app.py

import torch
import torch.nn as nn
import yaml
from transformers import GPT2Config, PreTrainedTokenizerFast
from huggingface_hub import HfApi, upload_file
import os
import json
import sentencepiece as spm
from tokenizers import SentencePieceBPETokenizer

# Define the TinyLlama Model
class TinyLlama(nn.Module):
    def __init__(self, config):
        super(TinyLlama, self).__init__()
        self.embedding = nn.Embedding(config.vocab_size, config.n_embd)
        self.transformer_blocks = nn.ModuleList([
            nn.TransformerEncoderLayer(
                d_model=config.n_embd,
                nhead=config.n_head,
                dim_feedforward=config.hidden_dim
            )
            for _ in range(config.n_layer)
        ])
        self.output_layer = nn.Linear(config.n_embd, config.vocab_size)

    def forward(self, x):
        x = self.embedding(x)
        for block in self.transformer_blocks:
            x = block(x)
        x = self.output_layer(x)
        return x

# Load Model Configuration from YAML file
model_config_path = "/home/jax/out/custom-model/final/model_config.yaml"
with open(model_config_path, 'r') as file:
    config_data = yaml.safe_load(file)

# Create Model Configuration
config = GPT2Config(
    vocab_size=config_data.get("vocab_size", 32000),    # Adjust to match TinyLlama's vocab size
    n_embd=config_data.get("n_embd", 2048),             # Embedding size for TinyLlama
    n_layer=config_data.get("n_layer", 24),             # Number of transformer layers
    n_head=config_data.get("n_head", 16),               # Number of attention heads
    hidden_dim=config_data.get("hidden_dim", 8192)      # Feedforward layer dimension
)

# Instantiate the TinyLlama Model
model = TinyLlama(config)

# Load Weights from the .pth file
model_weights_path = "/home/jax/out/custom-model/final/lit_model.pth"
model_weights = torch.load(model_weights_path, map_location=torch.device('cpu'), weights_only=True)
model.load_state_dict(model_weights, strict=False)  # strict=False to allow for minor mismatches

# Directory to Save the Model and Tokenizer for Hugging Face
model_dir = "./huggingface_tinyllama"
os.makedirs(model_dir, exist_ok=True)

# Save the Model Weights
model_weights_save_path = os.path.join(model_dir, "pytorch_model.bin")
torch.save(model.state_dict(), model_weights_save_path)

# Save the Configuration in JSON format
config_save_path = os.path.join(model_dir, "config.json")
with open(config_save_path, 'w') as f:
    json.dump(config.to_dict(), f)

# Load and Convert the SentencePiece model to Hugging Face-compatible format
# Load the SentencePiece model
tokenizer_path = "/home/jax/out/custom-model/final/tokenizer.model"
sp_tokenizer = spm.SentencePieceProcessor()
sp_tokenizer.load(tokenizer_path)

# Create a new tokenizer using the tokenizers library
hf_tokenizer = SentencePieceBPETokenizer(
    vocab=tokenizer_path,  # Load SentencePiece model
    add_prefix_space=True
)

# Train the tokenizer if needed (optional step, depending on your model and usage)
# hf_tokenizer.train([tokenizer_path])

# Set special tokens
hf_tokenizer.add_special_tokens(["<unk>", "<pad>", "<s>", "</s>"])

# Wrap the tokenizer with PreTrainedTokenizerFast for Hugging Face compatibility
hf_tokenizer = PreTrainedTokenizerFast.from_pretrained(tokenizer_path)

# Save the Hugging Face tokenizer
hf_tokenizer.save_pretrained(model_dir)

# Push to Hugging Face Hub
api.create_repo(repo_id=repo_id, token=os.getenv('HUGGINGFACE_API_TOKEN'), exist_ok=True)

# Create a Model Card for the Model
model_card = """
---
language: en
tags:
- tinyllama
- language-model
- chat
license: apache-2.0
---

# TinyLlama 1.1B Chat Model

## Model Description
TinyLlama is a lightweight LLaMA-based model with 1.1 billion parameters, designed to perform well on conversational and text generation tasks. It has been fine-tuned specifically for chat applications, providing coherent and context-aware responses.

## Training Data
The model was trained on a diverse dataset, including web text, books, and conversational data, to make it capable of handling a wide range of language styles.

## Usage
You can use this model for conversational AI, text completion, or other natural language generation tasks. Here’s a quick example:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("jacksonstrut/tinyllama-1.1B-chat")
model = AutoModelForCausalLM.from_pretrained("jacksonstrut/tinyllama-1.1B-chat")

input_ids = tokenizer("Hello, how are you?", return_tensors="pt").input_ids
output = model.generate(input_ids)
print(tokenizer.decode(output[0]))
```

## Limitations
- The model may produce biased or inappropriate outputs as it is trained on general datasets from the internet.
- It may not be suitable for all applications, especially those requiring factual accuracy.

## License
This model is licensed under the Apache 2.0 License.
"""

# Save the Model Card to README.md
readme_path = os.path.join(model_dir, "README.md")
with open(readme_path, 'w') as f:
    f.write(model_card)

# Upload the Model Card to Hugging Face Hub

# Upload files to Hugging Face Hub
upload_file(
    path_or_fileobj=model_weights_save_path,
    path_in_repo="pytorch_model.bin",
    repo_id=repo_id,
    token=os.getenv('HUGGINGFACE_API_TOKEN')
)
upload_file(
    path_or_fileobj=config_save_path,
    path_in_repo="config.json",
    repo_id=repo_id,
    token=os.getenv('HUGGINGFACE_API_TOKEN')
)

# Upload the tokenizer files
upload_file(
    path_or_fileobj=os.path.join(model_dir, "tokenizer.model"),
    path_in_repo="tokenizer.model",
    repo_id=repo_id,
    token=os.getenv('HUGGINGFACE_API_TOKEN')
)
upload_file(
    path_or_fileobj=os.path.join(model_dir, "tokenizer_config.json"),
    path_in_repo="tokenizer_config.json",
    repo_id=repo_id,
    token=os.getenv('HUGGINGFACE_API_TOKEN')
)