How does one set the pad token correctly (not to eos) during fine-tuning to avoid model not predicting EOS?

Question

**tldr; what I really want to know is what is the official way to set pad token for fine tuning it wasn't set during original training, so that it doesn't not learn to predict EOS. **

colab: https://colab.research.google.com/drive/1poFdFYmkR_rDM5U5Z2WWjTepMQ8hvzNc?usp=sharing

---

The HF falcon tutorial has the following line:

tokenizer.pad_token = tokenizer.eos_token

it looks strange to me. It make sense pad and eos are the same but then why even make a difference between them in the first place in general?

Note its wrong to do pad = eos. This means during fine-tuning the model will never be trained to output eos (most likely) since eos is treated as pad token and no back propagated:

I just observed that when I set tokenizer.pad_token = tokenizer.eos_token during training, the model won't stop generating during inference, since it was trained to not output the eos token (per discussions above).

I saw this (here https://github.com/huggingface/transformers/issues/22794):

tokenizer.add_special_tokens({'pad_token': '[PAD]'})

But this assumes the model has a pad_token. I think an additional check has to be done that it does have an embedding for pad_token so that there are no run time errors (~type errors in the matrix extraction from the embedding "table"/matrix).

But if one does that some care might be needed to initialize the new token so that it dominates the generation: https://nlp.stanford.edu/~johnhew/vocab-expansion.html

---

code:

def get_model_tokenizer_qlora_falcon7b(model_name: str = "ybelkada/falcon-7b-sharded-bf16",
                                       config: wand.Config,  # todo
                                       lora_alpha=16,  # todo
                                       lora_dropout=0.1,  # todo
                                       lora_r=64,  # todo
                                       bnb_4bit_compute_dtype=torch.float16,  # changed it from Guanaco hf
                                       ) -> tuple:
    """
    Load the Falcon 7B model, quantize it in 4bit and attach LoRA adapters on it.

    bf16 = 1S, 7Exp, 8Mantissa

    Do:
        pip install bitsandbytes
    ref:
        - https://colab.research.google.com/drive/1DOi8MFv4SWN9NImVornZ7t6BgmLoPQO-#scrollTo=AjB0WAqFSzlD
    """
    from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer

    # model_id = "tiiuae/falcon-7b"
    # model_name: str = "ybelkada/falcon-7b-sharded-bf16"

    # - get bnb config for bit-4 base model (bnb lib for using 4bit qlora quantization techniques by tim dettmers)
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,  # load (usually huge) base model in 4 bits
        bnb_4bit_quant_type="nf4",  # normal float 4 for the (usually huge) base model. introduces error but fixed by ft
        # ref: https://gist.github.com/pacman100/1731b41f7a90a87b457e8c5415ff1c14
        bnb_4bit_compute_dtype=bnb_4bit_compute_dtype,
    )

    # - get falcon 4bit model
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        quantization_config=bnb_config,
        trust_remote_code=True  # allows to execute custom code you download from the uploaded model code you are using
    )
    model.config.use_cache = False  # todo: why? https://stackoverflow.com/questions/76633335/why-does-hugging-face-falcon-model-use-mode-config-use-cache-false-why-wouldn

    # get falcon tockenizer
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)  # execs code downloaded from hf hub
    tokenizer.pad_token = tokenizer.eos_token

---

Modifying model gives issues

Darn this still not works:

 UserWarning: You have modified the pretrained model configuration to control generation. This is a deprecated strategy to control generation and will be removed soon, in a future version. Please use a generation configuration file (see https://huggingface.co/docs/transformers/main_classes/text_generation)

code:

"""
sfttrainer (likely using peft) best practices:
https://huggingface.co/docs/trl/main/en/sft_trainer#best-practices

Best practices

Pay attention to the following best practices when training a model with that trainer:

- SFTTrainer always pads by default the sequences to the max_seq_length argument of the SFTTrainer. If none is passed, the trainer will retrieve that value from the tokenizer. Some tokenizers do not provide default value, so there is a check to retrieve the minimum between 2048 and that value. Make sure to check it before training.
- For training adapters in 8bit, you might need to tweak the arguments of the prepare_model_for_int8_training method from PEFT, hence we advise users to use prepare_in_int8_kwargs field, or create the PeftModel outside the SFTTrainer and pass it.
- For a more memory-efficient training using adapters, you can load the base model in 8bit, for that simply add load_in_8bit argument when creating the SFTTrainer, or create a base model in 8bit outside the trainer and pass it.
- If you create a model outside the trainer, make sure to not pass to the trainer any additional keyword arguments that are relative to from_pretrained() method.

todo: why trust_remote_code? I want more details.
"""
import sys

import torch
from peft import LoraConfig

from transformers.modeling_utils import PreTrainedModel

from pdb import set_trace as st


def test_bfloat16_int4(compute_dtype: torch.dtype,
                       use_4bit,
                       ):
    """
python -c "import torch; print(torch.cuda.get_device_capability());"
    todo: check other code test_bfloat16() do we need use_4bit?
    """
    if compute_dtype == torch.float16 and use_4bit:
        major, _ = torch.cuda.get_device_capability()
        if major >= 8:
            print("=" * 80)
            print("Your GPU supports bfloat16, you can accelerate training with the argument --bfloat16")
            print("=" * 80)


def get_model_tokenizer_qlora_falcon7b(
        # -- mode args
        # model_id = "tiiuae/falcon-7b"
        pretrained_model_name_or_path: str = "ybelkada/falcon-7b-sharded-bf16",
        use_cache: bool = True,
        # -- lora args
        lora_alpha=16,  # todo
        lora_dropout=0.1,  # todo, evidence drop out really help? google, crfm, gpt4
        lora_r=64,  # todo
        bnb_4bit_compute_dtype=torch.float16,  # changed it from Guanaco hf

        # -- training args
        output_dir="./results",
        per_device_train_batch_size=4,
        gradient_accumulation_steps=4,
        # paging so that the sudden mem gpu spikes don't cause the run to shut down
        # (I think usually caused by too long seqs)
        # todo: why 32 bit opt?
        # todo: paged nadamw opt?
        optim="paged_adamw_32bit",
        save_steps=10,
        logging_steps=10,
        learning_rate=2e-4,
        max_grad_norm=0.3,
        max_steps=500,
        warmup_ratio=0.03,
        lr_scheduler_type="constant",
        # -- quant. args (not recommended to be changed unless you know what your doing?)
        load_in_4bit=True,  # load (usually huge) base model in 4 bits
        bnb_4bit_quant_type="nf4",  # normal float 4 for the (large) base models qlora
) -> tuple:
    """
    Load the Falcon 7B model, quantize it in 4bit and attach LoRA adapters on it.

    bf16 = 1S, 7Exp, 8Mantissa
    hypothesis: 7b trained due to 6.7 emergence rumour, I still don't think emergence is real.
    Notes:
        - ft a model is very specific to the model, tokenizer and training scheme. Thus we return
            - model, tokenizer, ft config (peft config), training args

    ref:
        - https://colab.research.google.com/drive/1DOi8MFv4SWN9NImVornZ7t6BgmLoPQO-#scrollTo=AjB0WAqFSzlD
    """
    from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer

    # - Get bnb config for bit-4 base model (bnb lib for using 4bit qlora quantization techniques by tim dettmers)
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=load_in_4bit,  # load (usually huge) base model in 4 bits
        bnb_4bit_quant_type=bnb_4bit_quant_type,  # normal float 4 for the (usually huge) base model
        bnb_4bit_compute_dtype=bnb_4bit_compute_dtype,  # if you can, during computation use bf16
    )

    # - Get falcon 4bit model
    # todo, where is this being saved & how to download quicker
    model = AutoModelForCausalLM.from_pretrained(
        pretrained_model_name_or_path=pretrained_model_name_or_path,
        quantization_config=bnb_config,
        trust_remote_code=True  # allows to execute custom code you download from the uploaded model code you are using
    )
    print(f'{type(model)=}')
    print(f'{model=}')
    # this is here to save gpu vram. Likely only needed when using 40b or when oom issues happen ref: https://stackoverflow.com/questions/76633335/why-does-hugging-face-falcon-model-use-mode-config-use-cache-false-why-wouldn
    model.config.use_cache = use_cache
    print(f'{type(model)=}')

    # - Get falcon tokenizer
    tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path,
                                              trust_remote_code=True)  # execs code downloaded from hf hub
    # tokenizer.pad_token = tokenizer.eos_token  # ref: https://stackoverflow.com/questions/76633368/why-does-the-falcon-qlora-tutorial-code-use-eos-token-as-pad-token
    # tokenizer.add_special_tokens({'pad_token': '[PAD]'})  # I think this is fine if during the training pad is ignored
    tokenizer.add_special_tokens({'pad_token': '<|pad|>'})  # I think this is fine if during the training pad is ignored

    # - Modify model
    # add pad token embed
    model.resize_token_embeddings(len(tokenizer))  # todo: I think this is fine if during the training pad is ignored
    model.transformer.word_embeddings.padding_idx = len(tokenizer) - 1
    model.config.max_new_tokens = len(tokenizer)
    # model.config.min_length = 1
    print(f'{model=}')
    print(f'{type(tokenizer)=}')
    print(f'{tokenizer.pad_token=}')
    # data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False) todo

    # - Get falcon lora config
    peft_config = LoraConfig(
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        r=lora_r,
        bias="none",
        task_type="CAUSAL_LM",
        # model card for falcon tiiuae/falcon-7b: https://huggingface.co/tiiuae/falcon-7b/blob/main/modelling_RW.py
        # does seem to include all trainable params as done by qlora on their own paper
        target_modules=[
            # word_embeddings,
            "query_key_value",
            "dense",
            "dense_h_to_4h",
            "dense_4h_to_h",
            # "lm_head"
        ]
    )
    print(f'{type(peft_config)=}')

    # todo: print the num params of the lora = D1*r + D2*r and num of bytes by prec. (bytes) * num params
    return model, tokenizer, peft_config


# -- tests

def example_test_model_already_has_pad_token():
    """
    if it already has pad token, it likely has a small prob, so we are done.

    compare it's norm with other tokens to verify this is true.

python ~/ultimate-utils/ultimate-utils-proj-src/uutils/hf_uu/model_tokenizer/falcon_uu_mdl_tok.py
    """
    # - the get datasets todo: preprocessing, padding, streaming
    from uutils.hf_uu.data_hf.common import get_guanaco_datsets_add_splits_train_test_only
    trainset, _, testset = get_guanaco_datsets_add_splits_train_test_only()

    # qlora flacon7b
    from uutils.hf_uu.model_tokenizer.falcon_uu_mdl_tok import get_model_tokenizer_qlora_falcon7b
    model, tokenizer, peft_config = get_model_tokenizer_qlora_falcon7b()
    model: PreTrainedModel = model
    print(f'{model=}')
    sent = 'Dogs are great because they are '
    print()

    # print to see if pad tokens are present and if it ignores the tokens at the end
    encoded_input = tokenizer(sent, padding='max_length', max_length=10, return_tensors='pt')
    print(f'{encoded_input=}')

    # Print all special tokens
    print('\n---- start Print all special tokens')
    for token_name, token in tokenizer.special_tokens_map.items():
        print(f"{token_name}: {token}")
    print('\n---- end Print all special tokens')

    # Get the ID for the '[PAD]' token
    try:
        pad_token_id = tokenizer.convert_tokens_to_ids('[PAD]')
    except KeyError:
        raise ValueError("Token [PAD] is not present in the tokenizer vocabulary.")

    # Index into the model's embedding table
    try:
        print(f'{model.get_input_embeddings().weight.size()=}')
        pad_embedding = model.get_input_embeddings().weight[pad_token_id]
    except IndexError:
        raise ValueError(f"Token ID {pad_token_id} is not present in the model's embedding matrix.")

    print(f'{pad_embedding=}')
    print('Success!\n')

    # check it generates something sensible
    # tokenizer.decode(model.generate(**tokenizer(sent, return_tensors='pt'), do_sample=True)[0])
    input_ids, attention_mask = encoded_input['input_ids'], encoded_input['attention_mask']
    predicted_tokens_ids_options = model.generate(input_ids=input_ids, attention_mask=attention_mask, do_sample=True)
    predicted_tokens_ids = predicted_tokens_ids_options[0]
    predicted_sent = tokenizer.decode(predicted_tokens_ids)
    print(f'original sentence: {sent=}')
    print(f'predicted sentence: {predicted_sent=}')
    print('Success2!')


if __name__ == '__main__':
    import time

    start_time = time.time()
    example_test_model_already_has_pad_token()
    print(f"The main function executed in {time.time() - start_time} seconds.\a")

it doesn't like the modifications to the model:

    model.transformer.word_embeddings.padding_idx = len(tokenizer) - 1
    model.config.max_new_tokens = len(tokenizer)

How to fix?

Errors:

/lfs/hyperturing1/0/brando9/miniconda/envs/data_quality/lib/python3.10/site-packages/transformers/generation/utils.py:1259: UserWarning: You have modified the pretrained model configuration to control generation. This is a deprecated strategy to control generation and will be removed soon, in a future version. Please use a generation configuration file (see https://huggingface.co/docs/transformers/main_classes/text_generation)
  warnings.warn(
Setting `pad_token_id` to `eos_token_id`:11 for open-end generation.
/lfs/hyperturing1/0/brando9/miniconda/envs/data_quality/lib/python3.10/site-packages/transformers/generation/utils.py:1452: UserWarning: You are calling .generate() with the `input_ids` being on a device type different than your model's device. `input_ids` is on cpu, whereas the model is on cuda. You may experience unexpected behaviors or slower generation. Please make sure that you have put `input_ids` to the correct device by calling for example input_ids = input_ids.to('cuda') before running `.generate()`.
  warnings.warn(
Traceback (most recent call last):
  File "/lfs/hyperturing1/0/brando9/ultimate-utils/ultimate-utils-proj-src/uutils/hf_uu/model_tokenizer/falcon_uu_mdl_tok.py", line 211, in <module>
    example_test_model_already_has_pad_token()
  File "/lfs/hyperturing1/0/brando9/ultimate-utils/ultimate-utils-proj-src/uutils/hf_uu/model_tokenizer/falcon_uu_mdl_tok.py", line 199, in example_test_model_already_has_pad_token
    predicted_tokens_ids_options = model.generate(input_ids=input_ids, attention_mask=attention_mask, do_sample=True)
  File "/lfs/hyperturing1/0/brando9/miniconda/envs/data_quality/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/lfs/hyperturing1/0/brando9/miniconda/envs/data_quality/lib/python3.10/site-packages/transformers/generation/utils.py", line 1572, in generate
    return self.sample(
  File "/lfs/hyperturing1/0/brando9/miniconda/envs/data_quality/lib/python3.10/site-packages/transformers/generation/utils.py", line 2633, in sample
    next_token_scores = logits_warper(input_ids, next_token_scores)
  File "/lfs/hyperturing1/0/brando9/miniconda/envs/data_quality/lib/python3.10/site-packages/transformers/generation/logits_process.py", line 92, in __call__
    scores = processor(input_ids, scores)
  File "/lfs/hyperturing1/0/brando9/miniconda/envs/data_quality/lib/python3.10/site-packages/transformers/generation/logits_process.py", line 302, in __call__
    indices_to_remove = scores < torch.topk(scores, top_k)[0][..., -1, None]
RuntimeError: "topk_cpu" not implemented for 'Half'

---

Bounty Section: Small GPT2 code example

Yes I agree that pad is assigned to eos. Eos is still eos. But during fine-tuning now the weights wrt to eos are unchanged. This might be an issue since the probability of eos has not shifted to the fine-tuning regime. One possibility is that eos is outputed with less chance. Yes we can still halt production when we see eos but we've not shifted the probability to output eos according to our fine-tuning distribution -- but all other tokens have changed distribution. I think this could be an issue because it's not like the old probability of eos is conserved since all tokens probs have changed except eos + even if the old eos prob was conserved, it's wrt wrong distribution (not the fine tuning one).

e.g.,

    if tokenizer.pad_token_id is None:
        tokenizer.pad_token = tokenizer.eos_token
...
    raw_text_batch='a'
    tokenize_batch={'input_ids': tensor([[   64, 50256, 50256, 50256, 50256]]), 'attention_mask': tensor([[1, 0, 0, 0, 0]])}

but it would have been better to have

    tokenize_batch={'input_ids': tensor([[   64, 50256, 50256, 50256, 50256]]), 'attention_mask': tensor([[1, 1, 0, 0, 0]])}

code

def test_eos_pad():
    from datasets import load_dataset
    import torch
    from transformers import GPT2Tokenizer, GPT2LMHeadModel

    raw_text_batch = 'a'

    tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
    # print(f'{tokenizer.eos_token=}')
    # print(f'{tokenizer.eos_token_id=}')
    # print(f'{tokenizer.pad_token=}')
    # print(f'{tokenizer.pad_token_id=}')

    # print(f'{raw_text_batch=}')
    # tokenize_batch = tokenizer(raw_text_batch, padding="max_length", max_length=5, truncation=True, return_tensors="pt")
    # print(f'{tokenize_batch=}')

    if tokenizer.pad_token_id is None:
        tokenizer.pad_token = tokenizer.eos_token
    probe_network = GPT2LMHeadModel.from_pretrained("gpt2")
    device = torch.device(f"cuda:{0}" if torch.cuda.is_available() else "cpu")
    probe_network = probe_network.to(device)

    print(f'{tokenizer.eos_token=}')
    print(f'{tokenizer.eos_token_id=}')
    print(f'{tokenizer.pad_token=}')
    print(f'{tokenizer.pad_token_id=}')

    print(f'{raw_text_batch=}')
    tokenize_batch = tokenizer(raw_text_batch, padding="max_length", max_length=5, truncation=True, return_tensors="pt")
    print(f'{tokenize_batch=}')
    print('Done')

---

cross:

• https://discord.com/channels/879548962464493619/1126681170957045770/1126681170957045770
• hf https://discuss.huggingface.co/t/why-does-the-falcon-qlora-tutorial-code-use-eos-token-as-pad-token/45954
• so How does one set the pad token correctly (not to eos) during fine-tuning to avoid model not predicting EOS?
• context peft pacman100 code: https://gist.github.com/pacman100/1731b41f7a90a87b457e8c5415ff1c14
• https://twitter.com/BrandoHablando/status/1693676898013061337?s=20

Answer 1

I have not done the falcon model finetuning using QLoRA but I did it using PEFT and bitsandbytes for the 7b variant by loading the model in 8bit and using LoRA rank of 16 with a micro batch size of 8 on a 24Gb GPU. So I would like to mention this if it helps you:

I did not find any <pad> token or <unk> token in falcon. (So e.g. the workaround in the alpaca-lora repository of using token id 0 for padding would not work as that is assigned to another token.) However, using tokenizer.add_special_tokens({'pad_token': '<PAD>'}) after loading the tokenizer and model.resize_token_embeddings(len(tokenizer)) after loading the model in 8-bit worked (at least I did not get any errors during finetuning and the text generation with the finetuned model also worked).

Answer 2

For next-token-prediction, setting EOS to PAD during fine-tuning is actually okay. Note that you may need to manually add the EOS token ([PAD] in this case) to the training data though. As an example, given the input sequence Hello world., we can tokenize the sequence into [Hello, world, ., [PAD]].

The attention mask will be [1, 1, 1, 0] and no next token will be predicted for the [PAD] token.

Remember however that in autoregressive language modelling, the token labels for an input sequence will always be the next token in that sequence. For example:

Input: [Hello, world, .]

Labels: [world, ., [PAD]]

In other words; the model is optimized to generate a [PAD] token after a . token by optimising the token-level cross-entropy loss, even though the [PAD] token (EOS) itself is masked during the attention calculation and thus not back-propagated.

Note that while you can use any token as the EOS token, using the same embedding for the PAD and EOS token is a slight optimisation to remove one entry from the embedding weight matrix. Due to the size of the vocabulary this memory saving is however often negligible (1 token out of e.g. ~50k tokens in GPT2).

Answer 3

Ok I think this is the code that train on first occurence of eos and makes sure the rest are NOT trained on (feedback welcomed):

def custom_collate_fn_train_on_first_eos_occurrence(data: list[dict[str, str]], tokenizer: PreTrainedTokenizer) -> dict[str, torch.Tensor]:
    # Ensure tokenizer has a padding token
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    # Extract sequences
    sequences: list[str] = [example.get("text", "") or "" for example in data]

    # Tokenize the sequences
    tokenized_data = tokenizer(sequences, padding="max_length", max_length=context_length, truncation=True, return_tensors="pt")
    
    # Clone input_ids to labels
    tokenized_data["labels"] = tokenized_data["input_ids"].clone()

    # Set the mask value for the first eos_token in each sequence to 1
    eos_token_id = tokenizer.eos_token_id
    for idx, input_ids in enumerate(tokenized_data["input_ids"]):
        # Find all occurrences of eos_token
        eos_positions = (input_ids == eos_token_id).nonzero(as_tuple=True)[0]
        if eos_positions.nelement() > 0:  # Check if eos_token is present
            first_eos_position = eos_positions[0]
            tokenized_data["attention_mask"][idx, first_eos_position] = 1  # Set the mask value to 1
            
            # Assert that the label for the first occurrence of eos_token is eos_token_id
            assert tokenized_data["labels"][idx, first_eos_position] == eos_token_id, "The label for the first eos_token is incorrect!"
            
            # For all subsequent occurrences of eos_token, set their labels to -100
            for subsequent_eos_position in eos_positions[1:]:
                assert tokenized_data["labels"][idx, subsequent_eos_position] == -100, "The label for the first eos_token is incorrect!"
                # tokenized_data["labels"][idx, subsequent_eos_position] = -100

    return tokenized_data

ref: https://discuss.huggingface.co/t/why-does-the-falcon-qlora-tutorial-code-use-eos-token-as-pad-token/45954/13?u=brando