[refactor] graduate custom_config_module and unify args/config naming (pytorch#1871)

In the past, the terms "args" and "config" have been used in a mix.

To make it unambiguous, in torchtitan we use
- "args" as in `ModelArgs` to refer to parameters used to define a model
in model code
- "config" as in `JobConfig` to refer to configurable training job
commands used in training script

This also PR also moves `custom_args_module` (which should be
`custom_config_module` according to the naming rule above) from
`Experimental` to `Job`, as it has been extensively used by various
models in torchtitan, especially those in the `experiments` folder.

Author: tianyu-l

torchtitan/experiments/simple_fsdp/llama3/__init__.py

@@ -10,11 +10,7 @@
 from torchtitan.components.tokenizer import build_hf_tokenizer
 from torchtitan.datasets.hf_datasets import build_hf_dataloader
 from torchtitan.distributed.pipeline_parallel import pipeline_llm
-<<<<<<< HEAD
 from torchtitan.models.llama3 import llama3_args
-=======
-from torchtitan.models.llama3 import llama3_configs
->>>>>>> cd16507b (move PP API to model agnostic file (#1868))
 from torchtitan.protocols.train_spec import TrainSpec
 
 from .model import SimpleFSDPTransformer

torchtitan/models/llama3_ft/__init__.py

@@ -22,7 +22,6 @@
     "Transformer",
     "llama3_configs",
 ]
->>>>>>> cd16507b (move PP API to model agnostic file (#1868))
 
 
 def get_train_spec() -> TrainSpec:

torchtitan/models/llama3_ft/train_configs/debug_model.toml

@@ -1,13 +1,8 @@
 [job]
 dump_folder = "./outputs"
-<<<<<<< HEAD
 description = "Llama 3 fault-tolerant debug training"
 print_config = false
 custom_config_module = "torchtitan.components.ft.config"
-=======
-description = "Llama 3 debug training"
-print_args = false
->>>>>>> f63037ff (add script to train with ft (#1812))
 
 [profiling]
 enable_profiling = true