GPT-2-from-scratch

Making GPT-2 from scracth with modifications of ROPE, sliding window attention and grouped query attention

Features

• Rotational Positional Embedding: Introduces rotational positional embeddings to improve the model's understanding of token positions.

• Sliding Window Attention: Implements sliding window attention to focus on local contexts, allowing the model to capture more fine-grained information.

• Multi-Query Attention: Enhances attention mechanisms with multi-query attention, enabling the model to consider multiple queries simultaneously.

Training Results

After extensive experimentation, it was observed that a context size of 25 yielded promising results compared to the original architecture. The following steps outline the training process:

1. Data Preparation:

   • For this project, the training data utilized Shakespearean text for tokenization. Ensure your dataset is preprocessed and tokenized using the provided Shakespearean text embeddings or any other suitable embeddings.

2. Model Configuration:

   • Adjust the configuration file to include your modifications (rotational positional embedding, sliding window attention, multi-query attention). Additionally, introduce random model configurations to add an element of variability.

3. Evaluation:

   • Evaluate the model on your test set to assess its performance.

Experimentations

After extensive experimentation, the custom GPT-2 architecture demonstrated notable advantages over the original architecture, especially when utilizing a smaller context window. Here are some key findings:

1. Context Window Optimization:

   • The introduction of sliding window attention, rotational positional embedding, and multi-query attention allowed for effective training with a smaller context window (context size 25). Despite the reduced context, the model exhibited enhanced understanding and capturing of intricate dependencies within the input data.

2. Token Attention Efficiency:

   • The customized attention mechanisms ensured that all tokens within the specified context received attention, leading to a more comprehensive understanding of local contexts. This improvement contributed significantly to the model's ability to capture nuanced relationships and generate more contextually relevant outputs.

3. Reduced Inference Time:

   • The optimized architecture not only excelled in training but also showcased reduced inference time during text generation. The combination of sliding window attention and multi-query attention led to more efficient computations, making the model well-suited for real-time or resource-constrained applications.

4. Performance Outcomes:

   • Comparative evaluations against the original GPT-2 architecture consistently demonstrated superior performance in terms of training convergence, overall model quality, and efficiency in utilizing smaller context windows, a slightly bigger size than original model

Example Usage

Here's a simple example of how to use the trained model for text generation:

#to load the final modified architecture
from gpt2_rope_slide_multiquery import Attention_rope_slide_group
from training_loop import TrainerConfig,Trainer,CharDataset,top_k_logits,sample


# make an instance of model
model = Attention_rope_slide_group()

#loading dataset
block_size=25

text = open('/content/drive/MyDrive/shakespeare.txt', 'r').read()
train_dataset = CharDataset(text, block_size = 25) # 25 is for context storing

#for training on single GPU, multi GPU needs to be tested later
model=model.to(device)
tconf = TrainerConfig(
    max_epochs=1,
    batch_size=8,
    learning_rate=6e-4,
    lr_decay=True,
    warmup_tokens=512,
    final_tokens=2*len(train_dataset)*block_size,
    num_workers=4,
)
trainer = Trainer(model, train_dataset, None, tconf)
trainer.train()


# Generate text
context = "When are you going to change the diaper?"

# Check if each character in the context is in the vocabulary
try:
    indices = [train_dataset.stoi[s] for s in context]
except KeyError as e:
    print(f"Error: Character '{str(e)}' not found in the vocabulary.")
    # Handle the error or skip this context

# Convert the indices to a torch tensor
x = torch.tensor(indices, dtype=torch.long)[None, ...].to(trainer.device)

# Rest of your code...
y = sample(trainable_model, x, 2000, temperature=1.0, sample=True, top_k=10)[0]
completion = ''.join([train_dataset.itos[int(i)] for i in y])
print(completion)