ggml icon indicating copy to clipboard operation
ggml copied to clipboard

Published 20 hours ago • verified publisher icon ggerganov

Add example for text-to-text transfer transformer (T5) inference

Open ggerganov opened this issue 1 year ago • 5 comments

WIP in progress

model: https://github.com/huggingface/transformers/blob/main/src/transformers/models/t5/modeling_t5.py

tokenizer: https://github.com/huggingface/transformers/blob/main/examples/flax/language-modeling/t5_tokenizer_model.py

gated-gelu: https://arxiv.org/pdf/2002.05202.pdf

google t5x: https://github.com/google-research/t5x / models: https://github.com/google-research/t5x/blob/main/docs/models.md

FLAN-T5 Small layers 
T5ForConditionalGeneration(
  (shared): Embedding(32128, 512)
  (encoder): T5Stack(
    (embed_tokens): Embedding(32128, 512)
    (block): ModuleList(
      (0): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
              (relative_attention_bias): Embedding(32, 6)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (1): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (2): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (3): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (4): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (5): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (6): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (7): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
    )
    (final_layer_norm): T5LayerNorm()
    (dropout): Dropout(p=0.1, inplace=False)
  )
  (decoder): T5Stack(
    (embed_tokens): Embedding(32128, 512)
    (block): ModuleList(
      (0): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
              (relative_attention_bias): Embedding(32, 6)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerCrossAttention(
            (EncDecAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (2): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (1): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerCrossAttention(
            (EncDecAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (2): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (2): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerCrossAttention(
            (EncDecAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (2): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (3): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerCrossAttention(
            (EncDecAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (2): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (4): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerCrossAttention(
            (EncDecAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (2): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (5): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerCrossAttention(
            (EncDecAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (2): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (6): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerCrossAttention(
            (EncDecAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (2): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (7): T5Block(
        (layer): ModuleList(
          (0): T5LayerSelfAttention(
            (SelfAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (1): T5LayerCrossAttention(
            (EncDecAttention): T5Attention(
              (q): Linear(in_features=512, out_features=384, bias=False)
              (k): Linear(in_features=512, out_features=384, bias=False)
              (v): Linear(in_features=512, out_features=384, bias=False)
              (o): Linear(in_features=384, out_features=512, bias=False)
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (2): T5LayerFF(
            (DenseReluDense): T5DenseGatedActDense(
              (wi_0): Linear(in_features=512, out_features=1024, bias=False)
              (wi_1): Linear(in_features=512, out_features=1024, bias=False)
              (wo): Linear(in_features=1024, out_features=512, bias=False)
              (dropout): Dropout(p=0.1, inplace=False)
              (act): NewGELUActivation()
            )
            (layer_norm): T5LayerNorm()
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
    )
    (final_layer_norm): T5LayerNorm()
    (dropout): Dropout(p=0.1, inplace=False)
  )
  (lm_head): Linear(in_features=512, out_features=32128, bias=False)
)

ggerganov avatar Dec 31 '22 11:12 ggerganov

Where does this stand? Interested to give the LaMini-Flan-T5 models a go as their paper shows promising results of the T5 models particularly for smaller model sizes.

tkon99 avatar May 10 '23 19:05 tkon99

This fell out of priority - feel free to give it a try

ggerganov avatar May 10 '23 22:05 ggerganov

Does the draft implementation work?

redthing1 avatar May 20 '23 20:05 redthing1

So I tested the t5 branch, it does indeed load the model, but there's no code for actual inference right now. And honestly it looks pretty complicated to implement.

redthing1 avatar May 21 '23 18:05 redthing1

The new upcoming StableDiffusion version 3 uses, beside others, T5.

https://stability.ai/news/stable-diffusion-3-research-paper

Green-Sky avatar Mar 05 '24 13:03 Green-Sky