Liquid-Codec / modeling_lfm3.py
paulpak58
[Misc] Remote code
840e7da 1 day ago
28.4 kB
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/lfm3/modular_lfm3.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_lfm3.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections.abc import Callable

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

from transformers import initialization as init
from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.integrations import use_kernel_forward_from_hub
from transformers.integrations.accelerate import force_accelerate_hooks
from transformers.masking_utils import create_causal_mask, create_recurrent_attention_mask, create_sliding_window_causal_mask
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel, slice_logits_to_keep
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple
from transformers.utils.generic import merge_with_config_defaults
from transformers.utils.output_capturing import capture_outputs
from .configuration_lfm3 import Lfm3Config


@use_kernel_forward_from_hub("RMSNorm")
class Lfm3RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        Lfm3RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        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)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"


class Lfm3RotaryEmbedding(nn.Module):
    def __init__(self, config: Lfm3Config, device=None):
        super().__init__()
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings
        self.config = config
        # diff -> only the attention layer types have a RoPE entry, `conv` layers have none
        self.layer_types = [layer_type for layer_type in config._rope_type_labels if layer_type in config.layer_types]
        self.rope_type = {}
        for layer_type in self.layer_types:
            self.rope_type[layer_type] = config.rope_parameters[layer_type]["rope_type"]
            rope_init_fn: Callable = self.compute_default_rope_parameters
            if self.rope_type[layer_type] != "default":
                rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type[layer_type]]
            curr_inv_freq, curr_attention_scaling = rope_init_fn(config, device, layer_type=layer_type)
            setattr(self, f"{layer_type}_inv_freq", nn.Buffer(curr_inv_freq, persistent=False))
            setattr(self, f"{layer_type}_original_inv_freq", nn.Buffer(curr_inv_freq.clone(), persistent=False))
            setattr(self, f"{layer_type}_attention_scaling", curr_attention_scaling)

    @staticmethod
    def compute_default_rope_parameters(
        config: Lfm3Config, device=None, layer_type: str | None = None, **kwargs
    ) -> tuple[torch.Tensor, float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            layer_type (`str`):
                The current layer type if the model has different RoPE parameters per type.
                Should not be used unless `config.layer_types is not None`
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        # For backward compatibility standardize the `rope_parameters_dict` if it uses old format
        base = config.rope_parameters[layer_type]["rope_theta"]
        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads

        attention_factor = 1.0  # Unused in this type of RoPE
        # Compute the inverse frequencies
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        return inv_freq.to(device), attention_factor

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids, layer_type):
        # diff -> returns cos/sin in fp32 without casting to `x.dtype`
        inv_freq = getattr(self, f"{layer_type}_inv_freq")
        attention_scaling = getattr(self, f"{layer_type}_attention_scaling")

        freqs = position_ids[..., None].float() * inv_freq.to(device=x.device, dtype=torch.float)
        emb = torch.cat((freqs, freqs), dim=-1)
        cos = emb.cos() * attention_scaling
        sin = emb.sin() * attention_scaling

        return cos, sin


class Lfm3MLP(nn.Module):
    def __init__(self, config: Lfm3Config, intermediate_size: int | None = None):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
        self.w1 = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.w3 = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.w2 = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)

    def forward(self, x):
        return self.w2(F.silu(self.w1(x)) * self.w3(x))


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)


@use_kernel_forward_from_hub("rotary_pos_emb")
def apply_rotary_pos_emb(q, k, cos, sin, 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.
        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.
    """
    q_type, k_type = q.dtype, k.dtype
    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.to(q_type), k_embed.to(k_type)


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: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    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:
        attn_weights = attn_weights + attention_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 Lfm3Attention(nn.Module):
    """Qwen3 attention with RoPE in float32 and a sigmoid output gate per head."""

    def __init__(self, config: Lfm3Config, layer_idx: int):
        super().__init__()
        self.layer_type = config.layer_types[layer_idx] if hasattr(config, "layer_types") else None
        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
        )
        self.q_norm = Lfm3RMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
        self.k_norm = Lfm3RMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape
        self.sliding_window = config.sliding_window if self.layer_type == "sliding_attention" else None
        self.g_proj = (
            nn.Linear(config.hidden_size, config.num_attention_heads, bias=False)
            if config.use_attention_gate
            else None
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None,
        past_key_values: Cache | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
        key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
        value_states = self.v_proj(hidden_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_values is not None:
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        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,
            sliding_window=self.sliding_window,
            **kwargs,
        )
        if self.g_proj is not None:
            attn_output = attn_output * torch.sigmoid(self.g_proj(hidden_states)).unsqueeze(-1)
        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


def apply_mask_to_padding_states(hidden_states, attention_mask):
    """
    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
    """
    # NOTE: attention mask is a 2D boolean tensor
    if attention_mask is not None:
        dtype = hidden_states.dtype
        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)

    return hidden_states


# torch only: the hub causal_conv1d kernels serve widths 2-4, LFM3 uses 7
def causal_conv1d_update(
    hidden_states: torch.Tensor,
    conv_state: torch.Tensor,
    weight: nn.Parameter,
    bias: nn.Parameter | None = None,
    activation: str | None = None,
):
    _, hidden_size, seq_len = hidden_states.shape
    state_len = conv_state.shape[-1]

    hidden_states_new = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype)
    conv_state.copy_(hidden_states_new[:, :, -state_len:])
    out = F.conv1d(hidden_states_new, weight.unsqueeze(1), bias, padding=0, groups=hidden_size)
    out = out[:, :, -seq_len:]
    if activation is not None:
        out = ACT2FN[activation](out)
    return out.to(hidden_states.dtype)


def causal_conv1d_fn(
    hidden_states: torch.Tensor,
    weight: nn.Parameter,
    bias: nn.Parameter | None = None,
    activation: str | None = None,
    **kwargs,
):
    _, hidden_size, seq_len = hidden_states.shape
    padding = weight.shape[-1] - 1

    out = F.conv1d(
        hidden_states.to(weight.dtype),
        weight=weight.unsqueeze(1),
        bias=bias,
        padding=padding,
        groups=hidden_size,
    )[:, :, :seq_len]
    if activation is not None:
        out = ACT2FN[activation](out)
    return out.to(hidden_states.dtype)


class Lfm3ShortConv(nn.Module):
    def __init__(
        self,
        config: Lfm3Config,
        layer_idx: int,
    ):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.conv_kernel_size = config.conv_L_cache
        self.bias = config.conv_bias

        self.conv = nn.Conv1d(
            in_channels=config.hidden_size,
            out_channels=config.hidden_size,
            kernel_size=self.conv_kernel_size,
            groups=config.hidden_size,
            bias=self.bias,
            padding=self.conv_kernel_size - 1,
        )
        self.in_proj = nn.Linear(config.hidden_size, 3 * config.hidden_size, bias=self.bias)
        self.out_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=self.bias)

        self.layer_type = config.layer_types[layer_idx]

    @force_accelerate_hooks("conv")
    def forward(
        self,
        hidden_states: torch.Tensor,
        past_key_values: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs,
    ):
        seq_len = hidden_states.shape[1]

        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
        BCx = self.in_proj(hidden_states).transpose(-1, -2)
        B, C, x = BCx.chunk(3, dim=-2)
        hidden_states = B * x

        if past_key_values is not None:
            conv_state = past_key_values.get_conv_state(self.layer_idx, state_idx=0)
        else:
            conv_state = None

        if conv_state is not None and seq_len == 1 and not past_key_values.is_recording_past(self.layer_idx):
            # Single-token cached decode: the fused per-step kernel updates the conv state in-place.
            hidden_states = causal_conv1d_update(
                hidden_states, conv_state, self.conv.weight.squeeze(1), self.conv.bias
            )
        else:
            if past_key_values is not None:
                hidden_states = past_key_values.update_conv_state(
                    hidden_states, self.layer_idx, conv_kernel_size=self.conv_kernel_size
                )

            hidden_states = causal_conv1d_fn(hidden_states, self.conv.weight.squeeze(1), self.conv.bias, **kwargs)

            # Drop the additional previous states
            if past_key_values is not None:
                hidden_states = hidden_states[:, :, -seq_len:]

        y = C * hidden_states
        y = y.transpose(-1, -2).contiguous()
        y = self.out_proj(y)
        return y


class Lfm3DecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Lfm3Config, layer_idx: int):
        super().__init__()
        self.is_attention_layer = config.layer_types[layer_idx] != "conv"

        if self.is_attention_layer:
            self.self_attn = Lfm3Attention(config, layer_idx)
        else:
            self.conv = Lfm3ShortConv(config, layer_idx)
        self.feed_forward = Lfm3MLP(config)
        self.operator_norm = Lfm3RMSNorm(config.hidden_size, eps=config.norm_eps)
        self.ffn_norm = Lfm3RMSNorm(config.hidden_size, eps=config.norm_eps)
        self.post_operator_norm = (
            Lfm3RMSNorm(config.hidden_size, eps=config.norm_eps) if config.use_sandwich_norm else nn.Identity()
        )
        self.post_ffn_norm = (
            Lfm3RMSNorm(config.hidden_size, eps=config.norm_eps) if config.use_sandwich_norm else nn.Identity()
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        **kwargs,
    ) -> torch.Tensor:
        residual = hidden_states
        if self.is_attention_layer:
            hidden_states, _ = self.self_attn(
                hidden_states=self.operator_norm(hidden_states),
                position_embeddings=position_embeddings,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                **kwargs,
            )
        else:
            hidden_states = self.conv(
                hidden_states=self.operator_norm(hidden_states),
                past_key_values=past_key_values,
                attention_mask=attention_mask,
                **kwargs,
            )
        hidden_states = self.post_operator_norm(hidden_states)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.ffn_norm(hidden_states)
        hidden_states = self.feed_forward(hidden_states)
        hidden_states = self.post_ffn_norm(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states


@auto_docstring
class Lfm3PreTrainedModel(PreTrainedModel):
    config: Lfm3Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Lfm3DecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": Lfm3DecoderLayer,
        "attentions": Lfm3Attention,
    }

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, Lfm3RotaryEmbedding):
            for layer_type in module.layer_types:
                rope_init_fn = module.compute_default_rope_parameters
                if module.rope_type[layer_type] != "default":
                    rope_init_fn = ROPE_INIT_FUNCTIONS[module.rope_type[layer_type]]
                curr_inv_freq, _ = rope_init_fn(module.config, layer_type=layer_type)
                init.copy_(getattr(module, f"{layer_type}_inv_freq"), curr_inv_freq)
                init.copy_(getattr(module, f"{layer_type}_original_inv_freq"), curr_inv_freq)


@auto_docstring
class Lfm3Model(Lfm3PreTrainedModel):
    """LFM2 model with a sliding-attention layer type; the mask and rotary embedding are built per layer type."""

    def __init__(self, config: Lfm3Config):
        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(
            [Lfm3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.rotary_emb = Lfm3RotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        self.embedding_norm = Lfm3RMSNorm(config.hidden_size, eps=config.norm_eps)

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

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        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(config=self.config)

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

        if not isinstance(causal_mask_mapping := attention_mask, dict):
            mask_kwargs = {
                "config": self.config,
                "inputs_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
            }
            mask_creation_functions = {
                "full_attention": lambda: create_causal_mask(**mask_kwargs),
                "sliding_attention": lambda: create_sliding_window_causal_mask(**mask_kwargs),
                "conv": lambda: create_recurrent_attention_mask(**mask_kwargs),
            }
            causal_mask_mapping = {}
            for layer_type in set(self.config.layer_types):
                causal_mask_mapping[layer_type] = mask_creation_functions[layer_type]()

        hidden_states = inputs_embeds
        position_embeddings = {
            layer_type: self.rotary_emb(hidden_states, position_ids=position_ids, layer_type=layer_type)
            for layer_type in self.rotary_emb.layer_types
        }

        for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]):
            layer_type = self.config.layer_types[i]
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping[layer_type],
                position_embeddings=position_embeddings.get(layer_type),
                position_ids=position_ids,
                past_key_values=past_key_values,
                **kwargs,
            )

        hidden_states = self.embedding_norm(hidden_states)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


@auto_docstring
class Lfm3ForCausalLM(Lfm3PreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
    _tp_plan = {"lm_head": "colwise_gather_output"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
    _fsdp_plan = {"lm_head": "keep_full_weight"}

    def __init__(self, config: Lfm3Config):
        super().__init__(config)
        self.model = Lfm3Model(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()

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> CausalLMOutputWithPast:
        r"""
        Example:
        ```python
        >>> from transformers import AutoTokenizer, Lfm3ForCausalLM
        >>> model = Lfm3ForCausalLM.from_pretrained("LiquidAI/LFM3-1.7B")
        >>> tokenizer = AutoTokenizer.from_pretrained("LiquidAI/LFM3-1.7B")
        >>> inputs = tokenizer.apply_chat_template(
        ...     [{"role": "user", "content": "What is C. elegans?"}], add_generation_prompt=True, return_tensors="pt"
        ... )
        >>> generate_ids = model.generate(**inputs, max_new_tokens=64)
        >>> tokenizer.decode(generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
        ```"""
        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,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        hidden_states = slice_logits_to_keep(hidden_states, logits_to_keep)
        logits = self.lm_head(hidden_states)

        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,
        )


__all__ = ["Lfm3ForCausalLM", "Lfm3Model", "Lfm3PreTrainedModel"]