modeling_q.py

from collections.abc import Callable

import torch
import torch.nn as nn
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
from transformers.models.mistral.modeling_mistral import (
    MistralAttention,
    MistralDecoderLayer,
    MistralForCausalLM,
    MistralMLP,
    MistralModel,
    MistralRMSNorm,
    apply_rotary_pos_emb,
    eager_attention_forward,
)

try:
    from .configuration_q import QConfig
except ImportError:  # запуск сгенерированного файла прямо из каталога ноутбука
    from configuration_q import QConfig


class QScalarGate(nn.Module):
    def __init__(self, hidden_size, multiplier=2.0):
        super().__init__()
        self.projection = nn.Linear(hidden_size, 1, bias=False)
        self.multiplier = multiplier

    def forward(self, branch, residual_input):
        return branch * self.multiplier * torch.sigmoid(self.projection(residual_input))


class QMLP(MistralMLP):
    def __init__(self, config):
        super().__init__(config)
        self.output_gate = (
            QScalarGate(config.hidden_size, config.gate_multiplier)
            if config.mlp_scalar_gate else None
        )

    def forward(self, hidden_states):
        output = super().forward(hidden_states)
        if self.output_gate is not None:
            output = self.output_gate(output, hidden_states)
        return output


class QAttention(MistralAttention):
    def __init__(self, config, layer_idx):
        super().__init__(config, layer_idx)
        self.use_rope = layer_idx not in config.nope_layers
        self.q_norm = MistralRMSNorm(self.head_dim, eps=config.rms_norm_eps) if config.qk_norm else None
        self.k_norm = MistralRMSNorm(self.head_dim, eps=config.rms_norm_eps) if config.qk_norm else None
        self.output_gate = (
            QScalarGate(config.hidden_size, config.gate_multiplier)
            if config.attention_scalar_gate else None
        )

    def forward(self, hidden_states, position_embeddings, attention_mask, past_key_values=None, **kwargs):
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)
        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        if self.q_norm is not None:
            query_states = self.q_norm(query_states)
            key_states = self.k_norm(key_states)
        if self.use_rope:
            cos, sin = position_embeddings
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_values is not None:
            key_states, value_states = past_key_values.update(
                key_states, value_states, self.layer_idx
            )

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )
        attn_output, attn_weights = attention_interface(
            self, query_states, key_states, value_states, attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            sliding_window=getattr(self.config, "sliding_window", None),
            **kwargs,
        )
        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        if self.output_gate is not None:
            attn_output = self.output_gate(attn_output, hidden_states)
        return attn_output, attn_weights


class QDecoderLayer(MistralDecoderLayer):
    def __init__(self, config, layer_idx):
        super().__init__(config, layer_idx)
        self.self_attn = QAttention(config, layer_idx)
        self.mlp = QMLP(config)


class QModel(MistralModel):
    config_class = QConfig

    def __init__(self, config):
        super().__init__(config)
        self.layers = nn.ModuleList(
            [QDecoderLayer(config, i) for i in range(config.num_hidden_layers)]
        )
        self.post_init()


class QForCausalLM(MistralForCausalLM):
    config_class = QConfig

    def __init__(self, config):
        super().__init__(config)
        self.model = QModel(config)
        self.post_init()