support chatglm2&3

[Agoniii]

What does this PR do ?

Support ChatGLM2&ChatGLM3

Collection: [Note which collection this PR will affect]

Changelog

• Add specific line by line info of high level changes in this PR.

Usage

• You can potentially add a usage example below

Model conversion from huggingface to nemo

python ./scripts/nlp_language_modeling/convert_hf_chatglm_to_nemo.py --in-file '/mount/data/chatglm3-6b' --out-file '/mount/nemo_models/chatglm3.nemo'

Inference

python ./examples/nlp/language_modeling/megatron_gpt_eval.py \
    --config-name=megatron_chatglm_inference \
    gpt_model_file=${CHATGLM_NEMO_MODEL} \
    trainer.devices=1 \
    trainer.num_nodes=1 \
    tensor_model_parallel_size=1 \
    pipeline_model_parallel_size=1 

Pretrain

 python3 -u examples/nlp/language_modeling/megatron_gpt_pretraining.py \
  --config-name=megatron_chatglm_config \
  trainer.devices=8 \
  trainer.num_nodes=1 \
  trainer.val_check_interval=100 \
  trainer.precision=bf16 \
 +model.data.data_prefix=${TRAIN_DS} \
  trainer.max_steps=1000 \
  trainer.max_epochs=1 \
  model.megatron_amp_O2=True \
  model.sequence_parallel=True \
  model.tensor_model_parallel_size=2 \
  model.pipeline_model_parallel_size=1 \
  model.global_batch_size=128 \
  model.micro_batch_size=1 \
  model.override_vocab_size=65024 \
  model.use_cpu_initialization=False

SFT

python ./examples/nlp/language_modeling/tuning/megatron_gpt_sft.py \
    trainer.devices=8 \
    trainer.val_check_interval=200 \
    model.restore_from_path=${CHATGLM_NEMO_MODEL} \
    model.tensor_model_parallel_size=${TP_SIZE} \
    model.pipeline_model_parallel_size=${PP_SIZE} \
    model.answer_only_loss=True \
    model.global_batch_size=128 \
    model.micro_batch_size=1 \
    model.data.train_ds.file_names=${TRAIN_DS} \
    model.data.train_ds.concat_sampling_probabilities=[1.0] \
    model.data.validation_ds.file_names=${VALID_DS} \
    model.data.validation_ds.global_batch_size=128\
    model.data.validation_ds.micro_batch_size=1 \
    model.megatron_amp_O2=True \
    exp_manager.exp_dir=exp \
    exp_manager.name=chatglm_test \
    trainer.precision=bf16

PEFT

./examples/nlp/language_modeling/tuning/megatron_gpt_peft_tuning.py \
  trainer.precision=bf16 \
  trainer.devices=8 \
  trainer.num_nodes=1 \
  trainer.max_steps=200 \
  trainer.val_check_interval=20 \
  ++trainer.limit_val_batches=10 \
  trainer.gradient_clip_val=1.0 \
  model.megatron_amp_O2=True \
  ++model.mcore_gpt=True \
  model.restore_from_path=${nemo_model} \
  model.peft.peft_scheme=${scheme} \
  model.data.train_ds.file_names=["${TRAIN_DS}"] \
  model.data.train_ds.concat_sampling_probabilities=[1.0] \
  model.data.train_ds.label_key='output' \
  model.data.train_ds.num_workers=0 \
  model.data.validation_ds.file_names=["${VALID_DS}"] \
  model.data.validation_ds.label_key='output' \
  model.data.validation_ds.num_workers=0 \
  model.tensor_model_parallel_size=${TP_SIZE} \
  model.pipeline_model_parallel_size=${PP_SIZE} \
  model.global_batch_size=${GBS} \
  model.micro_batch_size=1

Jenkins CI

To run Jenkins, a NeMo User with write access must comment jenkins on the PR.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to # (issue)

[yaoyu-33]

Can you follow this PR to update your argument names? https://github.com/NVIDIA/NeMo/pull/8435/files#diff-3c80bf9f00be20ecece478699ebbd05ddf7a5a88646ef056b540be5f33622ada

[yaoyu-33]

Also rename to convert_chatglm_hf_to_nemo (or nemo_to_hf)