modeling_llama_recovered.py

"""Custom loader for the compressed Llama-3.1 checkpoint.

A standard Llama with a reduced number of layers, plus two per-layer buffers
(`recover_scale`, `recover_bias`) applied at the start of each decoder layer's
forward. Layers carry scale=1, bias=0 where no correction is present.

Load with:  AutoModelForCausalLM.from_pretrained(path, trust_remote_code=True)
"""
import torch
import torch.nn as nn
from transformers.models.llama.configuration_llama import LlamaConfig
from transformers.models.llama.modeling_llama import (
    LlamaForCausalLM, LlamaModel, LlamaDecoderLayer)


class LlamaRecoveredConfig(LlamaConfig):
    model_type = "llama_recovered"


class RecoveredLlamaDecoderLayer(LlamaDecoderLayer):
    def __init__(self, config, layer_idx):
        super().__init__(config, layer_idx)
        h = config.hidden_size
        self.register_buffer("recover_scale", torch.ones(h), persistent=True)
        self.register_buffer("recover_bias", torch.zeros(h), persistent=True)

    def forward(self, hidden_states, *args, **kwargs):
        s = self.recover_scale.to(hidden_states.dtype)
        b = self.recover_bias.to(hidden_states.dtype)
        hidden_states = hidden_states * s + b
        return super().forward(hidden_states, *args, **kwargs)


class LlamaRecoveredModel(LlamaModel):
    config_class = LlamaRecoveredConfig
    _no_split_modules = ["RecoveredLlamaDecoderLayer"]

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


class LlamaRecoveredForCausalLM(LlamaForCausalLM):
    config_class = LlamaRecoveredConfig
    _no_split_modules = ["RecoveredLlamaDecoderLayer"]

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