src/run_t5.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Fine-tuning the library models for sequence to sequence.
"""
# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.

import logging
import os
import sys
import json
import time
from dataclasses import dataclass, field
from typing import Optional
import math
import torch
_original_load = torch.load
def _patched_load(*args, **kwargs):
    kwargs.setdefault("weights_only", False)
    return _original_load(*args, **kwargs)
torch.load = _patched_load
from torch import nn
try:
    import ipdb  # Optional debug dependency.
except ImportError:
    ipdb = None

import datasets
import nltk  # Here to have a nice missing dependency error message early on
import numpy as np
import pickle
from datasets import load_dataset
from copy import deepcopy

import transformers
from transformers import (
    AutoConfig,
    AutoTokenizer,
    HfArgumentParser,
    Seq2SeqTrainingArguments,
    set_seed, )

from transformers.trainer_utils import get_last_checkpoint

from cl_collator import DataCollator
from cl_dataset import gen_cache_path

from assets import task_config, lora_state_dict_A, lora_state_dict_B

from cl_trainer_gainlora import DenserEvalCallback, skip_instructions
from compute_metrics import compute_metrics, compute_grouped_metrics
from datasets.download import DownloadConfig

# off wandb
os.environ['WANDB_DISABLED'] = "True"
# os.environ['CUDA_VISIBLE_DEVICES'] = '0'
logger = logging.getLogger(__name__)
CURRENT_DIR = os.path.dirname(__file__)

local_data_path = "/home/work/nltk_data"
nltk.data.path.append(local_data_path)

@dataclass(frozen=False)
class ModelArguments:
    """
    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
    """

    model_name_or_path: str = field(
        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
    )
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
    )
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
    )
    cache_dir: Optional[str] = field(
        default=None,
        metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
    )
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
    )
    model_revision: str = field(
        default="main",
        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
    )
    use_auth_token: bool = field(
        default=False,
        metadata={
            "help": "Will use the token generated when running `transformers-cli login` (necessary to use this script "
                    "with private models)."
        },
    )
    resize_position_embeddings: Optional[bool] = field(
        default=None,
        metadata={
            "help": "Whether to automatically resize the position embeddings if `max_source_length` exceeds "
                    "the model's position embeddings."
        },
    )
    # added for AutoCL
    lora_dim: Optional[int] = field(
        default=8,
        metadata={
            "help": "Intrinsic dimension of the latent space."
        },
    )

    prefix_len: Optional[int] = field(
        default=10,
        metadata={
            "help": "Length of Prompt."
        },
    )

    mlp_hidden_dim: Optional[int] = field(
        default=100,
        metadata={
            "help": "Intrinsic dimension of the latent MLP space."
        },
    )

    attn_temperature: Optional[int] = field(
        default=1,
        metadata={
            "help": "Temperature to control attention weights."
        },
    )
    lora_r: Optional[int] = field(
        default=8,
        metadata={
            "help": "Temperature to control attention weights."
        },
    )
    lora_alpha: Optional[int] = field(
        default=1,
        metadata={
            "help": "Temperature to control attention weights."
        },
    )
    lora_dropout: Optional[float] = field(
        default=0.,
        metadata={
            "help": "Temperature to control attention weights."
        },
    )

    run_single: bool = field(
        default=False,
        metadata={
            "help": "Temperature to control attention weights."
        },
    )

    previous_lora_path: Optional[str] = field(
        default=None,
        metadata={"help": "the path to load previous prompts."}
    )

    current_lora_path: Optional[str] = field(
        default=None,
        metadata={"help": "Path to saved_weights/ of current task for eval-only resume. "
                          "Loads lora_weights_A.pt and lora_weights_B.pt into the current task's lora_A/B."}
    )

    previous_prompt_key_path: Optional[str] = field(
        default=None,
        metadata={"help": "the path to load previous prompts."}
    )

    load_checkpoint_from: str = field(
        default=None,
        metadata={"help": "Path to load previous checkpoints"}
    )


@dataclass(frozen=False)
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.
    """
    lang: str = field(default=None, metadata={"help": "Language id for multilingual model."})
    data_dir: str = field(
        default=None, metadata={"help": "The directory for saving the CL train/dev/test splits."}
    )
    gen_data_dir: str = field(
        default=None, metadata={"help": "The directory for saving the generated train/dev/test splits."}
    )
    task_order: str = field(
        default=None, metadata={"help": "order of the tasks"}
    )
    task_config_dir: str = field(
        default=None, metadata={"help": "The json file for config training and testing tasks"}
    )
    replay_task_list: Optional[str] = field(
        default='', metadata={
            "help": "Different tasks to replay"
        }
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
    input_record_file: str = field(
        default=None, metadata={"help": "file to record model input"}
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    max_source_length: Optional[int] = field(
        default=512,
        metadata={
            "help": "The maximum total input sequence length after tokenization. Sequences longer "
                    "than this will be truncated, sequences shorter will be padded."
        },
    )
    # for decoder model, it means max_new_tokens
    max_target_length: Optional[int] = field(
        default=50,
        metadata={
            "help": "The maximum total sequence length for target text after tokenization. Sequences longer "
                    "than this will be truncated, sequences shorter will be padded."
        },
    )
    repetition_penalty: Optional[float] = field(
        default=1.0,
        metadata={
            "help": "Penalty for repeat tokens in decode stage."
        },
    )
    num_beams: Optional[int] = field(
        default=1,
        metadata={
            "help": "Number of beams to use for evaluation. This argument will be passed to ``model.generate``, "
                    "which is used during ``evaluate`` and ``predict``."
        },
    )
    max_num_instances_per_task: int = field(
        default=10000, metadata={"help": "The maximum number of instances we will consider for each training task."}
    )
    max_num_instances_per_eval_task: int = field(
        default=200,
        metadata={"help": "The maximum number of instances we will consider for each validation/test task."}
    )
    max_train_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
                    "value if set."
        },
    )
    max_eval_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
                    "value if set."
        },
    )
    max_predict_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of prediction examples to this "
                    "value if set."
        },
    )
    num_examples: Optional[int] = field(
        default=0,
        metadata={"help": "number of in-context positive examples."}
    )
    ignore_pad_token_for_loss: bool = field(
        default=True,
        metadata={
            "help": "Whether to ignore the tokens corresponding to padded labels in the loss computation or not."
        },
    )
    add_task_name: Optional[bool] = field(
        default=False,
        metadata={"help": "whether to preappend task name before the task input."}
    )
    add_dataset_name: Optional[bool] = field(
        default=False,
        metadata={"help": "whether to preappend dataset name before the task input."}
    )
    add_instruction_replay: Optional[bool] = field(
        default=True,
        metadata={"help": "whether to preappend definition and few-shot cases before the task input during replay."}
    )

@dataclass(frozen=False)
class TrainingArguments(Seq2SeqTrainingArguments):
    gradient_checkpointing: Optional[bool] = field(
        default=False,
        metadata={"help": "Whether to use computing time to gain more memory"}
    )

    # ── SRT (Statistical Routing Theory) ──────────────────────────────
    use_srt_router: Optional[bool] = field(
        default=True,
        metadata={"help": "Enable SRT router: non-parametric routing via {μ_t, Σ_t} signatures."},
    )
    srt_metric_mode: Optional[str] = field(
        default='hard',
        metadata={
            "help": "SRT routing mode: "
                    "'hard'     = ZCA whitening + L2 (matches routing_analysis experiment), "
                    "'dynamics' = SRM metric selection (matches contribution_UNIFIED)."
        },
    )
    srt_shrink: Optional[bool] = field(
        default=False,   # FALSE for hard mode: raw covariance matches routing_analysis experiment
        metadata={"help": "Apply Ledoit-Wolf shrinkage to covariance estimation."},
    )
    srt_shrink_factor: Optional[float] = field(
        default=0.1,
        metadata={"help": "Ledoit-Wolf shrinkage intensity."},
    )
    srt_max_emb_samples: Optional[int] = field(
        default=500,
        metadata={"help": "Max training batches for embedding extraction."},
    )
    srt_load_path: Optional[str] = field(
        default=None,
        metadata={"help": "Path to load SRT signatures from a previous checkpoint (multi-task CL)."},
    )
    srt_skip_forward: Optional[bool] = field(
        default=False,
        metadata={
            "help": "Skip forward-pass embedding extraction. "
                    "Load pre-extracted embeddings from embeddings/{backbone}/{split}/{task}/train.npz instead. "
                    "Requires pre-extracted embeddings to exist."
        },
    )

    # ── C2: SGWI (SRT-Guided Warm Initialization) + Dual Fisher ─────────
    sgwi: Optional[bool] = field(
        default=True,
        metadata={
            "help": "Enable SGWI warm initialization from past LoRA adapters. "
                    "True = sgwi_full (warm-init A+B from weighted past adapters), "
                    "False = full_lora (standard LoRA, both A+B trainable, no warm-init)."
        },
    )
    dual_fisher: Optional[bool] = field(
        default=False,
        metadata={
            "help": "Enable Dual Fisher regularization (L2 penalty around past θ*). "
                    "Only effective when --sgwi True. "
                    "When True: uses --lambda_emb as regularization strength (default 0.01 if not set). "
                    "When False: forces lambda_emb=0.0 (no regularization)."
        },
    )
    lambda_emb: Optional[float] = field(
        default=0.0,
        metadata={
            "help": "Dual Fisher regularization strength. "
                    "Auto-set to 0.01 when --dual_fisher True and lambda_emb not explicitly provided. "
                    "Ignored when --dual_fisher False."
        },
    )

    denser_evaluation: Optional[bool] = field(
        default=False,
        metadata={"help": "If specifid, the model will do more evaluation at the beginning of training."}
    )
    do_demo: bool = field(default=False, metadata={"help": "Whether to run the model as a demo in the terminal."})
    lamda_1: float = field(default = 0.5)
    lamda_2: float = field(default = 0)
    
    kl_ratio: Optional[float] = field(
        default=0.5,
        metadata={"help": "ratio of the replay kl loss"}
    )
    data_replay_freq: Optional[int] = field(
        default=-1,
        metadata={"help": "replay frequency"}
    )
    replay_after_n_epoch: Optional[int] = field(
        default=0,
        metadata={"help": "replay after n epoch"}
    )
    remove_unused_columns: Optional[bool] = field(
        default=False,
    )
    attn_lr: Optional[float] = field(
        default=0,
        metadata={"help": "learning rate of the attention module"}
    )
    model_name: Optional[str] = field(
        default='SAPT',
        metadata={"help": "models' name"}
    )
    chunk: Optional[int] = field(
        default=1,
        metadata={"help": "models' name"}
    )
    threshold: Optional[float] = field(
        default=0.99,
        metadata={"help": "learning rate of the attention module"}
    )
    transthreshold: Optional[float] = field(
        default=0.99,
        metadata={"help": "learning rate of the attention module"}
    )
    op: Optional[int] = field(
        default=0,
        metadata={"help": "occupy"}
    )
    persent: Optional[float] = field(
        default=1.0,
        metadata={"help": "persent"}
    )
    n_groups: Optional[int] = field(
        default=1,
        metadata={"help": "persent"}
    )
    lambda1: Optional[float] = field(
        default=1.0,
        metadata={"help": "persent"}
    )
    lambda2: Optional[float] = field(
        default=1.0,
        metadata={"help": "persent"}
    )


def main():
    # See all possible arguments in src/transformers/training_args.py
    # or by passing the --help flag to this script.
    # We now keep distinct sets of args, for a cleaner separation of concerns.

    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
        # If we pass only one argument to the script and it's the path to a json file,
        # let's parse it to get our arguments.
        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
    else:
        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
    training_args._frozen = False

    # Safety: T5 models produce NaN/zero gradients with fp16. Force disable.
    if training_args.fp16:
        print("=" * 60)
        print("WARNING: --fp16 detected. T5 models are unstable with fp16")
        print("(causes NaN loss or zero gradients). Forcing fp16=False.")
        print("Use --gradient_checkpointing for memory savings instead.")
        print("=" * 60)
        training_args.fp16 = False

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        handlers=[logging.StreamHandler(sys.stdout)],
    )
    log_level = training_args.get_process_log_level()
    logger.setLevel(log_level)
    datasets.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.enable_default_handler()
    transformers.utils.logging.enable_explicit_format()

    # Log on each process the small summary:
    logger.warning(
        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
    )
    logger.info(f"Training/evaluation parameters {training_args}")

    # Detecting last checkpoint.
    last_checkpoint = None
    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
        last_checkpoint = get_last_checkpoint(training_args.output_dir)
        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
            raise ValueError(
                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                "Use --overwrite_output_dir to overcome."
            )
        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
            logger.info(
                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
            )

    # Set seed before initializing model.
    set_seed(training_args.seed)
    data_cache_dir = gen_cache_path(training_args.output_dir, data_args)

    task_order = data_args.task_order.split(',')
    
    cur_task = data_args.task_config_dir.split('/')[-1]
    cur_task_id = task_order.index(cur_task)

    download_config = DownloadConfig
    download_config.local_files_only = True
    # Get the CL dataset
    dataset_script_path = os.path.join(CURRENT_DIR, "cl_dataset.py")
    if not os.path.exists(dataset_script_path):
        raise FileNotFoundError(f"Dataset script not found: {dataset_script_path}")

    abs_data_dir = os.path.abspath(data_args.data_dir) if data_args.data_dir else None
    abs_task_config_dir = os.path.abspath(data_args.task_config_dir) if data_args.task_config_dir else None
    raw_datasets = load_dataset(
        dataset_script_path,
        data_dir=abs_data_dir,
        download_config=download_config,
        task_config_dir=abs_task_config_dir,
        trust_remote_code=True,
        # cache_dir=data_cache_dir,  # for debug, change dataset size, otherwise open it
        max_num_instances_per_task=data_args.max_num_instances_per_task,
        max_num_instances_per_eval_task=data_args.max_num_instances_per_eval_task,
        num_examples=data_args.num_examples
    )
    raw_datasets.cleanup_cache_files()

    # Load pretrained model and tokenizer
    config = AutoConfig.from_pretrained(
        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )
    tokenizer = AutoTokenizer.from_pretrained(
        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        use_fast=model_args.use_fast_tokenizer,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )

    prompt_config = {
        'seq_len': data_args.max_source_length,
        'mlp_hidden_dim': model_args.mlp_hidden_dim,
        'attn_temperature': model_args.attn_temperature,
        'previous_lora_path': model_args.previous_lora_path,
        'previous_prompt_key_path': model_args.previous_prompt_key_path,
        'task_id': cur_task_id,
        'run_single': model_args.run_single,
        'lora_r': model_args.lora_r,
        'lora_alpha': model_args.lora_alpha,
        'lora_dropout': model_args.lora_dropout
    }

    from t5_gainlora import T5ForConditionalGeneration

    model = T5ForConditionalGeneration.from_pretrained(
            model_args.model_name_or_path,
            prompt_config,
            from_tf=bool(".ckpt" in model_args.model_name_or_path),
            config=config,
            cache_dir=model_args.cache_dir,
            revision=model_args.model_revision,
            use_auth_token=True if model_args.use_auth_token else None,
        )

    model.persent = training_args.persent
    model.resize_token_embeddings(len(tokenizer))

    # FIX: from_pretrained wraps model construction in no_init_weights() context,
    # which replaces nn.init.kaiming_uniform_ with a no-op. This leaves lora_A
    # as all zeros (from torch.zeros in constructor), making LoRA output = 0
    # and all lora_B gradients = 0. Re-initialize lora_A here.
    _n_reinit = 0
    for _module in model.modules():
        if hasattr(_module, 'lora_A') and hasattr(_module, 'lora_B') and hasattr(_module, 'reset_parameters'):
            nn.init.kaiming_uniform_(_module.lora_A, a=math.sqrt(5))
            _n_reinit += 1
    print(f"[FIX] Re-initialized lora_A in {_n_reinit} LoRA layers with kaiming_uniform_")

    # FIX: Also re-initialize trans_input nn.Linear layers and prompt_key.
    # no_init_weights() makes nn.Linear.reset_parameters() a no-op,
    # leaving weights as torch.empty() (uninitialized). If weights=0,
    # trans_input output=0, then cal_attention does x/x.norm() = 0/0 = NaN,
    # which propagates to loss=NaN (reported as train_loss=0.0 by nan filter).
    if hasattr(model, 'encoder') and hasattr(model.encoder, 'trans_input'):
        _n_linear = 0
        for _layer in model.encoder.trans_input:
            if isinstance(_layer, nn.Linear):
                nn.init.kaiming_uniform_(_layer.weight, a=math.sqrt(5))
                if _layer.bias is not None:
                    fan_in, _ = nn.init._calculate_fan_in_and_fan_out(_layer.weight)
                    bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
                    nn.init.uniform_(_layer.bias, -bound, bound)
                _n_linear += 1
        print(f"[FIX] Re-initialized {_n_linear} trans_input Linear layers")
    if hasattr(model, 'encoder') and hasattr(model.encoder, 'prompt_key'):
        nn.init.uniform_(model.encoder.prompt_key.data, -1, 1)
        print(f"[FIX] Re-initialized prompt_key with uniform(-1, 1)")

    # ── SRT: attach frozen encoder for routing ─────────────────────────
    # Theory (contribution_UNIFIED.md): SRT signatures {μ_t, Σ_t} are computed
    # on FROZEN pretrained encoder hidden states — NOT adapted (LoRA) outputs.
    # Uses T5EncoderModel (same as routing_analysis/extract_embeddings_t5.py).
    if training_args.use_srt_router:
        from transformers import T5EncoderModel
        frozen_encoder = T5EncoderModel.from_pretrained(
            model_args.model_name_or_path,
            cache_dir=model_args.cache_dir,
        )
        frozen_encoder.eval()
        for p in frozen_encoder.parameters():
            p.requires_grad = False
        model.encoder.encoder_frozen = frozen_encoder
        print(f"[SRT] Attached frozen T5EncoderModel from {model_args.model_name_or_path}")

    try:
        local_rank = int(os.environ['LOCAL_RANK'])
        device = torch.device(f"cuda:{local_rank}")
    except:
        device = torch.device(f"cuda:0")
    if model_args.load_checkpoint_from:
        if not os.path.exists(model_args.load_checkpoint_from):
            logger.warning(f"load_checkpoint_from not found: {model_args.load_checkpoint_from}, skipping load")
        else:
            print("----------Loading Previous Query Projection Layer----------")
            model.encoder.trans_input.load_state_dict(torch.load(model_args.load_checkpoint_from, map_location=device, weights_only=True))
            model.encoder.previous_trans_input.input_linear[0].data.copy_(torch.load(model_args.load_checkpoint_from, map_location=device, weights_only=True)['0.weight'])
            model.encoder.previous_trans_input.output_linear[0].data.copy_(torch.load(model_args.load_checkpoint_from, map_location=device, weights_only=True)['2.weight'])
            model.encoder.previous_trans_input.state_dict()
            if cur_task_id > 1:
                model.encoder.previous_trans_input.input_linear[1:].data.copy_(torch.load(model_args.load_checkpoint_from.replace('trans_input.pt', 'previous_trans_input.pt'), map_location=device, weights_only=True)['input_linear'])
                model.encoder.previous_trans_input.output_linear[1:].data.copy_(torch.load(model_args.load_checkpoint_from.replace('trans_input.pt', 'previous_trans_input.pt'), map_location=device, weights_only=True)['output_linear'])
            print("----------Loading Previous Query Projection Layer Done----------")

    if model_args.previous_lora_path:
        previous_lora_list = model_args.previous_lora_path.split(',')
        previous_lora_list.reverse()
        print(previous_lora_list)
        print("----------Loading Previous LoRA Weights----------")
        for i, path in enumerate(previous_lora_list):
            lora_A = torch.load(os.path.join(path, "lora_weights_A.pt"), map_location=device, weights_only=True)
            lora_B = torch.load(os.path.join(path, "lora_weights_B.pt"), map_location=device, weights_only=True)
            ## Encoder Layer       
            for j in range(config.num_layers):
                model.encoder.block[j].layer[0].SelfAttention.previous_lora_weights_q[i].lora_A.data.copy_(
                    lora_A[f"encoder.block.{j}.layer.0.SelfAttention.lora_q.lora_A"]
                )
                model.encoder.block[j].layer[0].SelfAttention.previous_lora_weights_q[i].lora_B.data.copy_(
                    lora_B[f"encoder.block.{j}.layer.0.SelfAttention.lora_q.lora_B"]
                )
                model.encoder.block[j].layer[0].SelfAttention.previous_lora_weights_v[i].lora_A.data.copy_(
                    lora_A[f"encoder.block.{j}.layer.0.SelfAttention.lora_v.lora_A"]
                )
                model.encoder.block[j].layer[0].SelfAttention.previous_lora_weights_v[i].lora_B.data.copy_(
                    lora_B[f"encoder.block.{j}.layer.0.SelfAttention.lora_v.lora_B"]
                )
            ## Decoder Layaer
            for j in range(config.num_layers):
                model.decoder.block[j].layer[0].SelfAttention.previous_lora_weights_q[i].lora_A.data.copy_(
                    lora_A[f"decoder.block.{j}.layer.0.SelfAttention.lora_q.lora_A"]
                )
                model.decoder.block[j].layer[0].SelfAttention.previous_lora_weights_q[i].lora_B.data.copy_(
                    lora_B[f"decoder.block.{j}.layer.0.SelfAttention.lora_q.lora_B"]
                )
                model.decoder.block[j].layer[0].SelfAttention.previous_lora_weights_v[i].lora_A.data.copy_(
                    lora_A[f"decoder.block.{j}.layer.0.SelfAttention.lora_v.lora_A"]
                )
                model.decoder.block[j].layer[0].SelfAttention.previous_lora_weights_v[i].lora_B.data.copy_(
                    lora_B[f"decoder.block.{j}.layer.0.SelfAttention.lora_v.lora_B"]
                )
                model.decoder.block[j].layer[1].EncDecAttention.previous_lora_weights_q[i].lora_A.data.copy_(
                    lora_A[f"decoder.block.{j}.layer.1.EncDecAttention.lora_q.lora_A"]
                )
                model.decoder.block[j].layer[1].EncDecAttention.previous_lora_weights_q[i].lora_B.data.copy_(
                    lora_B[f"decoder.block.{j}.layer.1.EncDecAttention.lora_q.lora_B"]
                )
                model.decoder.block[j].layer[1].EncDecAttention.previous_lora_weights_v[i].lora_A.data.copy_(
                    lora_A[f"decoder.block.{j}.layer.1.EncDecAttention.lora_v.lora_A"]
                )
                model.decoder.block[j].layer[1].EncDecAttention.previous_lora_weights_v[i].lora_B.data.copy_(
                    lora_B[f"decoder.block.{j}.layer.1.EncDecAttention.lora_v.lora_B"]
                )

    # Move all previous LoRA weights to CPU to free GPU VRAM.
    # They will be moved to GPU temporarily during forward in agg_lora_states.
    if model_args.previous_lora_path:
        for module in model.modules():
            for attr in ('previous_lora_weights_q', 'previous_lora_weights_v'):
                prev = getattr(module, attr, None)
                if prev is not None:
                    prev.to('cpu')
        print("[FIX] Moved all previous LoRA weights to CPU")

    # ── Load current task's saved LoRA weights (eval-only resume) ───
    if model_args.current_lora_path:
        _cur_lora_A_path = os.path.join(model_args.current_lora_path, "lora_weights_A.pt")
        _cur_lora_B_path = os.path.join(model_args.current_lora_path, "lora_weights_B.pt")
        if os.path.exists(_cur_lora_A_path) and os.path.exists(_cur_lora_B_path):
            print(f"[EVAL-ONLY] Loading current task LoRA from {model_args.current_lora_path}")
            _cur_lora_A = torch.load(_cur_lora_A_path, map_location=device)
            _cur_lora_B = torch.load(_cur_lora_B_path, map_location=device)
            for _name, _param in model.named_parameters():
                if _name in _cur_lora_A:
                    _param.data.copy_(_cur_lora_A[_name].to(device))
                elif _name in _cur_lora_B:
                    _param.data.copy_(_cur_lora_B[_name].to(device))
            # Also load trans_input if available
            _trans_path = os.path.join(model_args.current_lora_path, "trans_input.pt")
            if os.path.exists(_trans_path):
                model.encoder.trans_input.load_state_dict(
                    torch.load(_trans_path, map_location=device, weights_only=True))
            # Load SRT signatures
            if training_args.use_srt_router and training_args.srt_load_path is None:
                training_args.srt_load_path = model_args.current_lora_path
            print(f"[EVAL-ONLY] Loaded current task weights for eval-only mode")
        else:
            logger.warning(f"[EVAL-ONLY] current_lora_path set but files not found: {_cur_lora_A_path}")

    for name, param in model.named_parameters():
        param.requires_grad = False
        if ("lora" in name and "previous_lora_weights" not in name) or ("trans_input" in name and "previous_trans_input" not in name) or "prompt_key" in name:
            param.requires_grad = True

    # ── C2: Always unfreeze lora_A (both full_lora and sgwi_full need trainable A+B) ──
    if training_args.model_name == 'gainlora':
        _n_unfrozen = 0
        for name, param in model.named_parameters():
            if "lora_A" in name and "previous_lora_weights" not in name:
                param.requires_grad = True
                _n_unfrozen += 1
        _sgwi_mode = 'sgwi_full' if training_args.sgwi else 'full_lora'
        print(f"[C2] Unfroze lora_A: {_n_unfrozen} params (sgwi={training_args.sgwi} → mode={_sgwi_mode})")

    total_params, params = 0, 0
    for n, p in model.named_parameters():
        if p.requires_grad:
            print(n)
            total_params += p.numel()
        params += p.numel()

    print(
        "Total number of parameters: {}M, rate: {}%".format(
            total_params // 1000 / 1000, round(total_params / params * 100, 2)
        )
    )

    if (
            hasattr(model.config, "max_position_embeddings")
            and model.config.max_position_embeddings < data_args.max_source_length
    ):
        if model_args.resize_position_embeddings is None:
            logger.warning(
                f"Increasing the model's number of position embedding vectors from {model.config.max_position_embeddings} "
                f"to {data_args.max_source_length}."
            )
            model.resize_position_embeddings(data_args.max_source_length)
        elif model_args.resize_position_embeddings:
            model.resize_position_embeddings(data_args.max_source_length)
        else:
            raise ValueError(
                f"`--max_source_length` is set to {data_args.max_source_length}, but the model only has {model.config.max_position_embeddings}"
                f" position encodings. Consider either reducing `--max_source_length` to {model.config.max_position_embeddings} or to automatically "
                "resize the model's position encodings by passing `--resize_position_embeddings`."
            )

    if training_args.do_train:
        if "train" not in raw_datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = raw_datasets["train"]
        if data_args.max_train_samples is not None:
            train_dataset = train_dataset.select(range(data_args.max_train_samples))

    if training_args.do_eval:
        if "validation" not in raw_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = raw_datasets["validation"]
        if data_args.max_eval_samples is not None:
            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))

    if training_args.do_predict:
        if "test" not in raw_datasets:
            raise ValueError("--do_predict requires a test dataset")
        predict_dataset = raw_datasets["test"]
        if data_args.max_predict_samples is not None:
            predict_dataset = predict_dataset.select(range(data_args.max_predict_samples))

    # Data collator
    label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
    data_collator = DataCollator(
        tokenizer,
        model=model,
        padding="longest",
        max_source_length=data_args.max_source_length,
        max_target_length=data_args.max_target_length,
        label_pad_token_id=label_pad_token_id,
        pad_to_multiple_of=8 if training_args.fp16 else None,
        add_task_name=data_args.add_task_name,
        add_dataset_name=data_args.add_dataset_name,
        num_examples=data_args.num_examples,
        input_record_file=data_args.input_record_file
    )
    # we don't want to remove unused columns because we will prepare each batch during training,
    # and some of the information will also be used in evaluation.
    training_args.remove_unused_columns = False

    replay_dataset_dict, replay_label_dict = None, None
    data_collator_replay = None
    if training_args.data_replay_freq != -1:
        data_dir = data_args.gen_data_dir
        data_collator_replay = DataCollator(
            tokenizer,
            model=model,
            padding="longest",
            max_source_length=data_args.max_source_length,
            max_target_length=data_args.max_target_length,
            label_pad_token_id=label_pad_token_id,
            pad_to_multiple_of=8 if training_args.fp16 else None,
            add_task_name=data_args.add_task_name,
            add_dataset_name=data_args.add_dataset_name,
            add_instruction_replay=data_args.add_instruction_replay,
            num_examples=data_args.num_examples,
            input_record_file=data_args.input_record_file)

        replay_dataset_dict, replay_label_dict = None, None
        if model_args.load_checkpoint_from:
            replay_dataset_dict = {}
            abs_data_dir_replay = os.path.abspath(data_dir) if data_dir else None
            for idx in range(cur_task_id):
                raw_datasets_gen = load_dataset(
                    dataset_script_path,
                    data_dir=abs_data_dir_replay,
                    download_config=download_config,
                    task_config_dir=os.path.abspath(task_config[task_order[idx]]) if task_config[task_order[idx]] else None,
                    trust_remote_code=True,
                    cache_dir=data_cache_dir,  # for debug, change dataset size, otherwise open it
                    max_num_instances_per_task=data_args.max_num_instances_per_task,
                    max_num_instances_per_eval_task=data_args.max_num_instances_per_eval_task,
                    num_examples=data_args.num_examples)

                replay_dataset_dict[task_order[idx]] = raw_datasets_gen["train"]
                print(raw_datasets_gen)

            replay_label_dict = {}
            for idx in range(0,cur_task_id):
                with open(os.path.join("../logs_and_outputs/" + training_args.run_name + "/outputs/", str(idx+1)+"-"+task_order[idx], "saved_weights", "attention_weights.pkl"), 'rb') as f:
                    attn_weights = pickle.load(f)
                replay_label_dict[task_order[idx]] = torch.cat([torch.tensor([0.] * (cur_task_id - idx)), torch.tensor(attn_weights)], dim=0).to(dtype=torch.bfloat16, device='cuda')
            print(replay_label_dict)
        print('-'*50)

    # Metric
    def compute_rouge_metrics(dataset, preds, save_prefix=None):
        decoded_preds = skip_instructions(model, preds, tokenizer)
        references = [e["Instance"]["label"] for e in dataset]
        result = compute_metrics(predictions=decoded_preds, references=references)
        result_per_task = compute_grouped_metrics(predictions=decoded_preds, references=references,
                                                  groups=dataset["Task"])
        result.update(result_per_task)
        categories = dataset["Dataset"]
        result_per_category = compute_grouped_metrics(predictions=decoded_preds, references=references,
                                                      groups=categories)
        result.update(result_per_category)
        prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
        result["gen_len"] = np.mean(prediction_lens)
        result = {k: round(v, 4) for k, v in result.items()}
        if save_prefix is not None:
            with open(os.path.join(training_args.output_dir, f"{save_prefix}_eval_predictions.jsonl"), "w") as fout:
                for example, pred in zip(dataset, decoded_preds):
                    fout.write(json.dumps({
                        "Task": example["Task"],
                        "Dataset": example["Dataset"],
                        "Instance": example["Instance"],
                        "Prediction": pred
                    }) + "\n")
        return result
    print(f"-----Gradient checkpointing: {training_args.gradient_checkpointing} -----")
    if training_args.gradient_checkpointing:
        model.gradient_checkpointing_enable(
            gradient_checkpointing_kwargs={"use_reentrant": False}
        )
        # REQUIRED for use_reentrant=False: if embedding inputs don't have requires_grad,
        # no grad_fn is created and loss.backward() crashes with
        # "element 0 of tensors does not require grad and does not have a grad_fn"
        model.enable_input_require_grads()

    world_size = int(os.environ.get("WORLD_SIZE", 1))
    training_args.step_per_epoch = math.ceil(len(raw_datasets["train"]) / training_args.per_device_train_batch_size / world_size / training_args.gradient_accumulation_steps)
    training_args.eval_steps = 5 * training_args.step_per_epoch
    training_args.save_steps = 5 * training_args.step_per_epoch

    for module in model.modules():
        if hasattr(module, 'get_feature'):
            module.get_chunk(training_args.chunk)
    model.encoder.get_chunk(training_args.chunk)
    if training_args.model_name == 'gainlora':
        # ── C2: Always use SGWI_DualFisher_Trainer (supports both full_lora and sgwi_full) ──
        _sgwi_mode = 'sgwi_full' if training_args.sgwi else 'full_lora'
        # Dual Fisher: --dual_fisher True enables regularization
        if training_args.dual_fisher:
            # Auto-set lambda_emb to 0.01 if not explicitly provided (still 0.0 default)
            if training_args.lambda_emb == 0.0:
                training_args.lambda_emb = 0.01
                print(f"[C2] dual_fisher=True, auto-setting lambda_emb=0.01")
        else:
            training_args.lambda_emb = 0.0  # Force disable
        _lambda_emb = training_args.lambda_emb
        print(f"[C2] sgwi={training_args.sgwi} → mode={_sgwi_mode}, dual_fisher={training_args.dual_fisher}, lambda_emb={_lambda_emb}")

        from sgwi_trainer import SGWI_DualFisher_Trainer
        trainer = SGWI_DualFisher_Trainer(
            model=model,
            args=training_args,
            train_dataset=train_dataset if training_args.do_train else None,
            cur_task_id=cur_task_id,
            task_order=task_order,
            data_collator_replay=data_collator_replay,
            replay_dataset_dict=replay_dataset_dict,
            replay_label_dict=replay_label_dict,
            eval_dataset=eval_dataset if training_args.do_eval else None,
            tokenizer=tokenizer,
            data_collator=data_collator,
            compute_metrics=compute_rouge_metrics,
            callbacks=[DenserEvalCallback] if training_args.denser_evaluation else None,
            srt_metric_mode=training_args.srt_metric_mode,
            srt_shrink=training_args.srt_shrink,
            srt_shrink_factor=training_args.srt_shrink_factor,
            srt_max_emb_samples=training_args.srt_max_emb_samples,
            srt_load_path=training_args.srt_load_path,
            srt_skip_forward=training_args.srt_skip_forward,
            sgwi_mode=_sgwi_mode,
            lambda_emb=_lambda_emb,
        )
        if training_args.do_train:
            trainer.get_reg_matrix()
    else:
        raise NotImplementedError(f"Unknown model_name: {training_args.model_name}")

    trainer.is_deepspeed_enabled = False
    print("is_deepspeed_enabled", trainer.is_deepspeed_enabled)

    # ============ QUICK SANITY CHECK: LoRA weights ============
    if training_args.do_train:
        print("=" * 60)
        print("[SANITY] Checking LoRA layer initialization...")
        model.to(device)
        _lora = None
        for _m in model.modules():
            if hasattr(_m, 'lora_A') and hasattr(_m, 'lora_B'):
                _lora = _m
                break
        if _lora is not None:
            _a_ok = _lora.lora_A.data.norm().item() > 0
            print(f"  lora_A norm={_lora.lora_A.data.norm().item():.6f}, requires_grad={_lora.lora_A.requires_grad} {'OK' if _a_ok else 'ZERO - BUG!'}")
            print(f"  lora_B norm={_lora.lora_B.data.norm().item():.6f}, requires_grad={_lora.lora_B.requires_grad}")
            # Quick forward+backward test to verify gradient flow
            _test_x = torch.randn(1, 3, _lora.lora_A.shape[1], device=device)
            _lora.lora_B.grad = None
            _y = _lora(_test_x)
            _y.sum().backward()
            _b_grad = _lora.lora_B.grad.norm().item() if _lora.lora_B.grad is not None else 0
            print(f"  lora_B.grad norm={_b_grad:.6e} {'OK' if _b_grad > 0 else 'ZERO - BUG!'}")
            model.zero_grad()
            if not _a_ok:
                raise RuntimeError("lora_A is all zeros! from_pretrained no_init_weights fix failed.")
        # Check trans_input weights
        if hasattr(model, 'encoder') and hasattr(model.encoder, 'trans_input'):
            for _i, _layer in enumerate(model.encoder.trans_input):
                if isinstance(_layer, nn.Linear):
                    _w_norm = _layer.weight.data.norm().item()
                    _w_zero = (_layer.weight.data == 0).all().item()
                    print(f"  trans_input[{_i}] weight norm={_w_norm:.6f}, all_zero={_w_zero} {'BUG!' if _w_zero else 'OK'}")
        print("=" * 60)
    # ============ END SANITY CHECK ============

    all_metrics = {"run_name": training_args.run_name}

    # Training
    if training_args.do_train:
        checkpoint = None
        if training_args.resume_from_checkpoint is not None:
            checkpoint = training_args.resume_from_checkpoint
        elif last_checkpoint is not None:
            checkpoint = last_checkpoint
        train_result = trainer.train(resume_from_checkpoint=checkpoint)

        save_path = training_args.output_dir + "/saved_weights"
        if not os.path.exists(save_path):
            try:
                os.makedirs(save_path)
            except:
                pass

        if not prompt_config["run_single"]:
            # Save previous_trans_input (needed for both SRT and non-SRT)
            if prompt_config["previous_prompt_key_path"] is not None:
                previous_trans_input = deepcopy(trainer.model.encoder.previous_trans_input.state_dict())
                torch.save(previous_trans_input, os.path.join(save_path, 'previous_trans_input.pt'))
            torch.save(trainer.model.encoder.trans_input.state_dict(), os.path.join(save_path, 'trans_input.pt'))

        if prompt_config["previous_prompt_key_path"] is not None:
            torch.save(lora_state_dict_A(model, task_name=cur_task), os.path.join(save_path, 'lora_weights_A.pt'))
            torch.save(lora_state_dict_B(model, task_name=cur_task), os.path.join(save_path, 'lora_weights_B.pt'))
            if not prompt_config["run_single"]:
                torch.save(torch.cat([trainer.model.encoder.prompt_key, trainer.model.encoder.previous_prompts_keys], dim=0).data, os.path.join(save_path, 'prompts_keys_till_now.pt'))
        else:
            torch.save(lora_state_dict_A(model, task_name=cur_task), os.path.join(save_path, 'lora_weights_A.pt'))
            torch.save(lora_state_dict_B(model, task_name=cur_task), os.path.join(save_path, 'lora_weights_B.pt'))
            if not prompt_config["run_single"]:
                torch.save(trainer.model.encoder.prompt_key.data, os.path.join(save_path, 'prompts_keys_till_now.pt'))
        tokenizer.save_pretrained(save_path)

        metrics = train_result.metrics
        max_train_samples = (
            data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
        )
        metrics["train_samples"] = min(max_train_samples, len(train_dataset))

        trainer.log_metrics("train", metrics)
        trainer.save_metrics("train", metrics)
        trainer.save_state()
        logger.info(f"Metrics {metrics}")
        all_metrics.update(metrics)

        # SRT: compute and store statistical signature AFTER training this task
        if training_args.model_name == 'gainlora':
            if hasattr(trainer, 'on_task_end'):
                trainer.on_task_end(task_order[cur_task_id])
            if hasattr(trainer, 'save_srt_signatures'):
                trainer.save_srt_signatures(save_path)

    # Evaluation
    results = {}
    # in case the batch is shorter than max length, the output should be padded
    max_new_tokens = (
        training_args.generation_max_length
        if training_args.generation_max_length is not None
        else data_args.max_target_length
    )

    num_beams = data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
    repetition_penalty = data_args.repetition_penalty

    if training_args.do_predict or training_args.do_train:
        print("*** Prediction ***")
        logger.info("*** Prediction ***")

        if data_args.max_predict_samples is not None:
            predict_dataset = predict_dataset.select(range(data_args.max_predict_samples))

        # SRT: single predict pass with is_inference=True to collect routing stats.
        # Routing is ALWAYS hard one-hot (same result regardless of is_inference).
        trainer.model.encoder.is_inference = True
        predict_results = trainer.predict(
            predict_dataset,
            metric_key_prefix="predict",
            max_new_tokens=max_new_tokens,
            num_beams=num_beams,
            repetition_penalty=repetition_penalty,
            pad_token_id=tokenizer.pad_token_id
        )

        # Save routing stats (attention_weights.pkl)
        if not prompt_config["run_single"]:
            save_path = training_args.output_dir + "/saved_weights"
            with open(os.path.join(save_path, "attention_weights.pkl"), 'wb') as f:
                # Filter out 1D arrays (incompatible batch sizes cause shape mismatch)
                all_2d = [x for x in trainer.model.encoder.all_attn_weights if x.ndim == 2]
                if all_2d:
                    attn_w = np.array(np.concatenate(all_2d)).mean(axis=0)
                    print(f"{'*'*20} Saving Attention Weights {'*'*20}")
                    print(attn_w)
                    pickle.dump(attn_w, f)
                else:
                    # No valid 2D weights → write empty placeholder to avoid EOFError on load
                    print(f"{'*'*20} No valid 2D Attention Weights — saving empty dict {'*'*20}")
                    pickle.dump({}, f)
        trainer.model.encoder.is_inference = False

        # Save metrics
        if training_args.do_predict:
            metrics = predict_results.metrics
            max_predict_samples = (
                data_args.max_predict_samples if data_args.max_predict_samples is not None else len(predict_dataset)
            )
            metrics["predict_samples"] = min(max_predict_samples, len(predict_dataset))

            trainer.log(metrics)
            trainer.log_metrics("predict", metrics)
            trainer.save_metrics("predict", metrics)
            all_metrics.update(metrics)

            outputs_dir = os.path.join("logs_and_outputs", training_args.run_name, "outputs")
            os.makedirs(outputs_dir, exist_ok=True)
            with open(os.path.join(outputs_dir, "task_order.txt"), 'w') as f:
                f.write(data_args.task_order)

    # Clean up redundant Hugging Face checkpoints to save disk space
    import shutil
    import glob
    if training_args.output_dir and os.path.exists(training_args.output_dir):
        checkpoint_dirs = glob.glob(os.path.join(training_args.output_dir, "checkpoint-*"))
        for ckpt_dir in checkpoint_dirs:
            try:
                shutil.rmtree(ckpt_dir)
                logger.info(f"Deleted redundant checkpoint directory: {ckpt_dir}")
            except Exception as e:
                logger.warning(f"Failed to delete {ckpt_dir}: {e}")

    return results


if __name__ == "__main__":
    main()