "sdpa" and not using_compilable_cache and not output_attentions:
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                is_training=self.training,
            ):
                return None

        dtype = input_tensor.dtype
        sequence_length = input_tensor.shape[1]
        if using_compilable_cache:
            target_length = past_key_values.get_max_cache_shape()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type in ["cuda", "xpu", "npu"]
            and not output_attentions
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            min_dtype = torch.finfo(dtype).min
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask

    @staticmethod
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        cache_position: torch.Tensor,
        batch_size: int,
        **kwargs,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
                    causal_mask.device
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask


class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...


class OlmoForCausalLM(OlmoPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

    def __init__(self, config):
        super().__init__(config)
        self.model = OlmoModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model

    @can_return_tuple
    @add_start_docstrings_to_model_forward(OLMO_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs: Unpack[KwargsForCausalLM],
    ) -> CausalLMOutputWithPast:
        r"""
            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]`.

            logits_to_keep (`int` or `torch.Tensor`, *optional*):
                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
                This is useful when using packed tensor format (single dimension for batch and sequence length).

        Returns:

        Example:

        ```python
        >>> from transformers import AutoTokenizer, OlmoForCausalLM

        >>> model = OlmoForCausalLM.from_pretrained("meta-olmo/Olmo-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-olmo/Olmo-2-7b-hf")

        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs: BaseModelOutputWithPast = 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,
            cache_position=cache_position,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

import torch
import torch.nn as nn
import torch.nn.functional as F


class OlmoMoERouter(nn.Module):
    """
    Router module that uses random importance sampling instead of deterministic top-k.
    
    This router computes logits for each expert, converts them to probabilities,
    and then randomly samples experts based on these probabilities.
    """
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.num_experts = config.num_experts
        self.router = nn.Linear(self.hidden_size, self.num_experts, bias=False)
        self.top_k = config.num_selected_experts
        self.temperature = config.router_temperature if hasattr(config, "router_temperature") else 1.0
        
    def forward(self, hidden_states):
        """
        Args:
            hidden_states: [batch_size, sequence_length, hidden_size]
            
        Returns:
            routing_weights: [batch_size, sequence_length, top_k]
            routing_indices: [batch_size, sequence_length, top_k]
        """
        batch_size, sequence_length, _ = hidden_states.shape
        
        # Compute router logits and apply temperature
        router_logits = self.router(hidden_states) / self.temperature  # [batch_size, sequence_length, num_experts]
        
        # Convert to probabilities using softmax
        router_probs = F.softmax(router_logits, dim=-1)  # [batch_size, sequence_length, num_experts]
        
        # For random importance sampling, we'll:
        # 1. Add Gumbel noise to the log probabilities to induce randomness
        # 2. Sample top-k experts using the perturbed probabilities
        
        # Add Gumbel noise
        gumbel_noise = -torch.log(-torch.log(torch.rand_like(router_probs) + 1e-10) + 1e-10)
        perturbed_logits = torch.log(router_probs + 1e-10) + gumbel_noise
        
        # Sample top-k experts based on perturbed probabilities
        routing_weights, routing_indices = torch.topk(perturbed_logits, self.top_k, dim=-1)
        
        # Re-normalize the selected probabilities
        routing_weights = router_probs.gather(-1, routing_indices)
        routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True)
        
        return routing_weights, routing_indices


class OlmoExpertMLP(nn.Module):
    """
    Expert MLP module similar to OlmoMLP but used in the MoE architecture.
    """
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


class OlmoMixtureOfExperts(nn.Module):
    """
    Mixture of Experts layer that replaces the standard MLP in OLMo.
    """
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.num_experts = config.num_experts
        self.num_selected_experts = config.num_selected_experts  # top_k
        
        # Create router
        self.router = OlmoMoERouter(config)
        
        # Create experts
        self.experts = nn.ModuleList([OlmoExpertMLP(config) for _ in range(self.num_experts)])
        
        # Expert capacity factor (to avoid load balancing issues)
        self.capacity_factor = config.expert_capacity_factor if hasattr(config, "expert_capacity_factor") else 1.0
        
    def forward(self, hidden_states):
        """
        Args:
            hidden_states: [batch_size, sequence_length, hidden_size]
            
        Returns:
            outputs: [batch_size, sequence_length, hidden_size]
        """
        batch_size, sequence_length, hidden_size = hidden_states.shape
        
        # Get routing weights and indices
        routing_weights, routing_indices = self.router(hidden_states)
        
        # Reshape tensors for processing
        flat_hidden_states = hidden_states.reshape(-1, hidden_size)  # [batch_size * sequence_length, hidden_size]
        
        # Initialize expert outputs
        final_output = torch.zeros_like(flat_hidden_states)
        
        # For each expert, compute its contribution
        for expert_idx in range(self.num_experts):
            # Create a mask to identify which tokens use this expert
            expert_mask = (routing_indices == expert_idx).any(dim=-1).reshape(-1)
            
            if not expert_mask.any():
                continue  # Skip if no token routes to this expert
                
            # Get the hidden states for tokens routed to this expert
            expert_inputs = flat_hidden_states[expert_mask]
            
            # Process these hidden states through the expert
            expert_outputs = self.experts[expert_idx](expert_inputs)
            
            # Find weights for this expert
            expert_weights = routing_weights[routing_indices == expert_idx].reshape(-1, 1)
            
            # Multiply outputs by the routing weights
            weighted_outputs = expert_outputs * expert_weights
            
            # Combine the expert outputs into the final output tensor
            final_output[expert_mask] += weighted_outputs
            
        # Reshape back to original dimensions
        final_output = final_output.reshape(batch_size, sequence_length, hidden_size)
        
        return final_output


# Modified OlmoDecoderLayer to use MoE instead of standard MLP
class OlmoMoEDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: OlmoConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = OlmoAttention(config=config, layer_idx=layer_idx)

        # Use MoE instead of standard MLP
        self.mlp = OlmoMixtureOfExperts(config)
        self.input_layernorm = OlmoLayerNorm(config.hidden_size)
        self.post_attention_layernorm = OlmoLayerNorm(config.hidden_size)

    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: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights = 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,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # MoE instead of Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)
        if output_attentions:
            outputs += (self_attn_weights,)

        return outputs


# Modified OlmoConfig to include MoE-specific parameters
class OlmoMoEConfig(OlmoConfig):
    def __init__(
        self,
        num_experts=8,
        num_selected_experts=2,
        expert_capacity_factor=1.0,
        router_temperature=0.1,
        **kwargs
    ):
        super().__init__(**kwargs)
        self.num_experts = num_experts
        self.num_selected_experts = num_selected_experts
        self.expert_capacity_factor = expert_capacity_factor
        self.router_temperature = router_temperature


# Modified OlmoModel to use MoE decoder layers
class OlmoMoEModel(OlmoModel):
    def __init__(self, config: OlmoMoEConfig):
        OlmoPreTrainedModel.__init__(self, 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)
        # Use MoE decoder layers
        self.layers = nn.ModuleList(
            [OlmoMoEDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = OlmoLayerNorm(config.hidden_size)
        self.rotary_emb = OlmoRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

        # Initialize weights and apply final processing
        self.post_init()


# Modified OlmoForCausalLM to use MoE model
class OlmoMoEForCausalLM(OlmoForCausalLM):
    def __init__(self, config):
        OlmoPreTrainedModel.__init__(self, config)
        self.model = OlmoMoEModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

__all__ = ["OlmoForCausalLM", "OlmoModel", "OlmoPreTrainedModel"]