FIX / Quantizaiton: Revert back previous replacement logic

[younesbelkada]

What does this PR do?

Fixes: huggingface/peft#1720

#29958 introduced a new logic for linear layer replacement for bnb which led to a corner case.

Before that PR, for a BertForSequenceClassification model:

from transformers import AutoModelForSequenceClassification, BitsAndBytesConfig

config = BitsAndBytesConfig(
    load_in_4bit=True, # quantize the model to 4-bits when you load it
    bnb_4bit_quant_type="nf4", # use a special 4-bit data type for weights initialized from a normal distribution
    bnb_4bit_use_double_quant=True, # nested quantization scheme to quantize the already quantized weights
    bnb_4bit_compute_dtype=torch.bfloat16, # use bfloat16 for faster computation,
)

model = AutoModelForSequenceClassification.from_pretrained(
    model_checkpoint,
    quantization_config=config
)

print(model)

Gives:

DistilBertForSequenceClassification(
  (distilbert): DistilBertModel(
    (embeddings): Embeddings(
      (word_embeddings): Embedding(30522, 768, padding_idx=0)
      (position_embeddings): Embedding(512, 768)
      (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
      (dropout): Dropout(p=0.1, inplace=False)
    )
    (transformer): Transformer(
      (layer): ModuleList(
        (0-5): 6 x TransformerBlock(
          (attention): MultiHeadSelfAttention(
            (dropout): Dropout(p=0.1, inplace=False)
            (q_lin): Linear4bit(in_features=768, out_features=768, bias=True)
            (k_lin): Linear4bit(in_features=768, out_features=768, bias=True)
            (v_lin): Linear4bit(in_features=768, out_features=768, bias=True)
            (out_lin): Linear4bit(in_features=768, out_features=768, bias=True)
          )
          (sa_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
          (ffn): FFN(
            (dropout): Dropout(p=0.1, inplace=False)
            (lin1): Linear4bit(in_features=768, out_features=3072, bias=True)
            (lin2): Linear4bit(in_features=3072, out_features=768, bias=True)
            (activation): GELUActivation()
          )
          (output_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
        )
      )
    )
  )
  (pre_classifier): Linear(in_features=768, out_features=768, bias=True)
  (classifier): Linear(in_features=768, out_features=2, bias=True)
  (dropout): Dropout(p=0.2, inplace=False)
)

But after #29958 :

DistilBertForSequenceClassification(
  (distilbert): DistilBertModel(
    (embeddings): Embeddings(
      (word_embeddings): Embedding(30522, 768, padding_idx=0)
      (position_embeddings): Embedding(512, 768)
      (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
      (dropout): Dropout(p=0.1, inplace=False)
    )
    (transformer): Transformer(
      (layer): ModuleList(
        (0-5): 6 x TransformerBlock(
          (attention): MultiHeadSelfAttention(
            (dropout): Dropout(p=0.1, inplace=False)
            (q_lin): Linear4bit(in_features=768, out_features=768, bias=True)
            (k_lin): Linear4bit(in_features=768, out_features=768, bias=True)
            (v_lin): Linear4bit(in_features=768, out_features=768, bias=True)
            (out_lin): Linear4bit(in_features=768, out_features=768, bias=True)
          )
          (sa_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
          (ffn): FFN(
            (dropout): Dropout(p=0.1, inplace=False)
            (lin1): Linear4bit(in_features=768, out_features=3072, bias=True)
            (lin2): Linear4bit(in_features=3072, out_features=768, bias=True)
            (activation): GELUActivation()
          )
          (output_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
        )
      )
    )
  )
  (pre_classifier): Linear4bit(in_features=768, out_features=768, bias=True)
  (classifier): Linear(in_features=768, out_features=2, bias=True)
  (dropout): Dropout(p=0.2, inplace=False)
)

The pre_classifier gets also quantized. In order to preserve BC (even though it might not be 100% what is intended) I propose to add that extra condition and make sure previous behaviour is maintained

cc @SunMarc @BenjaminBossan

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[SunMarc]

Before the PR, pre_classifier was not quantized because we have modules_to_not_convert = ['classifier'] and classifier is in pre_classifier. I'm not sure about the right fix. One solution would be to fix get_keys_to_not_convert in such way that we get modules_to_not_convert = ['classifier', 'pre_classifier'] or we can just pass llm_int8_skip_modules=["classifier", "pre_classifier"]. However, going back to the previous logic is not optimal.

[younesbelkada]

I agree, ideally we don't want to do that .. I also mentioned here: huggingface/peft#1720 (comment) that the user should use modules_to_not_convert, IMO pre_classifier shouldn't be un-quantized at first place .

[BenjaminBossan]

I'm a bit confused, as this does not really revert the other commit, but introduces some extra logic. So I'm not really sure if the exact same keys are matched as before or not :D (are there tests)?

Anyway, do what you think is right :)

[BenjaminBossan]

Why not use current_key_name_str here?

[younesbelkada]

Closing as I agree there is no point reverting a bug!