Why does the falcon QLoRA tutorial code use eos_token as pad_token?

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?


1 Like

From my understanding falcon doesn’t have a pad token defined in the model config, that’s why you define the if statement to avoid getting an error with missing pad token.

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
    model.config.pad_token_id = model.config.eos_token_id
1 Like

but why not use the pad token? are you implying the pad token was not used during training so there is no point in using the pad token as the pad oken and instead use the eos?

For now I’m convinced .pad_token = eos_token is fine for decoder models (even when fine-tuning).

Assume we do eos = pad. Then, the model is trained to predict eos more often, since the loss pad token = the eos token so it doesn’t mask out the extra eos tokens. However, transformers are conditional models. Therefore, in a decoder only model (which is the case I care about), the model only increases eos given it already predicted eos. But at inference we would stop anyway, so it doesn’t matter since we are conditionally weighting eos only if more eos have been seen. Also, if the pad token is never trained on it should always have a low chance so it likely won’t be an issue. In addition if for some reason the pad was trained on (assuming no bugs) then it will be predicted only after a eos. Worst case at inference treat pad as eos to stop generation. For decoders it’s fine. For encoders-decoders it might be an issue since the encoder will encode eos more than usual, which is more of an issue since longer seqs will get eos more often artificially.

See my argument here: Pad and eos distinction.

context: falcon_peft.py · GitHub

Actually I think this discussion is correct: LLaMA FastTokenizer does not add `eos_token_id` at the end. · Issue #22794 · huggingface/transformers · GitHub but need to think through it.

Why is this the case? seem really bizzare to me.

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)


sfttrainer (likely using peft) 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,
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
        # 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?
        # -- 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.
        - ft a model is very specific to the model, tokenizer and training scheme. Thus we return
            - model, tokenizer, ft config (peft config), training args

        - 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(
        trust_remote_code=True  # allows to execute custom code you download from the uploaded model code you are using
    # 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

    # - 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
    # data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False) todo

    # - Get falcon lora config
    peft_config = LoraConfig(
        # 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
            # word_embeddings,
            # "lm_head"

    # 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
    sent = 'Dogs are great because they are '

    # 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 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
        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
        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.")


    # 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=}')

if __name__ == '__main__':
    import time

    start_time = time.time()
    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?

Honestly no idea. Researching it

**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. **

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).


    if tokenizer.pad_token_id is None:
        tokenizer.pad_token = tokenizer.eos_token
    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]])}


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)


    tokenize_batch = tokenizer(raw_text_batch, padding="max_length", max_length=5, truncation=True, return_tensors="pt")
1 Like

I’m still confused

if a model does not have a padding token already (which is common for decoder-only models because they are trained on blocks which do not have any padding). So you never “unlearn” anything.

is true, but then during training eos and pad will be masked. So there is a “wrong” distribution shift for generating EOS now. How to fix this? See details above.

1 Like

Hi all! There’s an interesting story here.

In general you are correct that causal LMs like Falcon are not trained with a pad token, and so the tokenizer does not have one set. This is true for a lot of causal LMs in the Hub. During training, these models are often fed sequences that have been concatenated together and truncated at the maximum sequence length, and so there is never any empty space that needs padding.

The reason we add one later is because a lot of downstream methods use padding and attention masks in some way. However, in many cases it doesn’t really matter what you set the padding token to! This is because the padded tokens will generally be masked by setting the attention_mask to 0, so those tokens will not be attended to by the rest of the sequence.

However, one place the choice of padding token can matter is in the labels when fine-tuning the model. This is because in standard CLM training, the labels are the inputs, shifted by a single position. This would mean that in the final position of the sequence before the padding at the end, the label at that position will be the padding token. When training models with shorter sequences (such as for chat), we generally want them to mark the end of the text they’ve generated, using a token like eos_token. As a result, we commonly just use eos_token as the padding token.

However, depending on your fine-tuning task, you may not want the model to learn to predict eos_token at the end of a sequence - if this is the case, simply change the label at that position to the token you do want, or set the label to -100 to mask the label at that position.

Does that answer the questions you had? Feel free to let me know if I missed anything here!

1 Like

Yes this is what I was going to do because I’m doing fine-tuning for code where syntax matters.

But I need the code. I’ve not had time to write it down. When I do I will share here. To clarify this is what I plan to do:

In the collate function for all seqs in the batch switch the final mask to 1 where the first EOS token is at.

why -100? what does this achieve?

The value -100 is a special token ID used by HuggingFace’s Transformers library to indicate that a particular token should be ignored when computing the loss.

why not mask = 0 for the indices you want to not train on?

Ok I think this is the code:

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
1 Like

It seems like you know a lot about how this works. So, if setting tokenizer.pad_token = tokenizer.eos_token causes falcon to infinitely generate text up to the cutoff point, how do you stop this from happening? Do you have time to provide a code snippet? All I can think of is:

raw_pad_token = “<pad>”
processed_token = tokenizer(raw_pad_token)
tokenizer.pad_token = processed_token

But based on this thread, this isn’t enough to work

Hi @brando @maxolotl @Rocketknight1
Best way to fix this issue is to change the processing template:

from transformers import AutoTokenizer
from tokenizers.processors import TemplateProcessing

text = "Random text"
tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-7b")

print(tokenizer(text)) # base tokenizer
# {'input_ids': [25070, 2288], 'token_type_ids': [0, 0], 'attention_mask': [1, 1]}

tokenizer._tokenizer.post_processor = TemplateProcessing(
    single="$A " + tokenizer.eos_token,
    pair="$A "+ tokenizer.eos_token +" $B:1 "+ tokenizer.eos_token +":1",
    special_tokens=[(tokenizer.eos_token, tokenizer.eos_token_id)],

print(tokenizer(text)) # Updated tokenizer with EOS token
# {'input_ids': [25070, 2288, 11], 'token_type_ids': [0, 0, 0], 'attention_mask': [1, 1, 1]}

tokenizer.pad_token = tokenizer.eos_token
tokenizer.model_max_length = 5

print(tokenizer(text, padding="max_length")) # Updated tokenizer with EOS token and padding
# {'input_ids': [25070, 2288, 11, 11, 11], 'token_type_ids': [0, 0, 0, 0, 0], 'attention_mask': [1, 1, 1, 0, 0]}

Note that the model has to learn to predict the eos token through causal language modeling.

1 Like