CLI parser throws an error when providing a config via yaml flag

[ngoncharenko]

Describe the bug

CLI parser throws an error when providing a config via yaml flag.

Steps/Code to reproduce bug
App version: commit 01a9a8b

Invocation:

torchrun --standalone --nnodes=1 --nproc_per_node=1 recipe.py --factory 'my_simple_recipe(model_name="meta/llama-3.2-1b-instruct")' --yes --yaml config.yaml -v

The my_simple_recipe recipe is based on /opt/NeMo/nemo/collections/llm/recipes/llama32_1b.py

config.yaml

data:
  dataset_root: /workspace/datasets/test_dataset
  seq_length: 4096
  global_batch_size: 1
  micro_batch_size: 1
  dataset_kwargs:
    prompt_template: '{prompt} {completion}'
    truncation_field: prompt
    label_key: completion
    add_bos: false
    add_eos: true
trainer:
  accelerator: gpu
  max_epochs: 1
  limit_val_batches: 1.0
  log_every_n_steps: 10
  max_steps: 32
  strategy:
    tensor_model_parallel_size: 1
    pipeline_model_parallel_size: 1
    context_parallel_size: 1
    ckpt_async_save: false
  val_check_interval: 1
log:
  ckpt:
    save_last: link
    save_top_k: 1
    train_time_interval: null
    every_n_train_steps: 2
    filename: '{model_name}--{val_loss:.4f}-{step}-{consumed_samples}'
resume:
  restore_config:
    path: /mount/models/llama32_1b-instruct_2_0
peft:
  dim: 32
  alpha: 16
model:
  config:
    seq_length: 4096

Error:

nemo_run.cli.cli_parser.UnknownTypeError: Failed to parse 'Error parsing argument: Failed to parse '32' as int: Unsupported type: int (Argument: 32, Context: {'expected_type': 'int'})' as peft.dim=32: {'key': 'peft.dim', 'value': '32', 'expected_type': 'int'} (Argument: Error parsing argument: Failed to parse '32' as int: Unsupported type: int (Argument: 32, Context: {'expected_type': 'int'}), Context: {'expected_type': 'peft.dim=32'})

Expected behavior

Should be able to parse config.yaml correctly without errors

Additional context

Possible cause: _resolve_type_checking_annotation() does not resolve the type correctly returning 'int' back instead of int as a type.

[hemildesai]

@marcromeyn Can you take a look?