some status here. This PR does the following:

1. adds all flags outlined in new training library to ilab model train
2. The defaults for these flags come from a new _train class in the config. The train class is composed of the exact classes introduced in the training library, I imported them.
3. I added a --train-profile to ilab init allowing a user to specify a yaml file matching the exact format of the training library args. The two sections in the yaml are train_args and torch_args.
4. I cleaned up mismatched flags and outdated flags necessary for this new library.

the commits might be a bit messy rn. I will clean these up in the morning. Functionally, this is the structure we should go for. Eventually, all of the top level config stuff should live in a profile. The amount of flags being added to training is... quite a lot.

usage:

1. ilab config init --train-profile /path/to/trainargs.yaml
2. ilab train (no flags unless you want to override the yaml. use --legacy to use old linux train code.

a train profile takes the following format:

train_args:
  model_path: instructlab/granite-7b-lab
  data_path: sample-data/train_all_pruned_SDG.jsonl
  ckpt_output_dir: checkpoints
  data_output_dir: output
  max_seq_len: 4096
  max_batch_len: 10000
  num_epochs: 10
  effective_batch_size: 3840
  save_samples: 25000
  learning_rate: 2e-6
  warmup_steps: 800
  is_padding_free: False
  random_seed: 42
torch_args:
  node_rank: 0
  nnodes: 1
  nproc_per_node: 1
  rdzv_id: 123
  rdzv_endpoint: 127.0.0.1:12345

thats it :)

I rebased, and added one more timeout bump.

We're failing one MacOS test. I manually ran the test to see if it's a test problem or a real problem. The manual testing passed.

i.e. if the server is running and we are in a chat window, and we kill the server while chat is emitting, the chat window will exit and let the user know that it's ceasing to emit a response.

I used:
watch -n 1 sudo lsof -iTCP -sTCP:LISTEN -nP

to validate that the serving process that we killed was the one that the chat instance was communicating on.

For future reference, this commit history could have used a bit of squashing cleanup prior to merging.