Source code for torchtune.models.mistral._component_builders

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

from functools import partial
from torchtune.modules.common_utils import reparametrize_as_dtype_state_dict_post_hook
from typing import List

from torch import nn

from torchtune.modules import (
    CausalSelfAttention,
    FeedForward,
    RMSNorm,
    RotaryPositionalEmbeddings,
    TransformerDecoder,
    TransformerDecoderLayer,
)

from torchtune.modules.peft import LORA_ATTN_MODULES, LoRALinear

"""
Component builders for the Mistral 7B models and popular variants such as LoRA.

torchtune provides composable building blocks. Builder functions help
stitch these building blocks into higher-level components. This design has
two benefits:
- The building blocks themselves are very flexible. For example, ``CausalSelfAttention``
can take either nn.Linear or nn.LoRALinear for ``q_proj``.
- Builder functions expose a set of configurable params which keep the constructors of
the building blocks simple.
"""


def mistral(
    vocab_size: int,
    num_layers: int,
    num_heads: int,
    num_kv_heads: int,
    embed_dim: int,
    intermediate_dim: int,
    max_seq_len: int,
    attn_dropout: float = 0.0,
    norm_eps: float = 1e-5,
    rope_base: int = 10_000,
) -> TransformerDecoder:
    """
    Build the decoder associated with the mistral model. This includes:
    - Token embeddings
    - num_layers number of TransformerDecoderLayer blocks
    - RMS Norm layer applied to the output of the transformer
    - Final projection into token space

    This does NOT currently include inference-time optimizations such as
    sliding-window attention

    Args:
        vocab_size (int): number of tokens in vocabulary.
        num_layers (int): number of layers in the transformer decoder.
        num_heads (int): number of query heads. For MHA this is also the
            number of heads for key and value
        num_kv_heads (int): number of key and value heads. If specified,
            user should ensure `num_heads` % `num_kv_heads` == 0. Default value is
            `None`, in which case this is the same as MHA
        embed_dim (int): embedding dimension for self-attention
        intermediate_dim (int): intermediate dimension for MLP
        max_seq_len (int): maximum sequence length the model will be run with,
        attn_dropout (float): dropout value passed onto scaled_dot_product_attention.
            Default: 0.0
        norm_eps (float): epsilon in RMS norms
        rope_base (int): base for the rotary positional embeddings. Default: 10_000

    Returns:
        TransformerDecoder: Instantiation of mistral model.
    """
    head_dim = embed_dim // num_heads
    num_kv_heads = num_kv_heads if num_kv_heads else num_heads

    rope = RotaryPositionalEmbeddings(dim=head_dim, max_seq_len=max_seq_len, base=rope_base)
    self_attn = CausalSelfAttention(
        embed_dim=embed_dim,
        num_heads=num_heads,
        num_kv_heads=num_kv_heads,
        head_dim=head_dim,
        q_proj=nn.Linear(embed_dim, num_heads * head_dim, bias=False),
        k_proj=nn.Linear(embed_dim, num_kv_heads * head_dim, bias=False),
        v_proj=nn.Linear(embed_dim, num_kv_heads * head_dim, bias=False),
        output_proj=nn.Linear(embed_dim, embed_dim, bias=False),
        pos_embeddings=rope,
        kv_cache=None,
        max_seq_len=max_seq_len,
        attn_dropout=attn_dropout,
    )
    mlp = mistral_mlp(dim=embed_dim, hidden_dim=intermediate_dim)
    layer = TransformerDecoderLayer(
        attn=self_attn,
        mlp=mlp,
        sa_norm=RMSNorm(dim=embed_dim, eps=norm_eps),
        mlp_norm=RMSNorm(dim=embed_dim, eps=norm_eps),
    )
    tok_embeddings = nn.Embedding(vocab_size, embed_dim)
    output_proj = nn.Linear(embed_dim, vocab_size, bias=False)
    return TransformerDecoder(
        tok_embeddings=tok_embeddings,
        layer=layer,
        num_layers=num_layers,
        max_seq_len=max_seq_len,
        num_heads=num_heads,
        head_dim=head_dim,
        norm=RMSNorm(embed_dim, eps=norm_eps),
        output=output_proj,
    )


def mistral_mlp(dim: int, hidden_dim: int) -> FeedForward:
    """
    Build the MLP layer associated with the Mistral model.
    """
    gate_proj = nn.Linear(dim, hidden_dim, bias=False)
    down_proj = nn.Linear(hidden_dim, dim, bias=False)
    up_proj = nn.Linear(dim, hidden_dim, bias=False)
    return FeedForward(gate_proj=gate_proj, down_proj=down_proj, up_proj=up_proj)


def lora_mistral(
    lora_attn_modules: List[LORA_ATTN_MODULES],
    apply_lora_to_mlp: bool = False,
    apply_lora_to_output: bool = False,
    *,
    # mistral args
    vocab_size: int,
    num_layers: int,
    num_heads: int,
    num_kv_heads: int,
    embed_dim: int,
    max_seq_len: int,
    intermediate_dim: int,
    attn_dropout: float = 0.0,
    norm_eps: float = 1e-5,
    rope_base: int = 10_000,
    # LoRA args
    lora_rank: int,
    lora_alpha: float,
    lora_dropout: float = 0.0,
    quantize_base: bool = False,
) -> TransformerDecoder:
    """
    Return a version of Mistral (an instance of :func:`~torchtune.modules.TransformerDecoder`)
    with LoRA applied based on the passed in configuration.

    Args:
        lora_attn_modules (List[LORA_ATTN_MODULES]): list of which linear layers
            LoRA should be applied to in each self-attention block. Options are
            ``{"q_proj", "k_proj", "v_proj", "output_proj"}``.
        apply_lora_to_mlp (bool): whether to apply LoRA to the MLP in each transformer layer.
            Default: False
        apply_lora_to_output (bool): whether to apply LoRA to the model's final output projection.
            Default: False
        vocab_size (int): number of tokens in vocabulary.
        num_layers (int): number of layers in the transformer decoder.
        num_heads (int): number of query heads. For MHA this is also the
            number of heads for key and value
        num_kv_heads (int): number of key and value heads. If specified,
            user should ensure `num_heads` % `num_kv_heads` == 0. Default value is
            `None`, in which case this is the same as MHA
        embed_dim (int): embedding dimension for self-attention
        max_seq_len (int): maximum sequence length the model will be run with
        intermediate_dim (int): intermediate dimension for MLP.
        attn_dropout (float): dropout value passed onto scaled_dot_product_attention.
            Default: 0.0
        norm_eps (float): epsilon in RMS norms.
        rope_base (int): base for the rotary positional embeddings. Default: 10_000
        lora_rank (int): rank of each low-rank approximation
        lora_alpha (float): scaling factor for the low-rank approximation
        lora_dropout (float): LoRA dropout probability. Default: 0.0
        quantize_base: (bool): Whether to quantize base model weights or not. Only applied to base
            weights within linear layers LoRA is applied to. The final output linear projection is not
            supported for quantization currently.

    Returns:
        TransformerDecoder: Instantiation of Mistral model with LoRA applied to
        a subset of the attention projections in each layer.

    """

    self_attn = lora_mistral_self_attention(
        lora_modules=lora_attn_modules,
        embed_dim=embed_dim,
        num_heads=num_heads,
        num_kv_heads=num_kv_heads,
        max_seq_len=max_seq_len,
        attn_dropout=attn_dropout,
        rope_base=rope_base,
        lora_rank=lora_rank,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        quantize_base=quantize_base,
    )

    if apply_lora_to_mlp:
        mlp = lora_mistral_mlp(
            dim=embed_dim,
            hidden_dim=intermediate_dim,
            lora_rank=lora_rank,
            lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
            quantize_base=quantize_base,
        )
    else:
        mlp = mistral_mlp(dim=embed_dim, hidden_dim=intermediate_dim)

    layer = TransformerDecoderLayer(
        attn=self_attn,
        mlp=mlp,
        sa_norm=RMSNorm(dim=embed_dim, eps=norm_eps),
        mlp_norm=RMSNorm(dim=embed_dim, eps=norm_eps),
    )

    tok_embeddings = nn.Embedding(vocab_size, embed_dim)

    # TODO: quantize_base is not applied to final output_proj currently.
    output_proj = (
        LoRALinear(embed_dim, vocab_size, rank=lora_rank, alpha=lora_alpha)
        if apply_lora_to_output
        else nn.Linear(embed_dim, vocab_size, bias=False)
    )
    model = TransformerDecoder(
        tok_embeddings=tok_embeddings,
        layer=layer,
        num_layers=num_layers,
        max_seq_len=max_seq_len,
        num_heads=num_heads,
        head_dim=(embed_dim // num_heads),
        norm=RMSNorm(embed_dim, eps=norm_eps),
        output=output_proj,
    )

    if quantize_base:
        # For QLoRA, we reparametrize 4-bit tensors to higher precision, and offload to CPU on the fly
        # so as to not increase peak memory
        model._register_state_dict_hook(
            partial(
                reparametrize_as_dtype_state_dict_post_hook,
                # TODO this is clowny, figure out a better way to get what precision the rest
                # of the model is in
                dtype=tok_embeddings.weight.dtype,
                offload_to_cpu=True,
            )
        )

    return model


def lora_mistral_self_attention(
    lora_modules: List[LORA_ATTN_MODULES],
    *,
    # CausalSelfAttention args
    embed_dim: int,
    num_heads: int,
    num_kv_heads: int,
    max_seq_len: int,
    attn_dropout: float = 0.0,
    rope_base: int = 10_000,
    # LoRA args
    lora_rank: int,
    lora_alpha: float,
    lora_dropout: float = 0.0,
    quantize_base: bool = False,
) -> CausalSelfAttention:
    """
    Return an instance of :func:`~torchtune.modules.CausalSelfAttention` with LoRA
    applied to a subset of its linear layers

    Args:
        lora_modules (List[LORA_ATTN_MODULES]): list of which linear layers
            LoRA should be applied to. Options are ``{"q_proj", "k_proj", "v_proj",
            "output_proj"}``.
        embed_dim (int): embedding dimension for self-attention
        num_heads (int): number of query heads. For MHA this is also the
            number of heads for key and value
        num_kv_heads (int): number of key and value heads. If specified,
            user should ensure `num_heads` % `num_kv_heads` == 0. Default value is
            `None`, in which case this is the same as MHA
        max_seq_len (int): maximum sequence length the model will be run with
        attn_dropout (float): dropout value passed onto scaled_dot_product_attention.
            Default: 0.0
        rope_base (int): base for the rotary positional embeddings. Default: 10_000
        lora_rank (int): rank of each low-rank approximation
        lora_alpha (float): scaling factor for the low-rank approximation
        lora_dropout (float): LoRA dropout probability. Default: 0.0
        quantize_base (bool): Whether to quantize base model parameters for linear layers
            LoRA is being applied to. Default is ``False``.

    Returns:
        CausalSelfAttention: instantiation of self-attention module with LoRA
        applied to a subset of Q, K, V, output projections.

    Raises:
        ValueError: If lora_modules arg is an empty list
    """
    if not lora_modules:
        raise ValueError(f"Must pass one or more of {LORA_ATTN_MODULES} as lora_modules")

    head_dim = embed_dim // num_heads
    num_kv_heads = num_kv_heads if num_kv_heads else num_heads

    q_proj = (
        LoRALinear(
            embed_dim,
            num_heads * head_dim,
            rank=lora_rank,
            alpha=lora_alpha,
            dropout=lora_dropout,
            quantize_base=quantize_base,
        )
        if "q_proj" in lora_modules
        else nn.Linear(embed_dim, num_heads * head_dim, bias=False)
    )
    k_proj = (
        LoRALinear(
            embed_dim,
            num_kv_heads * head_dim,
            rank=lora_rank,
            alpha=lora_alpha,
            dropout=lora_dropout,
            quantize_base=quantize_base,
        )
        if "k_proj" in lora_modules
        else nn.Linear(embed_dim, num_kv_heads * head_dim, bias=False)
    )
    v_proj = (
        LoRALinear(
            embed_dim,
            num_kv_heads * head_dim,
            rank=lora_rank,
            alpha=lora_alpha,
            dropout=lora_dropout,
            quantize_base=quantize_base,
        )
        if "v_proj" in lora_modules
        else nn.Linear(embed_dim, num_kv_heads * head_dim, bias=False)
    )
    output_proj = (
        LoRALinear(
            embed_dim,
            embed_dim,
            rank=lora_rank,
            alpha=lora_alpha,
            dropout=lora_dropout,
            quantize_base=quantize_base,
        )
        if "output_proj" in lora_modules
        else nn.Linear(embed_dim, embed_dim, bias=False)
    )
    rope = RotaryPositionalEmbeddings(dim=head_dim, max_seq_len=max_seq_len, base=rope_base)
    self_attn = CausalSelfAttention(
        embed_dim=embed_dim,
        num_heads=num_heads,
        num_kv_heads=num_kv_heads,
        head_dim=head_dim,
        q_proj=q_proj,
        k_proj=k_proj,
        v_proj=v_proj,
        output_proj=output_proj,
        pos_embeddings=rope,
        max_seq_len=max_seq_len,
        attn_dropout=attn_dropout,
    )
    return self_attn


def lora_mistral_mlp(
    *,
    dim: int,
    hidden_dim: int,
    lora_rank: int,
    lora_alpha: float,
    lora_dropout: float = 0.0,
    quantize_base: bool = False,
) -> FeedForward:
    gate_proj = LoRALinear(
        in_dim=dim,
        out_dim=hidden_dim,
        rank=lora_rank,
        alpha=lora_alpha,
        dropout=lora_dropout,
        quantize_base=quantize_base,
    )
    down_proj = LoRALinear(
        in_dim=hidden_dim,
        out_dim=dim,
        rank=lora_rank,
        alpha=lora_alpha,
        dropout=lora_dropout,
        quantize_base=quantize_base,
    )
    up_proj = LoRALinear(
        in_dim=dim,
        out_dim=hidden_dim,
        rank=lora_rank,
        alpha=lora_alpha,
        dropout=lora_dropout,
        quantize_base=quantize_base,
    )
    return FeedForward(
        gate_proj=gate_proj,
        down_proj=down_proj,
        up_proj=up_proj,
    )


def mistral_classifier(
    num_classes: int,
    *,
    # base mistral args
    vocab_size: int,
    num_layers: int,
    num_heads: int,
    num_kv_heads: int,
    embed_dim: int,
    intermediate_dim: int,
    max_seq_len: int,
    attn_dropout: float = 0.0,
    norm_eps: float = 1e-5,
    rope_base: int = 10_000,
) -> TransformerDecoder:
    """
    Build a base mistral model with an added classification layer.
    See :func:`~torchtune.models.mistral.mistral_classifier`
    for details on the base mistral classifier model.

    Args:
        num_classes (int): number of classes for the classification layer.
        vocab_size (int): number of tokens in vocabulary.
        num_layers (int): number of layers in the transformer decoder.
        num_heads (int): number of query heads. For MHA this is also the
            number of heads for key and value
        num_kv_heads (int): number of key and value heads. If specified,
            user should ensure `num_heads` % `num_kv_heads` == 0. Default value is
            `None`, in which case this is the same as MHA
        embed_dim (int): embedding dimension for self-attention
        intermediate_dim (int): intermediate dimension for MLP
        max_seq_len (int): maximum sequence length the model will be run with,
        attn_dropout (float): dropout value passed onto scaled_dot_product_attention.
            Default: 0.0
        norm_eps (float): epsilon in RMS norms
        rope_base (int): base for the rotary positional embeddings. Default: 10_000

    Returns:
        TransformerDecoder: Instantiation of mistral classification model.
    """
    head_dim = embed_dim // num_heads
    num_kv_heads = num_kv_heads if num_kv_heads else num_heads

    rope = RotaryPositionalEmbeddings(dim=head_dim, max_seq_len=max_seq_len, base=rope_base)
    self_attn = CausalSelfAttention(
        embed_dim=embed_dim,
        num_heads=num_heads,
        num_kv_heads=num_kv_heads,
        head_dim=head_dim,
        q_proj=nn.Linear(embed_dim, num_heads * head_dim, bias=False),
        k_proj=nn.Linear(embed_dim, num_kv_heads * head_dim, bias=False),
        v_proj=nn.Linear(embed_dim, num_kv_heads * head_dim, bias=False),
        output_proj=nn.Linear(embed_dim, embed_dim, bias=False),
        pos_embeddings=rope,
        kv_cache=None,
        max_seq_len=max_seq_len,
        attn_dropout=attn_dropout,
    )
    mlp = mistral_mlp(dim=embed_dim, hidden_dim=intermediate_dim)
    layer = TransformerDecoderLayer(
        attn=self_attn,
        mlp=mlp,
        sa_norm=RMSNorm(dim=embed_dim, eps=norm_eps),
        mlp_norm=RMSNorm(dim=embed_dim, eps=norm_eps),
    )
    tok_embeddings = nn.Embedding(vocab_size, embed_dim)
    output_proj = nn.Linear(embed_dim, num_classes, bias=False)
    return TransformerDecoder(
        tok_embeddings=tok_embeddings,
        layer=layer,
        num_layers=num_layers,
        max_seq_len=max_seq_len,
        num_heads=num_heads,
        head_dim=head_dim,
        norm=RMSNorm(embed_dim, eps=norm_eps),
        output=output_proj,
    )


def lora_mistral_classifier(
    lora_attn_modules: List[LORA_ATTN_MODULES],
    apply_lora_to_mlp: bool = False,
    apply_lora_to_output: bool = False,
    *,
    # mistral classifier args
    num_classes: int,
    # mistral args
    vocab_size: int,
    num_layers: int,
    num_heads: int,
    num_kv_heads: int,
    embed_dim: int,
    max_seq_len: int,
    intermediate_dim: int,
    attn_dropout: float = 0.0,
    norm_eps: float = 1e-5,
    rope_base: int = 10_000,
    # LoRA args
    lora_rank: int,
    lora_alpha: float,
    lora_dropout: float = 0.0,
    quantize_base: bool = False,
) -> TransformerDecoder:
    """
    Return a version of Mistral classifier (an instance of :func:`~torchtune.modules.TransformerDecoder`)
    with LoRA applied to some of the linear layers in its self-attention modules.

    Args:
        lora_attn_modules (List[LORA_ATTN_MODULES]): list of which linear layers
            LoRA should be applied to in each self-attention block. Options are
            ``{"q_proj", "k_proj", "v_proj", "output_proj"}``.
        apply_lora_to_mlp (bool): whether to apply LoRA to the MLP in each transformer layer.
            Default: False
        apply_lora_to_output (bool): whether to apply LoRA to the model's final output projection.
            Default: False
        vocab_size (int): number of tokens in vocabulary.
        num_layers (int): number of layers in the transformer decoder.
        num_heads (int): number of query heads. For MHA this is also the
            number of heads for key and value
        num_kv_heads (int): number of key and value heads. If specified,
            user should ensure `num_heads` % `num_kv_heads` == 0. Default value is
            `None`, in which case this is the same as MHA
        embed_dim (int): embedding dimension for self-attention
        max_seq_len (int): maximum sequence length the model will be run with
        intermediate_dim (int): intermediate dimension for MLP.
        attn_dropout (float): dropout value passed onto scaled_dot_product_attention.
            Default: 0.0
        norm_eps (float): epsilon in RMS norms.
        rope_base (int): base for the rotary positional embeddings. Default: 10_000
        lora_rank (int): rank of each low-rank approximation
        lora_alpha (float): scaling factor for the low-rank approximation
        lora_dropout (float): LoRA dropout probability. Default: 0.0
        quantize_base: (bool): Whether to quantize base model weights or not. Only applied to base
            weights within linear layers LoRA is applied to. The final output linear projection is not
            supported for quantization currently.

    Returns:
        TransformerDecoder: Instantiation of Mistral classifier model with LoRA applied to
        a subset of the attention projections in each layer.

    """

    self_attn = lora_mistral_self_attention(
        lora_modules=lora_attn_modules,
        embed_dim=embed_dim,
        num_heads=num_heads,
        num_kv_heads=num_kv_heads,
        max_seq_len=max_seq_len,
        attn_dropout=attn_dropout,
        rope_base=rope_base,
        lora_rank=lora_rank,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        quantize_base=quantize_base,
    )

    if apply_lora_to_mlp:
        mlp = lora_mistral_mlp(
            dim=embed_dim,
            hidden_dim=intermediate_dim,
            lora_rank=lora_rank,
            lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
            quantize_base=quantize_base,
        )
    else:
        mlp = mistral_mlp(dim=embed_dim, hidden_dim=intermediate_dim)

    layer = TransformerDecoderLayer(
        attn=self_attn,
        mlp=mlp,
        sa_norm=RMSNorm(dim=embed_dim, eps=norm_eps),
        mlp_norm=RMSNorm(dim=embed_dim, eps=norm_eps),
    )

    tok_embeddings = nn.Embedding(vocab_size, embed_dim)

    # TODO: quantize_base is not applied to final output_proj currently.
    output_proj = (
        LoRALinear(embed_dim, num_classes, rank=lora_rank, alpha=lora_alpha, dropout=lora_dropout)
        if apply_lora_to_output
        else nn.Linear(embed_dim, num_classes, bias=False)
    )
    model = TransformerDecoder(
        tok_embeddings=tok_embeddings,
        layer=layer,
        num_layers=num_layers,
        max_seq_len=max_seq_len,
        num_heads=num_heads,
        head_dim=(embed_dim // num_heads),
        norm=RMSNorm(embed_dim, eps=norm_eps),
        output=output_proj,
    )

    if quantize_base:
        # For QLoRA, we reparametrize 4-bit tensors to higher precision, and offload to CPU on the fly
        # so as to not increase peak memory
        model._register_state_dict_hook(
            partial(
                reparametrize_as_dtype_state_dict_post_hook,
                # TODO this is clowny, figure out a better way to get what precision the rest
                # of the model is in
                dtype=tok_embeddings.weight.dtype,
                offload_to_cpu=True,
            )
        )

    return model