### Install BLT with uv for Reproducibility

Source: https://github.com/facebookresearch/blt/blob/main/README.md

This experimental method uses 'uv' for reproducible environment setup. It installs dependencies and prepares the environment for running BLT scripts.

```bash
uv pip install --group pre_build --no-build-isolation
uv pip install --group compile_xformers --no-build-isolation
uv sync
uv run python download_blt_weights.py
uv run python demo.py "A BLT has"
```

--------------------------------

### Install BLT with Conda and Pip

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Use this method to set up the BLT environment using conda for environment management and pip for package installation. Ensure you have the correct CUDA version for PyTorch.

```bash
git clone https://github.com/facebookresearch/blt
cd blt
conda create -n blt python=3.12
conda activate blt
pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu121
pip install ninja
pip install -v -U git+https://github.com/facebookresearch/xformers.git@de742ec3d64bd83b1184cc043e541f15d270c148
pip install -r requirements.txt
```

--------------------------------

### Generate Text with BLT Demo Script

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Execute the demo script to generate text using the BLT model. This script also serves as an example for loading BLT in custom code.

```bash
python demo.py "A BLT has"
```

--------------------------------

### Build BLT Environment with Slurm

Source: https://github.com/facebookresearch/blt/blob/main/README.md

This option uses a Slurm job script to create the BLT environment. It's suitable for clusters with Slurm workload manager. Activate the environment using the provided command after setup.

```bash
git clone https://github.com/facebookresearch/blt
cd blt
bash setup/create_env.sh
# or if you have access to a SLURM cluster
sbatch setup/create_env.sh
```

--------------------------------

### Activate Conda Environment after Slurm Setup

Source: https://github.com/facebookresearch/blt/blob/main/README.md

After setting up the environment using the Slurm job script, activate the created conda environment using this command. The environment name includes a date.

```bash
conda activate blt_<date>
```

--------------------------------

### Launch Debug Training Job (Single GPU)

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Launch a debug training job on a single GPU using the bytelatent.train module.

```bash
python -m bytelatent.train  config=bytelatent/configs/debug.yaml
```

--------------------------------

### Download and Prepare Hugging Face Data

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Use this script to download and prepare data from Hugging Face. Specify the dataset, data directory, seed, and number of chunks. Adjust nchunks based on the number of GPUs available.

```bash
python setup/download_prepare_hf_data.py fineweb_edu <MEMORY> --data_dir ./data --seed 42 --nchunks <NCHUNKS>
```

--------------------------------

### Launch Debug Training Job (SLURM)

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Launch a debug training job using the SLURM tool. Adapt configurations like dump_dir, data.root_dir, and data.tokenizer.path.

```bash
# stool stands for SLURM tool !
python -m bytelatent.stool script=bytelatent.train config=bytelatent/configs/debug.yaml nodes=1 partition=<partition>
```

--------------------------------

### Launch Debug Training Job (torchrun)

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Launch a debug training job locally using torchrun, specifying the number of processes per node and the configuration file.

```bash
torchrun --nproc-per-node 8 -m bytelatent.train config=bytelatent/configs/debug.yaml
```

--------------------------------

### Download Tokenizer

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Download a tokenizer, such as Llama3, to a specified save path using your Hugging Face API key.

```bash
python setup/download_tokenizer.py llama3 <SAVE_PATH> --api_key <HUGGINGFACE_TOKEN>
```

--------------------------------

### Load BLT Model and Tokenizer in Python

Source: https://github.com/facebookresearch/blt/blob/main/README.md

This Python code demonstrates how to load the BLT model, tokenizer, and patcher components from Hugging Face for custom use. It initializes the necessary classes for text generation.

```python
from bytelatent.transformer import LMTransformer
from bytelatent.model.blt import ByteLatentTransformer
from bytelatent.hf import BltTokenizerAndPatcher

entropy_repo = "facebook/blt-entropy"
blt_repo = "facebook/blt-1b"
entropy_model = LMTransformer.from_pretrained(entropy_repo)
blt_model = ByteLatentTransformer.from_pretrained(blt_repo)
tok_and_patcher = BltTokenizerAndPatcher.from_pretrained(blt_repo)
tokenizer = tok_and_patcher.tokenizer_args.build()
patcher = tok_and_patcher.patcher_args.build()
```

--------------------------------

### Download BLT Model Weights

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Run this Python script to download the pre-trained BLT model weights. The weights will be saved to the 'hf-weights' directory.

```bash
python download_blt_weights.py
```

--------------------------------

### Load BLT Weights via Hugging Face CLI

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Use the Hugging Face CLI to load BLT model weights and generate text from a prompt. Ensure you have requested access to the weights on the Hugging Face model page.

```bash
python -m bytelatent.hf load-transformers --entropy-repo facebook/blt-entropy --blt-repo facebook/blt-1b --prompt "My test prompt" hub
```

--------------------------------

### Lint Project Code

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Run the linting script to check for code style and quality issues.

```bash
bash dev/lint.sh
```

--------------------------------

### Relaunch Crashed SLURM Job

Source: https://github.com/facebookresearch/blt/blob/main/README.md

If a SLURM job crashes, it can be relaunched using the sbatch command with the path to the dump directory's submit script.

```bash
sbatch path/to/dump_dir/submit.slurm
```

--------------------------------

### BLT Paper Citation

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Citation for the Byte Latent Transformer paper.

```bibtex
@article{meta_blt,
  author = {Artidoro Pagnoni, Ram Pasunuru, Pedro Rodriguez, John Nguyen, Benjamin Muller, Margaret Li, Chunting Zhou, Lili Yu, Jason Weston, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Ari Holtzman†, Srinivasan Iyer},
  title = {Byte Latent Transformer: Patches Scale Better Than Tokens},
  url = {https://github.com/facebookresearch/blt},
  year = {2024}
}
```

--------------------------------

### Lingua Code Citation

Source: https://github.com/facebookresearch/blt/blob/main/README.md

Citation for the Meta Lingua code library.

```bibtex
@misc{meta_lingua,
  author = {Mathurin Videau, Badr Youbi Idrissi, Daniel Haziza, Luca Wehrstedt, Jade Copet, Olivier Teytaud, David Lopez-Paz},
  title = {{Meta Lingua}: A minimal {PyTorch LLM} training library},
  url = {https://github.com/facebookresearch/lingua},
  year = {2024}
}
```

=== COMPLETE CONTENT === This response contains all available snippets from this library. No additional content exists. Do not make further requests.