dcbark01

#!/bin/bash
#SBATCH --job-name=llm-swarm
#SBATCH --partition hopper-prod
#SBATCH --gpus={{gpus}}
#SBATCH --cpus-per-task=12
#SBATCH --mem-per-cpu=11G
#SBATCH -o slurm/logs/%x_%j.out

# See original source here:
# https://github.com/huggingface/llm-swarm/blob/main/templates/tgi_h100.template.slurm

dcbark01

"""
# Source: https://gist.github.com/jneuff/682d47b786329f19291d166957b3274a

/// Fix a huggingface tokenizer to which tokens have been added after training.
///
/// Adding tokens after training via `add_special_tokens` leads to them being added to the
/// `added_tokens` section but not to the `model.vocab` section. This yields warnings like:
/// ```
/// [2023-10-17T07:54:05Z WARN  tokenizers::tokenizer::serialization] Warning: Token '<|empty_usable_token_space_1023|>' was expected to have ID '129023' but was given ID 'None'
/// ```