f7ea140838
* make TiledMLP work with FSDP * cleanup/gc at start of train to prevent large VRAM spike * chore: lint * generic function for non-deepspeed training * unify patch to fix imports * update readme for ALST and add examples * make deepspeed attribute on params check more robust * update with new info from PR review
60 lines
1.4 KiB
YAML
60 lines
1.4 KiB
YAML
base_model: meta-llama/Llama-3.1-8B
|
|
# Automatically upload checkpoint and final model to HF
|
|
# hub_model_id: username/custom_model_name
|
|
|
|
datasets:
|
|
- path: togethercomputer/Long-Data-Collections
|
|
type: completion
|
|
field: text
|
|
data_files:
|
|
- pretrain/rp_sub.jsonl.zst
|
|
- path: princeton-nlp/TextbookChapters
|
|
type: completion
|
|
field: chapter
|
|
dataset_prepared_path: last_run_prepared
|
|
val_set_size: 0.0
|
|
output_dir: ./outputs/out
|
|
|
|
sequence_len: 500_000
|
|
min_sample_len: 200_000
|
|
sample_packing: true
|
|
|
|
tiled_mlp: true
|
|
context_parallel_size: 8
|
|
plugins:
|
|
- axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
|
|
|
|
gradient_accumulation_steps: 1
|
|
micro_batch_size: 1
|
|
num_epochs: 1
|
|
optimizer: adamw_torch_8bit
|
|
lr_scheduler: cosine
|
|
learning_rate: 2e-5
|
|
|
|
bf16: auto
|
|
tf32: true
|
|
|
|
gradient_checkpointing: true
|
|
activation_offloading: legacy
|
|
|
|
resume_from_checkpoint:
|
|
logging_steps: 1
|
|
flash_attention: true
|
|
|
|
warmup_steps: 100
|
|
saves_per_epoch: 1
|
|
evals_per_epoch: 2
|
|
weight_decay: 0.0
|
|
special_tokens:
|
|
pad_token: <|end_of_text|>
|
|
|
|
fsdp_version: 2
|
|
fsdp_config:
|
|
offload_params: false # offloading is currently not compatible with SP + torchao optimizer
|
|
state_dict_type: SHARDED_STATE_DICT
|
|
auto_wrap_policy: TRANSFORMER_BASED_WRAP
|
|
transformer_layer_cls_to_wrap: LlamaDecoderLayer
|
|
reshard_after_forward: true
|
|
|
|
# save_first_step: true # uncomment this to validate checkpoint saving works with your config
|