model.py

# model.py
"""
Full definition of a GPT Language Model, all of it in this single file.
References:
1) the official GPT-2 TensorFlow implementation released by OpenAI:
https://github.com/openai/gpt-2/blob/master/src/model.py
2) huggingface/transformers PyTorch implementation:
https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py
"""

import math
import inspect
import sys
from rich import print
import copy

import numpy as np

import torch
import torch.nn as nn
from torch.nn import functional as F

# Config
from gpt_conf import GPTConfig

# Checkpointing
import torch.utils.checkpoint as checkpoint

# Variations
from variations.attention_variations import attention_dictionary
from variations.mlp_variations import get_mlp_instance
from variations.moe_variations import MoELayer
from variations.lsv_variations import lsv_dictionary
from variations.softmax_variations import softmax_dictionary
from variations.norm_variations import norm_dictionary
from variations.position_encoding_variations import QuantizedEmbedding, RotaryEmbedding, SymmetricalOverlapAngularPositions, FIRE
from variations.absolute_position_variations import absolute_position_embedding_dict
from variations.activation_variations import activation_dictionary
from variations.linear_variations import linear_dictionary
from variations.router_variations import router_dictionary
from variations.output_vector_variants import output_vector_variant_dict
from variations.numerical_mapping_variations import get_numerical_embedding, get_numerical_output
from quantization.quantize import quantize_dictionary, dequantize, fake_quantize_act
from quantization.quant_utils import set_variant, create_activation_buffers

from initializations.initialization_variations import init_dictionary

from shared_param_utils import SharedParamGroupCreator
from variations.block_variations import Block
from variations.attention_residual_variations import FullAttentionResidual

class GPT(nn.Module):

    def __init__(self, config):
        super().__init__()
        assert config.vocab_size is not None
        assert config.block_size is not None

        self.config = config

        self.uses_numerical_multicontext = bool(config.numerical_multicontext)
        if self.uses_numerical_multicontext:
            if not config.multicontext:
                raise ValueError("numerical_multicontext requires multicontext mode")
            if config.n_embd_wte:
                raise ValueError("numerical_multicontext does not support factored embeddings")
            if not config.vocab_sizes:
                raise ValueError("numerical_multicontext requires vocab_sizes to be provided")

            self.numerical_embeddings = nn.ModuleDict()
            self.numerical_output_mlps = nn.ModuleDict()
            for idx in range(len(config.vocab_sizes)):
                key = str(idx)
                embedding_module = get_numerical_embedding(config)
                self.numerical_embeddings[key] = embedding_module
                self.numerical_output_mlps[key] = get_numerical_output(config, embedding_module=embedding_module)

        # Final-logit softcapping
        self.final_logit_softcapping = config.final_logit_softcapping

        # Optionally mix outputs of all blocks before final layer norm
        self.use_ln_f_input_mixer = config.use_ln_f_input_mixer
        if self.use_ln_f_input_mixer:
            variant_cls = output_vector_variant_dict[config.ln_f_input_mixer_variant]
            self.ln_f_mixer = variant_cls(config)

        # Use the new SharedParamGroupCreator for MLP and Attn layers
        spg_creator = SharedParamGroupCreator(config)
        shared_mlp_array = spg_creator.create_shared_param_group("mlp")
        shared_attn_array = spg_creator.create_shared_param_group("attn")

        # General weight tying
        self.wte_weight_tying = config.wte_weight_tying

        # Factorization Parameters
        self.n_embd_wte = config.n_embd_wte
        self.n_embd_wte_scale_tying = config.n_embd_wte_scale_tying

        # Embedding scale
        if config.use_embedding_scale:
            if config.embedding_scale_init is not None:
                self.embedding_scale = nn.Parameter(torch.tensor([config.embedding_scale_init]))
            else:
                self.embedding_scale = nn.Parameter(torch.sqrt(torch.tensor([config.n_embd])))

        # Learned Steering Vectors
        self.use_lsv = config.use_lsv
        self.lsv_index = config.lsv_index
        self.lsv_dataset_num = config.lsv_dataset_num

        if config.lsv_dataset_num is not None and config.use_lsv:
            self.num_datasets = config.lsv_dataset_num
            print(config.lsv_variant)
            self.lsv_variant = config.lsv_variant
            self.lsv_matrix = lsv_dictionary[self.lsv_variant](config)

        self.transformer = nn.ModuleDict(dict())
        # Configure wte, with optional quantization and factoring
        if config.quantize_wte:
            if config.n_embd_wte:
                # If factorization is set
                word_embd = QuantizedEmbedding(config.vocab_size, config.n_embd_wte, config.quantize_wte_method, config.quantize_wte_bits)
            else:
                # no factorization
                word_embd = QuantizedEmbedding(config.vocab_size, config.n_embd, config.quantize_wte_method, config.quantize_wte_bits)
            self.transformer['wte'] = word_embd
        else:
            if config.n_embd_wte:
                # If factorization is set
                word_embd = nn.Embedding(config.vocab_size, config.n_embd_wte)
                self.transformer['wte'] = word_embd
            else:
                #TODO: currently multicontext is in own category, add support later for WTE factorization
                if (config.multicontext or config.multidataset_wte) and not self.uses_numerical_multicontext:
                    for i, vocab_size in enumerate(self.config.vocab_sizes):
                        embedding_layer = nn.Embedding(vocab_size, config.n_embd)
                        self.transformer[f'wte_{i}'] = embedding_layer
                        self.transformer[f'lm_head_{i}'] = nn.Linear(config.n_embd, vocab_size, bias=False)
                else:
                    # no factorization
                    word_embd = nn.Embedding(config.vocab_size, config.n_embd)
                    self.transformer['wte'] = word_embd


        self.transformer['drop'] = nn.Dropout(config.dropout)
        self.transformer['h'] = nn.ModuleList([Block(config, mlp=shared_mlp_array[i], attn=shared_attn_array[i]) for i in range(config.n_layer)])
        self.attention_residual_variant = config.attention_residual_variant
        if self.attention_residual_variant == "full":
            self.attention_residual = FullAttentionResidual(
                2 * config.n_layer + 1, config.n_embd, config.attention_residual_eps
            )
        elif self.attention_residual_variant != "standard":
            raise ValueError(f"unknown attention_residual_variant: {self.attention_residual_variant}")
        self.transformer['ln_f'] = norm_dictionary[config.norm_variant_output](config)

        # Optional post-embedding normalizations
        if self.config.norm_variant_wte is not None:
            self.transformer['post_embedding_norm'] = self.build_norm_from_variant(config, "norm_variant_wte", "norm_wte")
        if self.config.norm_variant_abs is not None:
            self.transformer['post_abs_norm'] = self.build_norm_from_variant(config, "norm_variant_abs", "norm_abs")
        if self.config.norm_variant_lm_head is not None:
            self.transformer['lm_head_norm'] = self.build_norm_from_variant(config, "norm_variant_lm_head", "norm_lm_head")

        if self.config.use_abs_pos_embeddings:
            self.transformer['wpe'] = absolute_position_embedding_dict[config.absolute_pos_embedding_variant](config)

        # Select softmax variant for output layer
        self.softmax_variant_output = config.softmax_variant_output
        if self.softmax_variant_output != "softmax":
            self.softmax_layer_output = softmax_dictionary[config.softmax_variant_output](config)

        if config.n_embd_wte:
            self.lm_head = nn.Linear(config.n_embd_wte, config.vocab_size, bias=False)
        else:
            #TODO: currently multicontext is in own category, add support later for WTE factorization
            if (config.multicontext or config.multidataset_wte) and not self.uses_numerical_multicontext:
                for i, vocab_size in enumerate(self.config.vocab_sizes):
                    self.transformer[f'lm_head_{i}'].weight = self.transformer[f'wte_{i}'].weight
            else:
                self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)

        # Initialize and possibly import scale_up and scale_down matrices, if factorization is set
        if self.n_embd_wte:
            # TODO: make this linear set from variant dictionary
            # TODO: make this linear quantizable
            self.transformer['scale_up'] = nn.Linear(config.n_embd_wte, config.n_embd, bias=False)
            self.transformer['scale_down'] = nn.Linear(config.n_embd_wte, config.n_embd, bias=False)

            if self.n_embd_wte_scale_tying:
                self.transformer.scale_up.weight = self.transformer.scale_down.weight # Weight tying

            if config.import_scale_matrices_freeze:
                self.transformer.scale_up.weight.requires_grad = False
                self.transformer.scale_down.weight.requires_grad = False

        # init all weights
        self.apply(self._init_weights)

        # with weight tying when using torch.compile() some warnings get generated:
        # "UserWarning: functional_call was passed multiple values for tied weights.
        # This behavior is deprecated and will be an error in future versions"
        # not 100% sure what this is, so far seems to be harmless. TODO investigate
        if self.wte_weight_tying:
            if (config.multicontext or config.multidataset_wte) and not self.uses_numerical_multicontext:
                for i, vocab_size in enumerate(self.config.vocab_sizes):
                    self.transformer[f'lm_head_{i}'].weight = self.transformer[f'wte_{i}'].weight
            else:
                self.lm_head.weight = self.transformer.wte.weight # https://paperswithcode.com/method/weight-tying

        # import wte
        if self.config.import_wte_npy:
            # Replace wte with values from numpy and retie weights
            self.import_wte(self.config.import_wte_npy)

        # import scale_matrices
        if config.import_scale_matrices_npz:
            self.import_scale_matrices(config.import_scale_matrices_npz, config.n_embd_wte_scale_tying)

        for pn, p in self.named_parameters():
            # apply special scaled init to the residual projections, per GPT-2 paper
            if pn.endswith('c_proj.weight'):
                torch.nn.init.normal_(p, mean=0.0, std=0.02/math.sqrt(2 * config.n_layer))

        # report number of parameters
        print("number of parameters: %.2fM" % (self.get_num_params()/1e6,))

    def get_num_params(self, non_embedding=True):
        """
        Return the number of parameters in the model.
        For non-embedding count (default), the position embeddings get subtracted.
        The token embeddings would too, except due to the parameter sharing these
        params are actually used as weights in the final layer, so we include them.
        """
        n_params = sum(p.numel() for p in self.parameters())
        if non_embedding and self.config.use_abs_pos_embeddings:
            n_params -= sum(p.numel() for p in self.transformer.wpe.parameters())
        return n_params

    def update_block_size(self, new_block_size):
        # Function to increase block size dynamically
        if new_block_size > self.config.block_size:
            self.config.block_size = new_block_size
            if self.config.use_abs_pos_embeddings:
                self.transformer.wpe.update_block_size(new_block_size)
            for block in self.transformer.h:
                if hasattr(block.attn, 'bias'):
                    block.attn.bias = torch.tril(torch.ones(new_block_size, new_block_size)).view(1, 1, new_block_size, new_block_size)

    def build_norm_from_variant(self, config, variant_key: str, prefix: str):
        """Helper to deep-copy config and override hsnorm parameters if present."""
        norm_config = copy.deepcopy(config)
        for attr in ("radius", "scale", "gain", "radius_learning"):
            src = f"{prefix}_{attr}"
            if getattr(norm_config, src, None) is not None:
                setattr(norm_config, f"hsnorm_{attr}", getattr(norm_config, src))
        return norm_dictionary[getattr(config, variant_key)](norm_config)

    def apply_lm_head_norm(self, lm_head_weight):
        if self.config.norm_variant_lm_head is None:
            return lm_head_weight
        return self.transformer.lm_head_norm(lm_head_weight)

    def compute_lm_head_logits(self, x, lm_head_module):
        weight = self.apply_lm_head_norm(lm_head_module.weight)
        return F.linear(x, weight, lm_head_module.bias)

    def _forward_full_attention_residual(self, x, iter_num):
        """Run blocks while retaining each sublayer output as a depth source."""
        sources = [x]
        destination = 0
        for block in self.transformer.h:
            attn_input = self.attention_residual(sources, destination)
            sources.append(block.attention_residual_attn(attn_input, iter_num))
            destination += 1
            mlp_input = self.attention_residual(sources, destination)
            sources.append(block.attention_residual_mlp(mlp_input, iter_num))
            destination += 1
        return self.attention_residual(sources, destination)

    def _init_weights(self, module):
        """
        Custom weight initialization logic for GPT model.
        """
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=self.config.linear_mean_init, std=self.config.linear_std_init)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            if self.config.init_variant == "gaussian":
                torch.nn.init.normal_(
                    module.weight,
                    mean=self.config.embedding_mean_init,
                    std=self.config.embedding_std_init
                )
            elif 'wpe' in self.transformer.keys() and module is self.transformer['wpe']:
                torch.nn.init.normal_(
                    module.weight,
                    mean=self.config.embedding_mean_init,
                    std=self.config.embedding_std_init
                )
            else:
                init_fn = init_dictionary[self.config.init_variant]
                print(self.config.init_variant)

                # Generate custom init matrix
                weight_data = init_fn(self.config)

                # Copy into the module's weight
                with torch.no_grad():
                    if weight_data.shape != module.weight.shape:
                        raise ValueError(
                            f"Init shape {weight_data.shape} does not match embedding shape {module.weight.shape} "
                            f"for init_variant='{self.config.init_variant}'"
                        )
                    module.weight.copy_(weight_data)


    def update_num_angles(self, num_angles):
        """Update the number of angles for rotary embeddings in all attention layers."""
        device = next(self.parameters()).device
        for block in self.transformer.h:
            if hasattr(block.attn, 'rotary_emb_q') and hasattr(block.attn, 'rotary_emb_k'):
                block.attn.rotary_emb_q.update_num_angles(num_angles, device)
                block.attn.rotary_emb_k.update_num_angles(num_angles, device)

    def update_rope_length(self, rope_length):
        """Update the number of angles for rotary embeddings in all attention layers."""
        for block in self.transformer.h:
            if hasattr(block.attn, 'rotary_emb_q') and hasattr(block.attn, 'rotary_emb_k'):
                block.attn.rotary_emb_q.update_rope_length(rope_length)
                block.attn.rotary_emb_k.update_rope_length(rope_length)

    def import_wte(self, file_path):
        """ Replace wte with values from numpy and retie weights """

        #Load and format weights
        initial_embeddings = np.load(self.config.import_wte_npy)
        initial_embeddings_tensor = torch.from_numpy(initial_embeddings).float()

        # Initialize imported wte
        self.transformer.wte = nn.Embedding.from_pretrained(
                initial_embeddings_tensor,
                freeze=self.config.import_wte_freeze
                )

        # Redo the Weight tying
        self.lm_head.weight = self.transformer.wte.weight

    def export_wte(self, file_path):
        # TODO: Determine strategy with this and other means of export, possibly
        # replacing this with composition of existing means
        embedding_table = self.transformer.wte.weight.detach().cpu().numpy()
        np.save(file_path, embedding_table)
        print(f"Embedding table saved to {file_path}")

    def import_scale_matrices(self, file_path, weight_tying=False):
        """Import scale_up and scale_down matrices from a numpy file."""
        scale_matrices = np.load(file_path)
        scale_up_tensor = torch.from_numpy(scale_matrices['scale_up']).float().T
        scale_down_tensor = torch.from_numpy(scale_matrices['scale_down']).float().T

        print(scale_up_tensor.size())
        print(scale_down_tensor.size())
        self.transformer.scale_up.weight.data.copy_(scale_up_tensor)
        self.transformer.scale_down.weight.data.copy_(scale_down_tensor)

        if weight_tying:
            self.transformer.scale_up.weight = self.transformer.scale_down.weight

        print(f"Scale matrices loaded from {file_path} with weight tying: {weight_tying}")

    def export_scale_matrices(self, file_path):
        """Export scale_up and scale_down matrices to a numpy file."""
        scale_up_matrix = self.transformer.scale_up.weight.detach().cpu().numpy()
        scale_down_matrix = self.transformer.scale_down.weight.detach().cpu().numpy()

        np.savez(file_path, scale_up=scale_up_matrix, scale_down=scale_down_matrix)
        print(f"Scale matrices saved to {file_path}")

    def get_embedding_gaussian_noise_std(self, iter_num=None):
        if not self.config.embedding_gaussian_noise_in_eval and not self.training:
            return 0.0

        base_std = float(self.config.embedding_gaussian_noise_std or 0.0)
        start_std = self.config.embedding_gaussian_noise_start_std
        end_std = self.config.embedding_gaussian_noise_end_std
        start_std = base_std if start_std is None else float(start_std)
        end_std = base_std if end_std is None else float(end_std)

        end_iter = self.config.embedding_gaussian_noise_end_iter
        if end_iter is None or iter_num is None:
            return max(0.0, end_std)

        start_iter = int(self.config.embedding_gaussian_noise_start_iter or 0)
        end_iter = int(end_iter)
        if end_iter <= start_iter:
            return max(0.0, end_std)

        progress = (float(iter_num) - start_iter) / float(end_iter - start_iter)
        progress = min(max(progress, 0.0), 1.0)
        current_std = start_std + progress * (end_std - start_std)
        return max(0.0, current_std)

    def add_embedding_gaussian_noise(self, embeddings, iter_num=None):
        noise_std = self.get_embedding_gaussian_noise_std(iter_num=iter_num)
        if noise_std > 0:
            noise = torch.randn_like(embeddings)
            noise = noise / (noise.norm(dim=-1, keepdim=True) + 1e-6)
            noise = noise * noise_std
            embeddings = embeddings / (embeddings.norm(dim=-1, keepdim=True) + 1e-6)
            return embeddings + noise
        return embeddings

    def forward(self, idx, targets=None, iter_num=None, token_dict=None, target_dict=None, dataset_idx=None, loss_fn=None):
        if token_dict is not None:
            token_list = list(token_dict.values())
            # If target_dict is None (typical for inference), set target_list = None
            if target_dict is not None:
                target_list = list(target_dict.values())
            else:
                target_list = None
            device = token_list[0].device
            b, t = token_list[0].size()

            x = None

            # Add all of the input tokens
            for i, tokens in enumerate(token_list):
                if self.uses_numerical_multicontext:
                    module = self.numerical_embeddings[str(i)]
                    param = next(module.parameters())
                    if self.config.numerical_multicontext_input_format == "fp16_bits":
                        numeric_tokens = self._fp16bits_to_fp32(tokens).to(param.dtype)
                    else:
                        numeric_tokens = tokens.to(param.dtype)
                    token_repr = module(numeric_tokens.unsqueeze(-1))
                else:
                    token_repr = self.transformer[f'wte_{i}'](tokens)

                token_repr = self.add_embedding_gaussian_noise(token_repr, iter_num=iter_num)
                x = token_repr if x is None else x + token_repr

            if self.config.norm_variant_wte is not None:
                x = self.transformer.post_embedding_norm(x)

            if self.config.use_embedding_scale:
                x = x * self.embedding_scale

            if self.config.use_abs_pos_embeddings:
                pos_emb = self.transformer.wpe(t, device=device, training=self.training)  # (t, n_embd)
                x = self.transformer.drop(x + pos_emb)
            else:
                x = self.transformer.drop(x)

            x.requires_grad_(True)

            # 2. Possibly apply LSV on input
            if self.use_lsv and self.config.apply_lsv_at_layer_idx == 0:
                x = self.lsv_matrix(x)

            if self.use_ln_f_input_mixer:
                layer_outputs = [x]

            layer_idx = 1
            blocks = self.transformer.h
            if self.attention_residual_variant == "full":
                x = self._forward_full_attention_residual(x, iter_num)
                blocks = ()
            for block in blocks:
                x = block(x, iter_num)

                # Steering logic
                if self.use_lsv and layer_idx == self.config.apply_lsv_at_layer_idx:
                    x = self.lsv_matrix(x)
                if (self.config.apply_vector_at_layer_idx is not None
                        and layer_idx == self.config.apply_vector_at_layer_idx):
                    x = self.apply_vector_to_layer_output(x)
                if (self.config.obtain_vector_at_layer_idx is not None
                        and layer_idx == self.config.obtain_vector_at_layer_idx):
                    x = self.obtain_vector_from_layer_output(x)

                if self.use_ln_f_input_mixer:
                    layer_outputs.append(x)

                layer_idx += 1

            if self.use_ln_f_input_mixer:
                x = self.ln_f_mixer(layer_outputs)

            # 3. Final layer norm
            x = self.transformer.ln_f(x)

            # 4. Optionally scale down
            if self.n_embd_wte:
                x = F.linear(x, self.transformer.scale_down.weight.t())

            # 5. Compute separate logits
            if self.uses_numerical_multicontext:
                logits = [self.numerical_output_mlps[str(i)](x) for i in range(len(token_list))]

                if target_list is not None:
                    losses = []
                    for i, preds in enumerate(logits):
                        if self.config.numerical_multicontext_input_format == "fp16_bits":
                            decoded_targets = self._fp16bits_to_fp32(target_list[i])
                            targets = decoded_targets.to(preds.dtype)
                        else:
                            targets = target_list[i].to(preds.dtype)
                        mask = target_list[i] != -1

                        if mask.any():
                            pred_sel = preds[mask]
                            target_sel = targets.unsqueeze(-1)[mask]

                            huber = F.huber_loss(
                                pred_sel,
                                target_sel,
                                delta=self.config.numerical_loss_huber_delta,
                                reduction="mean",
                            )

                            if self.config.numerical_loss_use_cosine:
                                cos = 1.0 - F.cosine_similarity(
                                    pred_sel,
                                    target_sel,
                                    dim=-1,
                                    eps=1e-8,
                                ).mean()
                                loss_i = huber + (self.config.numerical_loss_cosine_coeff * cos)
                            else:
                                loss_i = huber
                        else:
                            loss_i = torch.zeros((), device=preds.device, dtype=preds.dtype)
                        losses.append(loss_i)
                else:
                    logits = [pred[:, [-1], :] for pred in logits]
                    losses = None
            else:
                logits = [self.compute_lm_head_logits(x, self.transformer[f'lm_head_{i}']) for i in range(len(token_list))]

                # Soft‑cap **each** logits tensor (training & inference)
                if self.config.final_logit_softcapping is not None:
                    logits = [
                        torch.tanh(logit_var / self.config.final_logit_softcapping) *
                        self.config.final_logit_softcapping
                        for logit_var in logits
                    ]

                # 6. Compute losses if targets are provided
                # If we only want the last token, adapt the slices as you prefer
                if target_list is not None:
                    # If we do want to compute losses for each context
                    losses = []
                    for i in range(len(token_list)):
                        if loss_fn is None:
                            loss_i = F.cross_entropy(
                                logits[i].view(-1, logits[i].size(-1)),
                                target_list[i].view(-1),
                                ignore_index=-1
                            )
                        else:
                            loss_i = loss_fn(logits[i], target_list[i], iter_num=iter_num)
                        losses.append(loss_i)

                else:
                    # only forward lm head on very last position in inference mode
                    logits = [logit[:, [-1], :] for logit in logits]
                    losses = None

            return logits, losses

        else:
            device = idx.device
            b, t = idx.size()
            # assert t <= self.config.block_size, f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"

            # forward the GPT model itself
            if self.config.multidataset_wte and dataset_idx is not None:
                tok_emb = self.transformer[f'wte_{dataset_idx}'](idx)
            else:
                tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
            x = None

            tok_emb = self.add_embedding_gaussian_noise(tok_emb, iter_num=iter_num)
            if self.n_embd_wte:
                tok_emb = self.transformer.scale_up(tok_emb)

            if self.config.use_embedding_scale:
                tok_emb = tok_emb * self.embedding_scale

            if self.config.norm_variant_wte is not None:
                tok_emb = self.transformer.post_embedding_norm(tok_emb)

            if self.config.use_abs_pos_embeddings:
                pos_emb = self.transformer.wpe(t, device=device, training=self.training) # position embeddings of shape (t, n_embd)
                x = tok_emb + pos_emb
                if self.config.norm_variant_abs is not None:
                    x = self.transformer.post_abs_norm(x)
                x = self.transformer.drop(x)
            else:
                x = self.transformer.drop(tok_emb)

            x.requires_grad_(True)  # Ensure requires_grad is True

            if self.use_lsv and self.config.apply_lsv_at_layer_idx == 0:
                x = self.lsv_matrix(x)

            if self.use_ln_f_input_mixer:
                layer_outputs = [x]

            layer_idx = 1
            blocks = self.transformer.h
            if self.attention_residual_variant == "full":
                x = self._forward_full_attention_residual(x, iter_num)
                blocks = ()
            for block in blocks:
                # Propagate tokens through layers
                x = block(x, iter_num)

                # Intercept for Learned Steering Vectors
                if self.use_lsv and layer_idx == self.config.apply_lsv_at_layer_idx:
                    x = self.lsv_matrix(x)
                    # x = self.apply_learned_vector_to_layer_output(x)

                # Intercept for Steering Vectors
                if self.config.apply_vector_at_layer_idx is not None and layer_idx == self.config.apply_vector_at_layer_idx:
                    x = self.apply_vector_to_layer_output(x)
                if self.config.obtain_vector_at_layer_idx is not None and layer_idx == self.config.obtain_vector_at_layer_idx:
                    print(layer_idx, self.config.obtain_vector_at_layer_idx)
                    x = self.obtain_vector_from_layer_output(x)

                if self.use_ln_f_input_mixer:
                    layer_outputs.append(x)

                layer_idx +=1

            if self.use_ln_f_input_mixer:
                x = self.ln_f_mixer(layer_outputs)

            x = self.transformer.ln_f(x)

            if self.n_embd_wte:
                x = F.linear(x, self.transformer.scale_down.weight.t())


            if targets is not None:
                # if we are given some desired targets also calculate the loss
                if self.config.multidataset_wte and dataset_idx is not None:
                    logits = self.compute_lm_head_logits(x, self.transformer[f'lm_head_{dataset_idx}'])
                else:
                    logits = self.compute_lm_head_logits(x, self.lm_head)

                if self.config.final_logit_softcapping is not None:
                    logits = logits / self.config.final_logit_softcapping
                    logits = torch.tanh(logits)
                    logits = logits * self.config.final_logit_softcapping

                if loss_fn is None:
                    loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
                else:
                    loss = loss_fn(logits, targets, iter_num=iter_num)
            else:
                # inference-time mini-optimization: only forward the lm_head on the very last position
                if self.config.multidataset_wte and dataset_idx is not None:
                    logits = self.compute_lm_head_logits(x[:, [-1], :], self.transformer[f'lm_head_{dataset_idx}'])
                else:
                    logits = self.compute_lm_head_logits(x[:, [-1], :], self.lm_head) # note: using list [-1] to preserve the time dim

                if self.config.final_logit_softcapping is not None:
                    logits = logits / self.config.final_logit_softcapping
                    logits = torch.tanh(logits)
                    logits = logits * self.config.final_logit_softcapping

                loss = None

            return logits, loss
    # ------------------------------------------------------------------
    #  LATENT-CHAINING
    # ------------------------------------------------------------------
    @torch.no_grad()
    def embed_tokens(self, idx, dataset_idx=None):
        """
        Return the (B,T,E) tensor right *after* token embeddings,
        factor-scale-up, positional embedding and dropout.  Exactly the
        same tensor that flows into the first transformer Block inside
        `forward()`.  Used by train_recurrent.py for the FIRST step.
        """
        device = idx.device
        if self.config.multidataset_wte and dataset_idx is not None:
            tok_emb = self.transformer[f'wte_{dataset_idx}'](idx)
        else:
            tok_emb = self.transformer.wte(idx)

        tok_emb = self.add_embedding_gaussian_noise(tok_emb, iter_num=None)

        if self.n_embd_wte:
            tok_emb = self.transformer.scale_up(tok_emb)

        if self.config.use_embedding_scale:
            tok_emb = tok_emb * self.embedding_scale

        if self.config.norm_variant_wte is not None:
            tok_emb = self.transformer.post_embedding_norm(tok_emb)

        if self.config.use_abs_pos_embeddings:
            t = idx.size(1)
            tok_emb = tok_emb + self.transformer.wpe(t, device=device, training=self.training)
            if self.config.norm_variant_abs is not None:
                tok_emb = self.transformer.post_abs_norm(tok_emb)


        return self.transformer.drop(tok_emb)

    def forward_embedded(self, x_emb, iter_num=None, return_hidden=False, dataset_idx=None):
        """
        Complete forward pass **starting from an already-embedded tensor**
        `x_emb` of shape (B,T,E).  Returns (`logits`, `loss`) identical to
        `forward`, and – if `return_hidden` – also the final hidden state
        right before `lm_head`.  No gradients are blocked; loss still
        back-propagates into `x_emb`.
        """
        # ---- copy–paste from the “else:” branch of forward() ---------
        b, t, _ = x_emb.size()
        x = x_emb

        # (steering vectors, etc.)
        if self.use_lsv and self.config.apply_lsv_at_layer_idx == 0:
            x = self.lsv_matrix(x)

        if self.use_ln_f_input_mixer:
            layer_outputs = [x]

        layer_idx = 1
        for block in self.transformer.h:
            x = block(x, iter_num)
            if self.use_lsv and layer_idx == self.config.apply_lsv_at_layer_idx:
                x = self.lsv_matrix(x)
            if self.use_ln_f_input_mixer:
                layer_outputs.append(x)
            layer_idx += 1

        if self.use_ln_f_input_mixer:
            x = self.ln_f_mixer(layer_outputs)

        x = self.transformer.ln_f(x)
        if self.n_embd_wte:
            x = F.linear(x, self.transformer.scale_down.weight.t())

        if self.config.multidataset_wte and dataset_idx is not None:
            logits = self.compute_lm_head_logits(x, self.transformer[f'lm_head_{dataset_idx}'])
        else:
            logits = self.compute_lm_head_logits(x, self.lm_head)
        if self.final_logit_softcapping is not None:
            logits = torch.tanh(logits / self.final_logit_softcapping) \
                     * self.final_logit_softcapping

        return (logits, x) if return_hidden else (logits, None)

    def set_lsv_scaling_factor(self, factor):
        self.lsv_matrix.update_lsv_scaling_factor(factor)

    def set_lsv_mode(self, mode):
        self.lsv_matrix.set_mode(mode)

    def set_lsv_mixture(self, mixture):
        """ Mixture is a list, allowing for mixing steering vectors """
        self.lsv_matrix.set_mixture(mixture)

    def get_lsv_scaling_factor(self):
        return self.lsv_matrix.get_lsv_scaling_factor()

    def set_lsv_index(self, index):
        self.lsv_matrix.update_lsv_index(index)

    def freeze_non_lsv_parameters(self):
        """Freeze all parameters except for lsv_matrix if lsv_focused_training is enabled."""

        print("Freezing all parameters except for lsv_matrix")

        # Freeze all parameters by setting requires_grad to False
        for name, param in self.named_parameters():
            if name != "lsv_matrix":
                param.requires_grad = False
            else:
                param.requires_grad = True  # Ensure lsv_matrix can still be trained

    def apply_learned_vector_to_layer_output(self, x):
        """Conditionally add a vector based on dataset index to the output of a specific layer."""

        # Use one-hot vector for the dataset and multiply by the learned parameter matrix
        one_hot_vector = torch.zeros(self.lsv_matrix.size(0), device=x.device)
        one_hot_vector[self.lsv_index] = 1.0

        # Multiply the one-hot vector by the learned parameter matrix
        selected_vector = torch.matmul(one_hot_vector, self.lsv_matrix)

        x = x + selected_vector

        return x

    def apply_vector_to_layer_output(self, x):
        """Conditionally add a vector from a file to the output of a specific layer."""

        # require this method has the vector file
        assert self.config.apply_vector_file is not None

        vector = np.load(self.config.apply_vector_file)
        vector_tensor = torch.from_numpy(vector).float().to(x.device)
        x = x + self.config.apply_vector_scaling_factor * vector_tensor

        return x

    def obtain_vector_from_layer_output(self, x):
        """Append a vector to an existing .npy file."""

        # Convert the tensor back to a numpy array
        y = x
        y = torch.mean(y, dim=1, keepdim=True)
        result_vector = y.detach().cpu().numpy()

        # Save the vector to file
        np.save(self.config.obtain_vector_file, result_vector)
        print(f"Updated avg vector saved to {self.config.obtain_vector_file}")

    @staticmethod
    def _fp16bits_to_fp32(bits: torch.Tensor) -> torch.Tensor:
        """Convert IEEE-754 half-precision bit patterns to float32 tensors."""

        b = bits.to(torch.int32)

        sign = (b >> 15) & 0x1
        exponent = (b >> 10) & 0x1F
        mantissa = b & 0x3FF

        sign_f = torch.where(sign == 0, 1.0, -1.0).to(torch.float32)
        exponent_i = exponent.to(torch.int32)
        mantissa_f = mantissa.to(torch.float32)

        # Subnormal: sign * 2^-14 * (mantissa / 2^10)
        subnormal_mag = torch.ldexp(mantissa_f / 1024.0, torch.full_like(exponent_i, -14))

        # Normal: sign * 2^(exp-15) * (1 + mantissa / 2^10)
        normal_mag = torch.ldexp(1.0 + (mantissa_f / 1024.0), exponent_i - 15)

        is_zero = (exponent_i == 0) & (mantissa == 0)
        is_subnormal = (exponent_i == 0) & (mantissa != 0)
        is_normal = (exponent_i > 0) & (exponent_i < 0x1F)
        is_special = exponent_i == 0x1F
        is_inf = is_special & (mantissa == 0)
        is_nan = is_special & (mantissa != 0)

        out = torch.zeros_like(sign_f, dtype=torch.float32)
        out = torch.where(is_subnormal, sign_f * subnormal_mag, out)
        out = torch.where(is_normal, sign_f * normal_mag, out)
        out = torch.where(is_inf, sign_f * float("inf"), out)
        out = torch.where(is_nan, float("nan"), out)
        out = torch.where(is_zero, torch.zeros_like(out), out)

        return out

    def crop_block_size(self, block_size):
        # model surgery to decrease the block size if necessary
        # e.g. we may load the GPT2 pretrained model checkpoint (block size 1024)
        # but want to use a smaller block size for some smaller, simpler model
        assert block_size <= self.config.block_size
        self.config.block_size = block_size
        if self.config.use_abs_pos_embeddings:
            self.transformer.wpe.crop_block_size(block_size)
        for block in self.transformer.h:
            if hasattr(block.attn, 'bias'):
                block.attn.bias = block.attn.bias[:,:,:block_size,:block_size]

    @classmethod
    def from_pretrained(cls, config, model_type):
        # assert model_type in {'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'}
        from transformers import GPT2LMHeadModel

        print(f"loading weights from pretrained gpt: {model_type}")

        # create a from-scratch initialized minGPT model
        model = GPT(config)
        sd = model.state_dict()
        sd_keys = sd.keys()
        sd_keys = [k for k in sd_keys if not k.endswith('.attn.bias')] # discard this mask / buffer, not a param

        # init a huggingface/transformers model
        model_hf = GPT2LMHeadModel.from_pretrained(model_type)
        sd_hf = model_hf.state_dict()

        # copy while ensuring all of the parameters are aligned and match in names and shapes
        sd_keys_hf = sd_hf.keys()
        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.masked_bias')] # ignore these, just a buffer
        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.bias')] # same, just the mask (buffer)
        transposed = ['attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight']

        # assert len(sd_keys_hf) == len(sd_keys), f"mismatched keys: {len(sd_keys_hf)} != {len(sd_keys)}"
        for key in sd_keys_hf:
            # START FIX: Rename keys to match nanoGPT's convention
            my_key = key
            if 'ln_1' in my_key:
                my_key = my_key.replace('ln_1', 'ln1')
            if 'ln_2' in my_key:
                my_key = my_key.replace('ln_2', 'ln2')
            # END FIX

            if any(key.endswith(w) for w in transposed):
                # special treatment for the Conv1D weights we need to transpose
                assert sd_hf[key].shape[::-1] == sd[my_key].shape
                with torch.no_grad():
                    sd[my_key].copy_(sd_hf[key].t())
            elif key.endswith('attn.c_attn.weight') or key.endswith('attn.c_attn.bias'):
                # split into c_attn_q/k/v
                q, k, v  = sd_hf[key].split(config.n_embd, dim=-1) # Note: HF stores as (3 * n_embd, n_embd) for weights

                # Adjust for bias shape if it exists
                if key.endswith('.bias'):
                    q_key_str = my_key.replace("c_attn", "c_attn_q")
                    k_key_str = my_key.replace("c_attn", "c_attn_k")
                    v_key_str = my_key.replace("c_attn", "c_attn_v")
                    sd[q_key_str].copy_(q)
                    sd[k_key_str].copy_(k)
                    sd[v_key_str].copy_(v)
                else: # it's a weight
                    q, k, v = q.t(), k.t(), v.t() # Transpose weights
                    q_key_str = my_key.replace("c_attn", "c_attn_q")
                    k_key_str = my_key.replace("c_attn", "c_attn_k")
                    v_key_str = my_key.replace("c_attn", "c_attn_v")
                    sd[q_key_str].copy_(q)
                    sd[k_key_str].copy_(k)
                    sd[v_key_str].copy_(v)
            else:
                # vanilla copy over the other parameters
                if config.n_embd_wte:
                    if "wte" in key or "lm_head" in key:
                        continue
                if not config.use_abs_pos_embeddings:
                    if "wpe" in key:
                        continue

                # Ensure the key exists in your model before trying to copy
                if my_key in sd:
                    assert sd_hf[key].shape == sd[my_key].shape, f"Shape mismatch for key {my_key}: HF is {sd_hf[key].shape}, yours is {sd[my_key].shape}"
                    with torch.no_grad():
                        sd[my_key].copy_(sd_hf[key])

        return model

    def configure_optimizers(self, weight_decay, learning_rate, betas, device_type):
        # start with all of the candidate parameters
        param_dict = {pn: p for pn, p in self.named_parameters()}
        # filter out those that do not require grad
        param_dict = {pn: p for pn, p in param_dict.items() if p.requires_grad}
        # create optim groups. Any parameters that is 2D will be weight decayed, otherwise no.
        # i.e. all weight tensors in matmuls + embeddings decay, all biases and layernorms don't.
        decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]
        nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]
        optim_groups = [
            {'params': decay_params, 'weight_decay': weight_decay},
            {'params': nodecay_params, 'weight_decay': 0.0}
        ]
        num_decay_params = sum(p.numel() for p in decay_params)
        num_nodecay_params = sum(p.numel() for p in nodecay_params)
        print(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters")
        print(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters")
        # Create AdamW optimizer and use the fused version if it is available
        fused_available = 'fused' in inspect.signature(torch.optim.AdamW).parameters
        use_fused = fused_available and device_type == 'cuda'
        extra_args = dict(fused=True) if use_fused else dict()
        optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas, **extra_args)
        print(f"using fused AdamW: {use_fused}")

        return optimizer

    def estimate_mfu(self, fwdbwd_per_iter, dt):
        """ estimate model flops utilization (MFU) in units of A100 bfloat16 peak FLOPS """
        # first estimate the number of flops we do per iteration.
        # see PaLM paper Appendix B as ref: https://arxiv.org/abs/2204.02311
        N = self.get_num_params()
        cfg = self.config
        L, H, Q, T = cfg.n_layer, cfg.n_head, cfg.n_embd//cfg.n_head, cfg.block_size
        flops_per_token = 6*N + 12*L*H*Q*T
        flops_per_fwdbwd = flops_per_token * T
        flops_per_iter = flops_per_fwdbwd * fwdbwd_per_iter
        # express our flops throughput as ratio of A100 bfloat16 peak flops
        flops_achieved = flops_per_iter * (1.0/dt) # per second
        flops_promised = 312e12 # A100 GPU bfloat16 peak flops is 312 TFLOPS
        mfu = flops_achieved / flops_promised
        return mfu

    @torch.no_grad()
    def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None):
        """
        Take a conditioning sequence of indices idx (LongTensor of shape (b,t)) and complete
        the sequence max_new_tokens times, feeding the predictions back into the model each time.
        Most likely you'll want to make sure to be in model.eval() mode of operation for this.
        """
        for _ in range(max_new_tokens):
            # if the sequence context is growing too long we must crop it at block_size
            idx_cond = idx if idx.size(1) <= self.config.block_size else idx[:, -self.config.block_size:]
            # forward the model to get the logits for the index in the sequence
            logits, _ = self(idx_cond)
            # pluck the logits at the final step and scale by desired temperature
            logits = logits[:, -1, :] / temperature
            # optionally crop the logits to only the top k options
            if top_k is not None:
                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                logits[logits < v[:, [-1]]] = -float('Inf')

            probs = None
            if self.config.softmax_variant_output != 'softmax':
                probs = self.softmax_layer_output(logits)
            else:
                probs = F.softmax(logits, dim=-1)
            assert probs != None
            idx_next = torch.multinomial(probs, num_samples=1)
            # append sampled index to the running sequence and continue
            idx = torch.cat((idx, idx_next), dim=1)

        return idx

    @torch.no_grad()
    def generate_with_stop(self, idx, max_new_tokens, stop_strings, decode, temperature=1.0, top_k=None):
        """
        Generate tokens and stop on any fixed string match from a list of stop strings.
        """
        if isinstance(stop_strings, str):
            stop_strings = [stop_strings]  # make it a list if a single string

        generated_text = ""
        buffer = ""
        for _ in range(max_new_tokens):
            idx_cond = idx if idx.size(1) <= self.config.block_size else idx[:, -self.config.block_size:]
            logits, _ = self(idx_cond)
            logits = logits[:, -1, :] / temperature
            if top_k is not None:
                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                logits[logits < v[:, [-1]]] = -float('Inf')
            probs = F.softmax(logits, dim=-1)
            idx_next = torch.multinomial(probs, num_samples=1)
            idx = torch.cat((idx, idx_next), dim=1)

            next_token_text = decode(idx_next[0].tolist())
            generated_text += next_token_text
            buffer += next_token_text

            # Check if buffer ends with any stop string
            for stop_string in stop_strings:
                if buffer.endswith(stop_string):
                    return idx, generated_text

        return idx, generated_text