text-generation-inference/server/text_generation_server/models/custom_modeling/mllama.py

# coding=utf-8
# Copyright 2024 the HuggingFace Inc. 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.
"""PyTorch Mllama model."""

from typing import Optional, Tuple, List

import torch
import torch.utils.checkpoint
from torch import nn
import math

from transformers.activations import ACT2FN
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
from transformers.modeling_outputs import (
    CausalLMOutputWithPast,
    BaseModelOutputWithPast,
)
from transformers.cache_utils import (
    StaticCache,
    DynamicCache,
)
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
import torch.nn.functional as F

from text_generation_server.layers import (
    TensorParallelColumnLinear,
    TensorParallelEmbedding,
    TensorParallelRowLinear,
    SpeculativeHead,
    FastLinear,
)


# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
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


# 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)


# 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)


def _prepare_aspect_ratio_attention_mask(
    aspect_ratio_mask: torch.Tensor,
    num_patches: int,
    target_length: int,
    dtype: torch.dtype,
) -> torch.Tensor:
    # Expand aspect ratio mask to target_length
    batch_size, max_num_tiles = aspect_ratio_mask.shape
    attention_mask = aspect_ratio_mask.view(batch_size, max_num_tiles, 1, 1).to(dtype)
    attention_mask = attention_mask.repeat(1, 1, target_length, 1)

    # Mask padding patches
    pad_patches = target_length - num_patches
    attention_mask[:, :, -pad_patches:] = 0

    # Invert the mask (0 -> 1, 1 -> 0)
    attention_mask = 1 - attention_mask

    # Reshape to 2D and create 4D attention mask
    # (batch_size, 1, max_num_tiles * target_length, max_num_tiles * target_length)
    attention_mask = attention_mask.reshape(
        batch_size, max_num_tiles * target_length, 1
    )
    attention_mask = (
        attention_mask @ attention_mask.transpose(-1, -2) * torch.finfo(dtype).min
    )
    attention_mask = attention_mask.unsqueeze(1)

    return attention_mask


# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
def _prepare_4d_causal_attention_mask_with_cache_position(
    attention_mask: torch.Tensor,
    sequence_length: int,
    target_length: int,
    dtype: torch.dtype,
    device: torch.device,
    min_dtype: float,
    cache_position: torch.Tensor,
    batch_size: int,
):
    """
    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.
        device (`torch.device`):
            The device to plcae the 4D attention mask on.
        min_dtype (`float`):
            The minimum value representable with the dtype `dtype`.
        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:
        causal_mask = torch.full(
            (sequence_length, target_length),
            fill_value=min_dtype,
            dtype=dtype,
            device=device,
        )
        if sequence_length != 1:
            causal_mask = torch.triu(causal_mask, diagonal=1)
        causal_mask *= torch.arange(
            target_length, device=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, :]
            )
            padding_mask = padding_mask == 0
            causal_mask[:, :, :, :mask_length] = causal_mask[
                :, :, :, :mask_length
            ].masked_fill(padding_mask, min_dtype)

    return causal_mask


def _prepare_cross_attention_mask(
    cross_attention_mask: torch.Tensor,
    num_vision_tokens: int,
    dtype: str,
) -> Tuple[torch.Tensor, torch.Tensor]:
    # reshape so it can be used by attn module
    batch_size, text_total_length, *_ = cross_attention_mask.shape
    cross_attention_mask = cross_attention_mask.repeat_interleave(
        num_vision_tokens, dim=3
    )
    cross_attention_mask = cross_attention_mask.view(batch_size, text_total_length, -1)
    cross_attention_mask = cross_attention_mask.unsqueeze(1)

    # invert the mask
    inverted_cross_attn_mask = (1.0 - cross_attention_mask).to(dtype)
    cross_attention_mask = inverted_cross_attn_mask.masked_fill(
        inverted_cross_attn_mask.to(torch.bool), torch.finfo(dtype).min
    )

    # apply full-row bias, which return 4D tensor of shape [B, H, S1, 1] where value is 0 if the a full row in cross attn mask's
    # last dimension contains negative infinity values, otherwise it's 1
    negative_inf_value = torch.finfo(dtype).min
    full_text_row_masked_out_mask = (
        (cross_attention_mask != negative_inf_value)
        .any(dim=-1)
        .type_as(cross_attention_mask)[..., None]
    )
    cross_attention_mask *= full_text_row_masked_out_mask

    return cross_attention_mask, full_text_row_masked_out_mask


# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->MllamaVision
class MllamaVisionMLP(nn.Module):
    def __init__(self, *, prefix, config, weights):
        super().__init__()
        self.config = config
        self.activation_fn = ACT2FN[config.hidden_act]
        self.fc1 = TensorParallelColumnLinear.load(
            prefix=f"{prefix}.fc1", weights=weights, config=config, bias=True
        )
        self.fc2 = TensorParallelRowLinear.load(
            prefix=f"{prefix}.fc2", weights=weights, config=config, bias=True
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states = self.fc2(hidden_states)
        return hidden_states


class MllamaVisionSdpaAttention(nn.Module):
    def __init__(self, *, prefix, config, weights):
        super().__init__()

        self.embed_dim = config.hidden_size
        self.head_dim = config.hidden_size // config.attention_heads
        self.num_heads = config.attention_heads // weights.process_group.size()

        self.qkv_proj = TensorParallelColumnLinear.load_multi(
            config,
            prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"],
            dim=0,
            weights=weights,
            bias=False,
        )
        self.o_proj = TensorParallelRowLinear.load(
            config,
            prefix=f"{prefix}.o_proj",
            weights=weights,
            bias=False,
        )

    def forward(
        self,
        hidden_state: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        qkv = self.qkv_proj(hidden_state)
        query, key, value = qkv.split(
            [
                self.head_dim * self.num_heads,
                self.head_dim * self.num_heads,
                self.head_dim * self.num_heads,
            ],
            dim=2,
        )

        batch_size, q_seq_len, _ = query.shape
        _, kv_seq_len, _ = key.shape

        query = query.view(batch_size, q_seq_len, self.num_heads, self.head_dim)
        key = key.view(batch_size, kv_seq_len, self.num_heads, self.head_dim)
        value = value.view(batch_size, kv_seq_len, self.num_heads, self.head_dim)

        query = query.transpose(1, 2)
        key = key.transpose(1, 2)
        value = value.transpose(1, 2)

        attn_output = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask
        )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(batch_size, q_seq_len, -1)

        output = self.o_proj(attn_output)
        return output


class MllamaVisionEncoderLayer(nn.Module):
    def __init__(self, *, prefix, config, weights, is_gated: bool):
        super().__init__()

        self.hidden_size = config.hidden_size
        self.num_attention_heads = config.attention_heads
        self.is_gated = is_gated
        self.intermediate_size = config.intermediate_size

        self.self_attn = MllamaVisionSdpaAttention(
            prefix=f"{prefix}.self_attn", config=config, weights=weights
        )
        self.mlp = MllamaVisionMLP(
            prefix=f"{prefix}.mlp", config=config, weights=weights
        )

        self.input_layernorm = nn.LayerNorm.load(
            prefix=f"{prefix}.input_layernorm", weights=weights, eps=1e-05
        )
        self.post_attention_layernorm = nn.LayerNorm.load(
            prefix=f"{prefix}.post_attention_layernorm", weights=weights, eps=1e-05
        )

        # there used to be an if else here, no code path
        if is_gated:
            self.gate_attn = nn.Parameter(
                weights.get_tensor(f"{prefix}.gate_attn"), requires_grad=False
            )
            self.gate_ffn = nn.Parameter(
                weights.get_tensor(f"{prefix}.gate_ffn"), requires_grad=False
            )

    def forward(
        self,
        hidden_state: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
    ):
        # Self Attention
        residual = hidden_state
        hidden_state = self.input_layernorm(hidden_state)
        hidden_state = self.self_attn(hidden_state, attention_mask=attention_mask)
        gate_attn = 1 if not self.is_gated else self.gate_attn.tanh()
        hidden_state = residual + gate_attn * hidden_state

        # Feed forward
        residual = hidden_state
        hidden_state = self.post_attention_layernorm(hidden_state)
        hidden_state = self.mlp(hidden_state)
        gate_ffn = 1 if not self.is_gated else self.gate_ffn.tanh()
        hidden_state = residual + gate_ffn * hidden_state
        return hidden_state


class MllamaVisionEncoder(nn.Module):
    def __init__(self, *, prefix, config, weights, is_gated: bool, num_layers: int):
        super().__init__()
        self.config = config
        self.layers = [
            MllamaVisionEncoderLayer(
                prefix=f"{prefix}.layers.{i}",
                config=config,
                weights=weights,
                is_gated=is_gated,
            )
            for i in range(num_layers)
        ]

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
    ):
        encoder_states = [hidden_states]
        for encoder_layer in self.layers:
            layer_outputs = encoder_layer(
                hidden_states,
                attention_mask,
            )

            hidden_states = layer_outputs
            encoder_states.append(hidden_states)

        return hidden_states, encoder_states


class MllamaPrecomputedAspectRatioEmbedding(nn.Module):
    def __init__(self, *, prefix, config, weights):
        super().__init__()
        self.max_num_tiles = config.max_num_tiles
        self.hidden_size = config.hidden_size
        self.max_aspect_ratio_id = config.max_aspect_ratio_id

        self.embedding = TensorParallelEmbedding(
            prefix=f"{prefix}.embedding", weights=weights
        )
        self.gate = nn.Parameter(
            weights.get_tensor(f"{prefix}.gate"), requires_grad=False
        )

    def forward(
        self, hidden_state: torch.Tensor, aspect_ratio_ids: torch.Tensor
    ) -> torch.Tensor:
        embeddings = self.embedding(aspect_ratio_ids)
        embeddings = embeddings.reshape(-1, self.max_num_tiles, 1, self.hidden_size)

        # Always gated.
        embeddings = embeddings * self.gate.tanh()

        hidden_state = hidden_state + embeddings
        return hidden_state


class MllamaPrecomputedPositionEmbedding(nn.Module):
    def __init__(self, *, prefix, config, weights):
        super().__init__()
        self.max_num_tiles = config.max_num_tiles
        self.max_aspect_ratio_id = config.max_aspect_ratio_id
        self.num_patches = (config.image_size // config.patch_size) ** 2 + 1
        self.hidden_size = config.hidden_size
        self.scale = config.hidden_size**-0.5

        self.gate = nn.Parameter(
            weights.get_tensor(f"{prefix}.gate"), requires_grad=False
        )

        # position embedding
        embedding = nn.Parameter(
            weights.get_tensor(f"{prefix}.embedding"), requires_grad=False
        )
        self.gated_position_embedding = (1 - self.gate.tanh()) * embedding
        self.tile_embedding = TensorParallelEmbedding(
            prefix=f"{prefix}.tile_embedding", weights=weights
        )

    def forward(
        self, hidden_state: torch.Tensor, aspect_ratio_ids: torch.Tensor
    ) -> torch.Tensor:
        # position embeddings
        hidden_state = hidden_state + self.gated_position_embedding.view(
            1, 1, self.num_patches, self.hidden_size
        )

        # precomputed tile position embeddings
        tile_position_embedding = self.tile_embedding(aspect_ratio_ids)
        batch_size = hidden_state.shape[0]
        tile_position_embedding = tile_position_embedding.reshape(
            batch_size, self.max_num_tiles, self.num_patches, self.hidden_size
        )
        gated_tile_position_embedding = self.gate.tanh() * tile_position_embedding
        hidden_state = hidden_state + gated_tile_position_embedding

        return hidden_state


class MllamaVisionModel(nn.Module):
    def __init__(self, *, prefix, config, weights):
        super().__init__()
        self.image_size = config.image_size
        self.patch_size = config.patch_size
        self.max_num_tiles = config.max_num_tiles
        self.hidden_size = config.hidden_size
        self.in_channels = config.in_channels
        self.intermediate_layers_indices = config.intermediate_layers_indices

        self.num_patches = (self.image_size // self.patch_size) ** 2 + 1
        self.scale = config.hidden_size**-0.5
        self.dtype = weights.dtype

        self.patch_embedding = nn.Conv2d(
            in_channels=config.in_channels,
            out_channels=self.hidden_size,
            kernel_size=self.patch_size,
            stride=self.patch_size,
            padding="valid",
            bias=False,
        )
        self.patch_embedding.weight = nn.Parameter(
            weights.get_tensor(f"{prefix}.patch_embedding.weight"), requires_grad=False
        )

        self.class_embedding = nn.Parameter(
            weights.get_tensor(f"{prefix}.class_embedding"), requires_grad=False
        )

        self.gated_positional_embedding = MllamaPrecomputedPositionEmbedding(
            prefix=f"{prefix}.gated_positional_embedding",
            config=config,
            weights=weights,
        )

        self.pre_tile_positional_embedding = MllamaPrecomputedAspectRatioEmbedding(
            prefix=f"{prefix}.pre_tile_positional_embedding",
            config=config,
            weights=weights,
        )
        self.post_tile_positional_embedding = MllamaPrecomputedAspectRatioEmbedding(
            prefix=f"{prefix}.post_tile_positional_embedding",
            config=config,
            weights=weights,
        )

        ## layer norms
        self.layernorm_pre = nn.LayerNorm.load(
            prefix=f"{prefix}.layernorm_pre",
            weights=weights,
            # torch default
            eps=1e-05,
        )
        self.layernorm_post = nn.LayerNorm.load(
            prefix=f"{prefix}.layernorm_post",
            weights=weights,
            # torch default
            eps=1e-05,
        )

        ## encoders
        self.transformer = MllamaVisionEncoder(
            prefix=f"{prefix}.transformer",
            config=config,
            weights=weights,
            is_gated=False,
            num_layers=config.num_hidden_layers,
        )
        self.global_transformer = MllamaVisionEncoder(
            prefix=f"{prefix}.global_transformer",
            config=config,
            weights=weights,
            is_gated=True,
            num_layers=config.num_global_layers,
        )

    def apply_class_embedding(self, hidden_state: torch.Tensor) -> torch.Tensor:
        batch_size, _, hidden_size = hidden_state.shape
        class_embedding = self.class_embedding.expand(batch_size, 1, hidden_size)
        hidden_state = torch.cat([class_embedding, hidden_state], dim=1)
        return hidden_state

    def forward(
        self,
        pixel_values: torch.Tensor,
        aspect_ratio_ids: torch.Tensor,
        attention_mask: torch.Tensor,
    ) -> torch.Tensor:
        batch_size, num_concurrent_media, num_tiles, num_channels, height, width = (
            pixel_values.shape
        )

        pixel_values = pixel_values.reshape(
            batch_size * num_concurrent_media * num_tiles, num_channels, height, width
        )
        aspect_ratio_ids = aspect_ratio_ids.reshape(
            batch_size * num_concurrent_media, -1
        )

        # patch embedding
        patch_embeds = self.patch_embedding(pixel_values)
        hidden_state = patch_embeds.flatten(2).transpose(1, 2)

        # tile embeddings
        _, num_patches, dim = hidden_state.shape
        hidden_state = hidden_state.reshape(
            batch_size * num_concurrent_media, num_tiles, -1, dim
        )
        hidden_state = self.pre_tile_positional_embedding(
            hidden_state, aspect_ratio_ids
        )

        # apply cls token
        hidden_state = hidden_state.reshape(
            batch_size * num_concurrent_media * num_tiles, num_patches, dim
        )
        hidden_state = self.apply_class_embedding(hidden_state)
        num_patches += 1

        # apply position embeddings
        hidden_state = hidden_state.reshape(
            batch_size * num_concurrent_media, num_tiles, num_patches, dim
        )
        hidden_state = self.gated_positional_embedding(hidden_state, aspect_ratio_ids)

        # apply encoder
        hidden_state = self.layernorm_pre(hidden_state)

        # Compute the number of tokens to pad
        num_padding_patches = (8 - (hidden_state.shape[-2] % 8)) % 8
        # Compute padding tuple for pad function
        padding = (
            0,
            0,
            0,
            num_padding_patches,
        )  # (pad_left, pad_right, pad_left for dim -2, pad_right for dim -2)
        # Pad the tensor
        hidden_state = F.pad(hidden_state, padding, mode="constant", value=0)
        slice_index = -num_padding_patches if num_padding_patches > 0 else None

        if attention_mask is not None:
            attention_mask = attention_mask.reshape(
                batch_size * num_concurrent_media, -1
            )
            attention_mask = _prepare_aspect_ratio_attention_mask(
                aspect_ratio_mask=attention_mask,
                num_patches=self.num_patches,
                target_length=hidden_state.shape[2],
                dtype=self.dtype,
            )

        hidden_state = hidden_state.view(batch_size * num_concurrent_media, -1, dim)
        hidden_state, all_intermediate_hidden_states = self.transformer(
            hidden_state,
            attention_mask=attention_mask,
        )
        intermediate_hidden_states = [
            hidden_state
            for idx, hidden_state in enumerate(all_intermediate_hidden_states)
            if idx in self.intermediate_layers_indices
        ]
        intermediate_hidden_states = torch.stack(intermediate_hidden_states, dim=-1)

        # apply global encoder
        hidden_state = self.layernorm_post(hidden_state)
        hidden_state = hidden_state.reshape(
            batch_size * num_concurrent_media,
            num_tiles,
            num_patches + num_padding_patches,
            dim,
        )
        hidden_state = self.post_tile_positional_embedding(
            hidden_state, aspect_ratio_ids
        )
        hidden_state = hidden_state.reshape(
            batch_size * num_concurrent_media,
            num_tiles * (num_patches + num_padding_patches),
            dim,
        )
        hidden_state, _ = self.global_transformer(
            hidden_state, attention_mask=attention_mask
        )
        hidden_state = hidden_state.reshape(
            batch_size * num_concurrent_media,
            num_tiles,
            num_patches + num_padding_patches,
            dim,
        )
        hidden_state = hidden_state[:, :, :slice_index]

        # adding intermediate layer outputs
        hidden_state = hidden_state.reshape(
            batch_size, num_concurrent_media, num_tiles, num_patches, dim
        )
        intermediate_hidden_states = intermediate_hidden_states.reshape(
            batch_size * num_concurrent_media,
            num_tiles,
            num_patches + num_padding_patches,
            -1,
        )
        intermediate_hidden_states = intermediate_hidden_states[:, :, :slice_index]
        intermediate_hidden_states = intermediate_hidden_states.reshape(
            batch_size, num_concurrent_media, num_tiles, num_patches, -1
        )
        hidden_state = torch.cat([hidden_state, intermediate_hidden_states], dim=-1)
        return hidden_state


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

    def __init__(self, *, prefix, config, weights, layer_idx):
        super().__init__()
        self.config = config
        self.num_heads = self.config.num_attention_heads
        self.num_key_value_heads = self.config.num_key_value_heads
        self.dropout = config.dropout
        self.hidden_size = config.hidden_size
        self.head_dim = config.hidden_size // self.num_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.layer_idx = layer_idx

        self.num_heads = self.num_heads // weights.process_group.size()
        self.num_key_value_heads = (
            self.num_key_value_heads // weights.process_group.size()
        )

        self.q_proj = TensorParallelColumnLinear.load(
            config,
            prefix=f"{prefix}.q_proj",
            weights=weights,
            bias=False,
        )
        self.k_proj = TensorParallelColumnLinear.load(
            config,
            prefix=f"{prefix}.k_proj",
            weights=weights,
            bias=False,
        )
        self.v_proj = TensorParallelColumnLinear.load(
            config,
            prefix=f"{prefix}.v_proj",
            weights=weights,
            bias=False,
        )
        self.o_proj = TensorParallelRowLinear.load(
            config,
            prefix=f"{prefix}.o_proj",
            weights=weights,
            bias=False,
        )

        self.q_norm = MllamaTextRMSNorm.load(
            prefix=f"{prefix}.q_norm", weights=weights, eps=config.rms_norm_eps
        )
        self.k_norm = MllamaTextRMSNorm.load(
            prefix=f"{prefix}.k_norm", weights=weights, eps=config.rms_norm_eps
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        cross_attention_states: Optional[torch.Tensor] = None,
        past_key_value=None,
        attention_mask: Optional[torch.Tensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""
        bsz, q_len, _ = hidden_states.size()
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(
            bsz, q_len, self.num_heads, self.head_dim
        ).transpose(1, 2)
        query_states = self.q_norm(query_states)

        if cross_attention_states is not None:
            key_states = self.k_proj(cross_attention_states)
            value_states = self.v_proj(cross_attention_states)
            key_states = key_states.view(
                bsz, -1, self.num_key_value_heads, self.head_dim
            ).transpose(1, 2)
            value_states = value_states.view(
                bsz, -1, self.num_key_value_heads, self.head_dim
            ).transpose(1, 2)
            key_states = repeat_kv(key_states, self.num_key_value_groups)
            value_states = repeat_kv(value_states, self.num_key_value_groups)

            key_states = self.k_norm(key_states)

            if past_key_value is not None:
                # if we have a new image + new tokens, we only computed key_states on that new image
                # we still update the cross key states, past_image, new_image. And use it!
                key_states, value_states = past_key_value.update(
                    key_states,
                    value_states,
                    self.layer_idx,
                    {"cache_position": cache_position},
                )

        elif cache_position[0] != 0:
            key_states, value_states = (
                past_key_value.key_cache[self.layer_idx],
                past_key_value.value_cache[self.layer_idx],
            )
        else:
            raise ValueError(
                "Cross attention layer can't find neither `cross_attn_states` nor cached values for key/values!"
            )

        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[:, :, :, : key_states.shape[-2]]
            attn_weights = attn_weights + causal_mask

        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.dropout, training=self.training
        )
        attn_output = torch.matmul(attn_weights, value_states)
        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, -1)
        attn_output = self.o_proj(attn_output)

        return attn_output, attn_weights, past_key_value


# Copied from transformers.models.gemma2.modeling_gemma2.Gemma2MLP with Gemma2->MllamaText
class MllamaTextMLP(nn.Module):
    def __init__(self, *, prefix, config, weights):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = (
            config.intermediate_size // weights.process_group.size()
        )
        self.gate_up_proj = TensorParallelColumnLinear.load_multi(
            config,
            prefixes=[f"{prefix}.gate_proj", f"{prefix}.up_proj"],
            weights=weights,
            dim=0,
            bias=False,
        )
        self.down_proj = TensorParallelRowLinear.load(
            config,
            prefix=f"{prefix}.down_proj",
            weights=weights,
            bias=False,
        )
        self.act_fn = ACT2FN[config.hidden_activation]

    def forward(self, x):
        shape = x.shape
        gate_up_states = self.gate_up_proj(x)
        gate_up_states = gate_up_states.view(*shape[:-1], 2, self.intermediate_size)
        result = self.down_proj(
            self.act_fn(gate_up_states[:, :, 0]) * gate_up_states[:, :, 1]
        )
        return result


class MllamaCrossAttentionDecoderLayer(torch.nn.Module):
    """Cross-attention transformer block with tanh-gated attention and feedforward."""

    def __init__(self, *, prefix, config, weights, layer_idx) -> None:
        super().__init__()
        self.cross_attn = MllamaTextCrossAttention(
            prefix=f"{prefix}.cross_attn",
            config=config,
            weights=weights,
            layer_idx=layer_idx,
        )

        self.input_layernorm = MllamaTextRMSNorm.load(
            prefix=f"{prefix}.input_layernorm", weights=weights, eps=config.rms_norm_eps
        )
        self.cross_attn_attn_gate = torch.nn.Parameter(
            weights.get_tensor(f"{prefix}.cross_attn_attn_gate"), requires_grad=False
        )

        self.mlp = MllamaTextMLP(prefix=f"{prefix}.mlp", config=config, weights=weights)
        self.post_attention_layernorm = MllamaTextRMSNorm.load(
            prefix=f"{prefix}.post_attention_layernorm",
            weights=weights,
            eps=config.rms_norm_eps,
        )
        self.cross_attn_mlp_gate = torch.nn.Parameter(
            weights.get_tensor(f"{prefix}.cross_attn_mlp_gate"), requires_grad=False
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        cross_attention_states: torch.Tensor,
        cross_attention_mask: torch.Tensor,
        attention_mask: torch.Tensor,
        full_text_row_masked_out_mask: Tuple[torch.Tensor, torch.Tensor],
        past_key_value=None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)

        hidden_states, attn_weights, past_key_value = self.cross_attn(
            hidden_states=hidden_states,
            attention_mask=cross_attention_mask,
            cross_attention_states=cross_attention_states,
            past_key_value=past_key_value,
            cache_position=cache_position,
        )
        hidden_states = residual + self.cross_attn_attn_gate.tanh() * hidden_states

        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        if full_text_row_masked_out_mask is not None:
            hidden_states = full_text_row_masked_out_mask[:, 0] * hidden_states  # type: ignore
        hidden_states = residual + self.cross_attn_mlp_gate.tanh() * hidden_states

        return hidden_states


class MllamaTextSelfAttention(nn.Module):
    def __init__(self, *, prefix, config, weights, layer_idx):
        super().__init__()
        self.config = config
        self.num_heads = config.num_attention_heads
        self.dropout = config.dropout
        self.hidden_size = config.hidden_size
        self.num_key_value_heads = config.num_key_value_heads
        self.head_dim = config.hidden_size // self.num_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads

        self.num_heads = self.num_heads // weights.process_group.size()
        self.num_key_value_heads = (
            self.num_key_value_heads // weights.process_group.size()
        )
        self.layer_idx = layer_idx

        self.qkv_proj = TensorParallelColumnLinear.load_multi(
            config,
            prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"],
            dim=0,
            weights=weights,
            bias=False,
        )
        self.o_proj = TensorParallelRowLinear.load(
            config,
            prefix=f"{prefix}.o_proj",
            weights=weights,
            bias=False,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        position_embeddings: torch.Tensor,
        past_key_value=None,
        cache_position=None,
        **kwargs,
    ):
        bsz, q_len, _ = hidden_states.size()

        qkv = self.qkv_proj(hidden_states)
        query_states, key_states, value_states = qkv.split(
            [
                self.head_dim * self.num_heads,
                self.head_dim * self.num_key_value_heads,
                self.head_dim * self.num_key_value_heads,
            ],
            dim=2,
        )

        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 = 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
            )

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        causal_mask = attention_mask
        if attention_mask is not None:
            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]

        # 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()

        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
        is_causal = True if causal_mask is None and q_len > 1 else False

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            # TODO
            # attn_mask=causal_mask,
            dropout_p=self.dropout if self.training else 0.0,
            is_causal=is_causal,
        )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.view(bsz, q_len, -1)

        attn_output = self.o_proj(attn_output)
        return attn_output, None, past_key_value


# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->MllamaText
class MllamaTextRMSNorm(nn.Module):
    def __init__(self, weight, eps):
        super().__init__()
        self.weight = weight
        self.variance_epsilon = eps

    @classmethod
    def load(cls, *, prefix, weights, eps):
        weight = nn.Parameter(
            weights.get_tensor(f"{prefix}.weight"), requires_grad=False
        )
        return cls(weight=weight, eps=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)

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


# Copied from transformers.models.llama.modeling_llama.LlamaDecoderLayer with LlamaDecoder->MllamaSelfAttentionDecoder, Llama->MllamaText, LLAMA->MLLAMA_TEXT
class MllamaSelfAttentionDecoderLayer(nn.Module):
    def __init__(self, *, prefix, config, weights, layer_idx):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = MllamaTextSelfAttention(
            prefix=f"{prefix}.self_attn",
            config=config,
            weights=weights,
            layer_idx=layer_idx,
        )

        self.mlp = MllamaTextMLP(prefix=f"{prefix}.mlp", config=config, weights=weights)
        self.input_layernorm = MllamaTextRMSNorm.load(
            prefix=f"{prefix}.input_layernorm", weights=weights, eps=config.rms_norm_eps
        )
        self.post_attention_layernorm = MllamaTextRMSNorm.load(
            prefix=f"{prefix}.post_attention_layernorm",
            weights=weights,
            eps=config.rms_norm_eps,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value=None,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[
            Tuple[torch.Tensor, torch.Tensor]
        ] = None,  # will become mandatory in v4.45
        **kwargs,
    ) -> 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, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            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

        return hidden_states


class MllamaRotaryEmbedding(nn.Module):
    def __init__(
        self,
        *,
        config,
        weights,
    ):
        super().__init__()
        device = weights.device
        self.rope_type = config.rope_scaling["rope_type"]
        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)
        inv_freq.to(device=device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            inv_freq, self.attention_scaling = self.rope_init_fn(
                self.config, device, seq_len=seq_len, **self.rope_kwargs
            )
            self.register_buffer(
                "inv_freq", inv_freq, persistent=False
            )  # TODO joao: may break with compilation
            self.max_seq_len_cached = seq_len

        if (
            seq_len < self.original_max_seq_len
            and self.max_seq_len_cached > self.original_max_seq_len
        ):  # reset
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.original_max_seq_len

    @torch.no_grad()
    def forward(self, x, position_ids):
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # Core RoPE block
        inv_freq_expanded = (
            self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        )
        position_ids_expanded = position_ids[:, None, :].float()
        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
        device_type = x.device.type
        device_type = (
            device_type
            if isinstance(device_type, str) and device_type != "mps"
            else "cpu"
        )

        freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(
            1, 2
        )
        emb = torch.cat((freqs, freqs), dim=-1)
        cos = emb.cos()
        sin = emb.sin()

        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
        cos = cos * self.attention_scaling
        sin = sin * self.attention_scaling

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


class MllamaTextModel(nn.Module):
    def __init__(self, *, prefix, config, weights):
        super().__init__()
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size
        self.config = config
        self.embed_tokens = TensorParallelEmbedding(
            prefix=f"{prefix}.embed_tokens", weights=weights
        )
        self.cross_attention_layers = config.cross_attention_layers

        self.layers = []
        for layer_idx in range(config.num_hidden_layers):
            if layer_idx in self.cross_attention_layers:
                self.layers.append(
                    MllamaCrossAttentionDecoderLayer(
                        prefix=f"{prefix}.layers.{layer_idx}",
                        config=config,
                        weights=weights,
                        layer_idx=layer_idx,
                    )
                )
            else:
                self.layers.append(
                    MllamaSelfAttentionDecoderLayer(
                        prefix=f"{prefix}.layers.{layer_idx}",
                        config=config,
                        weights=weights,
                        layer_idx=layer_idx,
                    )
                )

        # TODO Should we use this slow norm ?
        # self.norm = MllamaTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.norm = MllamaTextRMSNorm.load(
            prefix=f"{prefix}.norm",
            weights=weights,
            eps=config.rms_norm_eps,
        )
        # TODO Anything specific ?
        head_size = config.hidden_size // config.num_attention_heads
        self.rotary_emb = MllamaRotaryEmbedding(config=config, weights=weights)

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        cross_attention_states: Optional[torch.FloatTensor] = None,
        cross_attention_mask: Optional[torch.Tensor] = None,
        full_text_row_masked_out_mask: Optional[
            Tuple[torch.Tensor, torch.Tensor]
        ] = None,
        past_key_values=None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ):
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
            )

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

        hidden_states = inputs_embeds

        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,
        )

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

        # decoder layers

        for idx, decoder_layer in enumerate(self.layers):
            if (
                idx in self.cross_attention_layers
                and cross_attention_states is None
                and (
                    past_key_values is None
                    or (
                        past_key_values is not None
                        and past_key_values.get_seq_length(idx) == 0
                    )
                )
            ):
                continue

            layer_outputs = decoder_layer(
                hidden_states,
                cross_attention_states=cross_attention_states,
                cross_attention_mask=cross_attention_mask,
                attention_mask=causal_mask,
                full_text_row_masked_out_mask=full_text_row_masked_out_mask,
                position_ids=position_ids,
                past_key_value=past_key_values,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
            )

            hidden_states = layer_outputs

        hidden_states = self.norm(hidden_states)

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

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and 0.0 in attention_mask:
                return attention_mask
            return None

        # 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_static_cache = isinstance(past_key_values, StaticCache)

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        # TODO: we have only SDPA currently and there's a bug when attn-bias is passed. Need to add eager attn and return the line
        # self.config._attn_implementation == "sdpa" and
        # if self.config._attn_implementation == "sdpa" and not using_static_cache:
        if self.config._attn_implementation == "sdpa" and not using_static_cache:
            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, device = input_tensor.dtype, input_tensor.device
        min_dtype = torch.finfo(dtype).min
        sequence_length = input_tensor.shape[1]
        if using_static_cache:
            target_length = past_key_values.get_max_length()
        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 = _prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            min_dtype=min_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 == "cuda"
        ):
            # 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
            causal_mask = AttentionMaskConverter._unmask_unattended(
                causal_mask, min_dtype
            )

        return causal_mask


class MllamaForCausalLM(nn.Module):
    def __init__(self, *, prefix, config, weights):
        super().__init__()
        self.vocab_size = config.vocab_size
        self.model = MllamaTextModel(
            prefix=f"{prefix}.model", config=config, weights=weights
        )
        self.lm_head = SpeculativeHead.load(
            prefix=f"{prefix}.lm_head",
            config=config,
            weights=weights,
        )

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        cross_attention_states: Optional[torch.LongTensor] = None,
        cross_attention_mask: Optional[torch.LongTensor] = None,
        full_text_row_masked_out_mask: Optional[
            Tuple[torch.Tensor, torch.Tensor]
        ] = None,
        past_key_values=None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        num_logits_to_keep: int = 0,
    ):
        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        # TODO
        outputs = self.model(
            input_ids=input_ids,
            cross_attention_states=cross_attention_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            cross_attention_mask=cross_attention_mask,
            full_text_row_masked_out_mask=full_text_row_masked_out_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            cache_position=cache_position,
        )

        hidden_states = outputs.last_hidden_state
        # if lm_head_indices is not None:
        #     hidden_states = hidden_states[lm_head_indices]
        logits, speculative_logits = self.lm_head(hidden_states)
        return (
            CausalLMOutputWithPast(
                logits=logits,
                past_key_values=outputs.past_key_values,
            ),
            speculative_logits,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        num_logits_to_keep=None,
        **kwargs,
    ):
        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
        # Exception 1: when passing input_embeds, input_ids may be missing entries
        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
        if past_key_values is not None:
            if inputs_embeds is not None:  # Exception 1
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif (
                input_ids.shape[1] != cache_position.shape[0]
            ):  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]

        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] :]

                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
                position_ids = position_ids.clone(memory_format=torch.contiguous_format)

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and cache_position[0] == 0:
            model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None}
        else:
            # The clone here is for the same reason as for `position_ids`.
            model_inputs = {
                "input_ids": input_ids.clone(memory_format=torch.contiguous_format),
                "inputs_embeds": None,
            }

        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
            if model_inputs["inputs_embeds"] is not None:
                batch_size, sequence_length, _ = model_inputs["inputs_embeds"].shape
                device = model_inputs["inputs_embeds"].device
            else:
                batch_size, sequence_length = model_inputs["input_ids"].shape
                device = model_inputs["input_ids"].device

            dtype = self.lm_head.weight.dtype
            min_dtype = torch.finfo(dtype).min

            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
                attention_mask,
                sequence_length=sequence_length,
                target_length=past_key_values.get_max_length(),
                dtype=dtype,
                device=device,
                min_dtype=min_dtype,
                cache_position=cache_position,
                batch_size=batch_size,
            )

        if num_logits_to_keep is not None:
            model_inputs["num_logits_to_keep"] = num_logits_to_keep

        model_inputs.update(
            {
                "position_ids": position_ids,
                "cache_position": cache_position,
                "past_key_values": past_key_values,
                "attention_mask": attention_mask,
            }
        )
        return model_inputs


class MllamaForConditionalGeneration(nn.Module):
    def __init__(self, prefix, config, weights):
        super().__init__()
        config.vision_config.quantize = None
        config.vision_config.speculator = config.speculator
        config.text_config.quantize = config.quantize
        config.text_config.speculator = config.speculator
        # TODO check how this is determined
        config.text_config._attn_implementation = "sdpa"
        # self.hidden_size = (
        #     config.text_config.hidden_size // weights.process_group.size()
        # )
        self.hidden_size = config.text_config.hidden_size
        self.vision_model = MllamaVisionModel(
            prefix="vision_model", config=config.vision_config, weights=weights
        )
        self.language_model = MllamaForCausalLM(
            prefix="language_model", config=config.text_config, weights=weights
        )
        self.multi_modal_projector = FastLinear.load(
            prefix="multi_modal_projector", config=config, weights=weights, bias=True
        )
        self.config = config
        self.dtype = weights.dtype
        self.device = weights.device

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        aspect_ratio_mask: Optional[List[List[int]]] = None,
        aspect_ratio_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[List[List[List[int]]]] = None,
        cross_attention_mask: Optional[torch.Tensor] = None,
        cross_attention_states: 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,
        cache_position: Optional[torch.LongTensor] = None,
        num_logits_to_keep: int = 0,
        image_hidden_states=None,
        image_attention_mask=None,
    ):
        if past_key_values is None:
            past_key_values = DynamicCache(
                num_hidden_layers=self.config.text_config.num_hidden_layers
            )
        elif isinstance(past_key_values, list):
            past_key_values = DynamicCache.from_legacy_cache(past_key_values)

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
            )

        if pixel_values is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
            )

        if pixel_values is not None and cross_attention_states is not None:
            raise ValueError(
                "`pixel_values` and `cross_attention_states` cannot be provided simultaneously"
            )

        if pixel_values is not None:
            if aspect_ratio_ids is None:
                raise ValueError(
                    "`aspect_ratio_ids` must be provided if `pixel_values` is provided"
                )
            # get vision tokens from vision model

            vision_states = self.vision_model(
                pixel_values, aspect_ratio_ids, aspect_ratio_mask
            )
            cross_attention_states = self.multi_modal_projector(vision_states).reshape(
                -1, vision_states.shape[-2], self.hidden_size
            )

        if cross_attention_mask is not None:
            cross_attention_mask, full_text_row_masked_out_mask = (
                _prepare_cross_attention_mask(
                    cross_attention_mask,
                    num_vision_tokens=self.vision_model.num_patches,
                    dtype=self.dtype,
                )
            )
        else:
            full_text_row_masked_out_mask = None

        if cross_attention_mask is not None and cache_position is not None:
            cross_attention_mask = cross_attention_mask[:, :, cache_position]
            full_text_row_masked_out_mask = full_text_row_masked_out_mask[
                :, :, cache_position
            ]

        outputs = self.language_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            cross_attention_states=cross_attention_states,
            cross_attention_mask=cross_attention_mask,
            full_text_row_masked_out_mask=full_text_row_masked_out_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            labels=labels,
            cache_position=cache_position,
            num_logits_to_keep=num_logits_to_keep,
        )

        return outputs