unsloth_compiled_cache/unsloth_compiled_module_mllama.py

# Unsloth Zoo - Utilities for Unsloth
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU Lesser General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with this program.  If not, see <https://www.gnu.org/licenses/>.

import torch
from unsloth_zoo.loss_utils import fused_linear_cross_entropy

scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
@torch.compiler.disable(recursive = False)
def disable_compile_scaled_dot_product_attention(*args, **kwargs):
    return scaled_dot_product_attention(*args, **kwargs)
pass


torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False}
from torch import Tensor
import torch
from torch.nn import functional as F
from transformers.models.mllama.modeling_mllama import (F, math, Optional, Tuple, torch, nn, ACT2FN, Cache, ROPE_INIT_FUNCTIONS, MllamaTextConfig, MllamaVisionConfig)

@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
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!=torch.finfo(cross_attention_mask.dtype).min, full_text_row_masked_out_mask

@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
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!=torch.finfo(attention_mask.dtype).min

@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
def MllamaPrecomputedAspectRatioEmbedding_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)

    if self.is_gated:
        embeddings = embeddings * self.gate.tanh()

    hidden_state = hidden_state + embeddings
    return hidden_state

class MllamaPrecomputedAspectRatioEmbedding(nn.Module):
    def __init__(self, config: MllamaVisionConfig, is_gated: bool = True):
        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.is_gated = is_gated

        self.embedding = nn.Embedding(self.max_aspect_ratio_id + 1, self.max_num_tiles * self.hidden_size)
        if is_gated:
            self.gate = nn.Parameter(torch.zeros(1))

    def forward(self, hidden_state: torch.Tensor, aspect_ratio_ids: torch.Tensor) -> torch.Tensor:
        return MllamaPrecomputedAspectRatioEmbedding_forward(self, hidden_state, aspect_ratio_ids)


@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
def MllamaPrecomputedPositionEmbedding_forward(self, hidden_state: torch.Tensor, aspect_ratio_ids: torch.Tensor) -> torch.Tensor:
    # position embeddings
    gated_position_embedding = (1 - self.gate.tanh()) * self.embedding
    hidden_state = hidden_state + 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 MllamaPrecomputedPositionEmbedding(nn.Module):
    def __init__(self, config: MllamaVisionConfig):
        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(torch.zeros(1))

        # position embedding
        position_embedding = torch.randn(self.num_patches, self.hidden_size)
        self.embedding = nn.Parameter(self.scale * position_embedding)

        # tile position embedding
        self.tile_embedding = nn.Embedding(
            self.max_aspect_ratio_id + 1, self.max_num_tiles * self.num_patches * self.hidden_size
        )

    def forward(self, hidden_state: torch.Tensor, aspect_ratio_ids: torch.Tensor) -> torch.Tensor:
        return MllamaPrecomputedPositionEmbedding_forward(self, hidden_state, aspect_ratio_ids)


@torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
def MllamaVisionMLP_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 MllamaVisionMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.activation_fn = ACT2FN[config.hidden_act]
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)

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


@torch.compiler.disable(recursive = False)
def MllamaVisionAttention_forward(
    self,
    hidden_state: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    output_attentions: bool = None,
) -> torch.Tensor:
    query = self.q_proj(hidden_state)
    key = self.k_proj(hidden_state)
    value = self.v_proj(hidden_state)

    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).transpose(1, 2)
    key = key.view(batch_size, kv_seq_len, self.num_heads, self.head_dim).transpose(1, 2)
    value = value.view(batch_size, kv_seq_len, self.num_heads, self.head_dim).transpose(1, 2)

    attn_weights = torch.matmul(query, key.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.shape[-2]]
        attn_weights = attn_weights + causal_mask

    # upcast attention to fp32
    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_output = torch.matmul(attn_weights, value)

    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)

    if not output_attentions:
        attn_weights = None

    return output, attn_weights

class MllamaVisionAttention(nn.Module):
    def __init__(self, config: MllamaVisionConfig):
        super().__init__()

        self.embed_dim = config.hidden_size
        self.num_heads = config.attention_heads
        self.head_dim = config.hidden_size // config.attention_heads

        self.q_proj = nn.Linear(self.embed_dim, self.num_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.embed_dim, self.num_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.embed_dim, self.num_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.embed_dim, bias=False)

    def forward(
        self,
        hidden_state: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = None,
    ) -> torch.Tensor:
        return MllamaVisionAttention_forward(self, hidden_state, attention_mask, output_attentions)


@torch.compiler.disable(recursive = False)
def MllamaVisionSdpaAttention_forward(
    self,
    hidden_state: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    output_attentions: bool = None,
) -> torch.Tensor:
    # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
    if output_attentions: raise RuntimeError('Unsloth: Not supported')

    query = self.q_proj(hidden_state)
    key = self.k_proj(hidden_state)
    value = self.v_proj(hidden_state)

    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, None

class MllamaVisionSdpaAttention(MllamaVisionAttention):
    # Adapted from MllamaVisionAttention
    def forward(
        self,
        hidden_state: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = None,
    ) -> torch.Tensor:
        return MllamaVisionSdpaAttention_forward(self, hidden_state, attention_mask, output_attentions)


@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
def MllamaTextRMSNorm_forward(self, hidden_states):
    input_dtype = hidden_states.dtype
    hidden_states = hidden_states.to(torch.float32)
    variance = hidden_states.pow(2).mean(-1, keepdim=True)
    hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
    return self.weight * hidden_states.to(input_dtype)

class MllamaTextRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        MllamaTextRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        return MllamaTextRMSNorm_forward(self, hidden_states)

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


@torch.compiler.disable(recursive = False)
def MllamaTextCrossAttention_forward(
    self,
    hidden_states: torch.Tensor,
    cross_attention_states: Optional[torch.Tensor] = None,
    past_key_value: Optional[Cache] = None,
    attention_mask: Optional[torch.Tensor] = None,
    output_attentions: bool = False,
    use_cache: bool = 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)

    if not output_attentions:
        attn_weights = None

    return attn_output, attn_weights, past_key_value

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

    def __init__(
        self,
        config: Optional[MllamaTextConfig] = None,
        layer_idx: Optional[int] = None,
    ):
        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.layer_idx = layer_idx
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads

        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

        self.q_norm = MllamaTextRMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.k_norm = MllamaTextRMSNorm(self.head_dim, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cross_attention_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
        use_cache: bool = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        return MllamaTextCrossAttention_forward(self, hidden_states, cross_attention_states, past_key_value, attention_mask, output_attentions, use_cache, cache_position)


@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
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)


@torch.compiler.disable(recursive = False)
def MllamaTextCrossSdpaAttention_forward(
    self,
    hidden_states: torch.Tensor,
    cross_attention_states: Optional[torch.Tensor] = None,
    past_key_value: Optional[Cache] = None,
    attention_mask: Optional[torch.Tensor] = None,
    output_attentions: bool = False,
    use_cache: bool = None,
    cache_position: Optional[torch.LongTensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
    """Input shape: Batch x Time x Channel"""
    if output_attentions: raise RuntimeError('Unsloth: Not supported')

    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)

        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!"
        )

    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)

    # 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 attention_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 attention_mask is None and q_len > 1 else False

    attn_output = torch.nn.functional.scaled_dot_product_attention(
        query_states,
        key_states,
        value_states,
        attn_mask=attention_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.reshape(bsz, q_len, -1)
    attn_output = self.o_proj(attn_output)

    return attn_output, None, past_key_value

class MllamaTextCrossSdpaAttention(MllamaTextCrossAttention):
    """
    Mllama attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `MllamaTextCrossAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """

    # Adapted from MllamaTextCrossAttention.forward
    def forward(
        self,
        hidden_states: torch.Tensor,
        cross_attention_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
        use_cache: bool = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        return MllamaTextCrossSdpaAttention_forward(self, hidden_states, cross_attention_states, past_key_value, attention_mask, output_attentions, use_cache, cache_position)


@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
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)


@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
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.
        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


@torch.compiler.disable(recursive = False)
def MllamaTextSelfAttention_forward(
    self,
    hidden_states: torch.Tensor,
    attention_mask: torch.Tensor,
    position_embeddings: torch.Tensor,
    output_attentions: bool = False,
    use_cache: bool = False,
    past_key_value=None,
    cache_position=None,
    **kwargs,
):
    bsz, q_len, _ = hidden_states.size()

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

    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)

    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

    # upcast attention to fp32
    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
    attn_output = torch.matmul(attn_weights, value_states)

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

    attn_output = self.o_proj(attn_output)

    if not output_attentions:
        attn_weights = None

    return attn_output, attn_weights, past_key_value

class MllamaTextSelfAttention(nn.Module):
    def __init__(self, config: MllamaTextConfig, layer_idx: int):
        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.rope_theta = config.rope_theta
        self.layer_idx = layer_idx

        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        position_embeddings: torch.Tensor,
        output_attentions: bool = False,
        use_cache: bool = False,
        past_key_value=None,
        cache_position=None,
        **kwargs,
    ):
        return MllamaTextSelfAttention_forward(self, hidden_states, attention_mask, position_embeddings, output_attentions, use_cache, past_key_value, cache_position, **kwargs)


@torch.compiler.disable(recursive = False)
def MllamaTextSelfSdpaAttention_forward(
    self,
    hidden_states: torch.Tensor,
    attention_mask: torch.Tensor,
    position_embeddings: torch.Tensor,
    output_attentions: bool = False,
    use_cache: bool = False,
    past_key_value=None,
    cache_position=None,
    **kwargs,
):
    if output_attentions: raise RuntimeError('Unsloth: Not supported')

    bsz, q_len, _ = hidden_states.size()

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

    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,
        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

class MllamaTextSelfSdpaAttention(MllamaTextSelfAttention):
    # Adapted from MllamaTextSelfAttention
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        position_embeddings: torch.Tensor,
        output_attentions: bool = False,
        use_cache: bool = False,
        past_key_value=None,
        cache_position=None,
        **kwargs,
    ):
        return MllamaTextSelfSdpaAttention_forward(self, hidden_states, attention_mask, position_embeddings, output_attentions, use_cache, past_key_value, cache_position, **kwargs)


@torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
def MllamaTextMLP_forward(self, x):
    return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))

class MllamaTextMLP(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)
        # Ignore copy
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        return MllamaTextMLP_forward(self, x)


@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
@torch.no_grad()
def MllamaRotaryEmbedding_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"
    with torch.autocast(device_type=device_type, enabled=False):
        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 MllamaRotaryEmbedding(nn.Module):
    def __init__(self, config: MllamaTextConfig, device=None):
        super().__init__()
        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)
        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


    def forward(self, x, position_ids):
        return MllamaRotaryEmbedding_forward(self, x, position_ids)