feat: Add ALCF/examples/finetune_llama3/*

[saforem2]

Copilot Generated Summary:

This pull request includes several updates to the ALCF/examples/finetune_llama3 directory, focusing on adding a new README file, configuration files, and a shell script for fine-tuning the Llama3 model. Additionally, there is a minor update in the megatron/checkpointing.py file to ensure directory creation when saving the learning rate state dictionary.

Documentation and setup:

• ALCF/examples/finetune_llama3/README.md: Added comprehensive instructions for setting up the environment, installing dependencies, downloading data, and converting Hugging Face checkpoints for fine-tuning Llama3.

Configuration files:

• ALCF/examples/finetune_llama3/ds_config.json: Introduced a new DeepSpeed configuration file with parameters for batch size, optimizer, and bf16 support.
• ALCF/examples/finetune_llama3/ds_config_empty.json: Added an empty DeepSpeed configuration file template with similar parameters as ds_config.json.

Shell script:

• ALCF/examples/finetune_llama3/finetune_llama.sh: Added a script for setting up the environment, configuring DeepSpeed, and running the fine-tuning process for Llama3. This script includes logic for converting Hugging Face models to Megatron-Deepspeed format and vice versa.

Minor update:

• megatron/checkpointing.py: Ensured the parent directory is created if it does not exist when saving the learning rate state dictionary.

[saforem2]

There was an issue being caused by the following block:

for name, param in hf_auto_model.named_parameters():

which was failing to capture the hf_auto_model.lm_head.weight, which was preventing the checkpoint from being converted successfully.

Replacing this block with

>>> for name, submodule in lmodel.named_children():
...     for pname, param in submodule.named_parameters():
...         named_submods[f'{name}.{pname}'] = param.clone()

fixes this issue, as shown explicitly below:

>>> from transformers import AutoModelForCausalLM, LlamaConfig, AutoTokenizer, LlamaForCausalLM, AutoConfig
>>> lmodel = AutoModelForCausalLM.from_pretrained('meta-llama/Llama-3.2-1B')
>>> named_params = []
>>> named_params = {}
>>> for name, param in lmodel.named_parameters():
...     named_params[name] = param.clone()

>>> named_submods = {}
>>> for name, submodule in lmodel.named_children():
...     for pname, param in submodule.named_parameters():
...         named_submods[f'{name}.{pname}'] = param.clone()

>>> len(named_submods.keys())
147

>>> len(named_params.keys())
146

>>> list(named_submods.keys())[-3:]
['model.layers.15.post_attention_layernorm.weight',
 'model.norm.weight',
 'lm_head.weight']

>>> list(named_params.keys())[-3:]
['model.layers.15.input_layernorm.weight',
 'model.layers.15.post_attention_layernorm.weight',
 'model.norm.weight']

[saforem2]

Mismatch between self.tokenizer.vocab_size hf_w.shape[0] when using Llama3 tokenizers.

This discrepancy causes the following assertion:

assert hf_w.shape[0] == self.padded_vocab_size

to fail since hf_w.shape[0] = 128256 but self.tokenizer.vocab_size = 128000.

Explicitly:

>>> type(self.hf_tokenizer)
<class 'transformers.tokenization_utils_fast.PreTrainedTokenizerFast'>
>>> type(self.tokenizer)
<class 'megatron.tokenizer.tokenizer._HFTokenizer'>
>>> self.tokenizer.vocab_size
128000
>>> len(self.hf_tokenizer)
128256

so, replacing self.token_vocab = len(self.hf_tokenizer) seems to resolve this issue.