Fix Sparsity Logs on FSDP Model Save

[Satrat]

Model sparsity was not being logged correctly on FSDP save. This was because when we save an FSDP model we gather it onto the CPU in the main process and call save_pretrained on the wrapped model.

    with FullyShardedDataParallel.state_dict_type(
        model, StateDictType.FULL_STATE_DICT, full_state_dict_config
    ):
        state_dict = accelerator.get_state_dict(model, unwrap=False)

    if accelerator.is_main_process:
        accelerator.unwrap_model(model).save_pretrained(
            output_dir,
            is_main_process=accelerator.is_main_process,
            save_function=accelerator.save,
            state_dict=state_dict

This results in the wrapped model being passed to the global sparsity function with incomplete information. We need the full state dict as well to properly calcualte the sparsity. I also add some other log cleanup so we aren't printing out as much sparsity information between train stages

Example

Launch with FSDP: accelerate launch --config_file integrations/huggingface-transformers/finetuning/example_fsdp_config.yaml test.py

from sparseml.transformers.finetune.text_generation import train

model_name = "neuralmagic/TinyLlama-1.1B-Chat-v1.0-pruned2.4"
output_dir = "llama-1.1b_compressed_test_out"
recipe_stub = """
test_stage:
  pruning_modifiers:
    ConstantPruningModifier:
      targets: ['re:.*q_proj.weight', 're:.*k_proj.weight', 're:.*v_proj.weight', 're:.*o_proj.weight',
        're:.*gate_proj.weight', 're:.*up_proj.weight', 're:.*down_proj.weight']
      start: 0
"""

train(
  save_compressed=True,
  model=model_name,
  recipe=recipe_stub,
  dataset="open_platypus",
  max_steps=5,
  output_dir=output_dir,
  overwrite_output_dir=True,
  splits = "train",
  logging_steps = 25,
  precision="bfloat16",
  gradient_checkpointing = True,
  bf16 = True,
)

Output:
All the parameters get picked up, previously only the 23 wrapped transformers layers were showing up as parameters

2024-03-28 21:18:09 sparseml.transformers.compression.utils.compress_save INFO     Inferring a sparsity configuration requires a global sparsity calculation. This can be costly for large models. To skip the calculation of compression statistics set skip_compression_stats=True
Calculating model sparsity: 100%|███████████████████████████████████████████████████████████████████████████████████████████████| 201/201 [00:04<00:00, 48.54it/s]
Compressing model: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████| 201/201 [00:16<00:00, 12.24it/s]