use pythia-12b, neox-20b is flaky

examples/pythia-12b/README.md

@@ -0,0 +1,10 @@
+# Python 12B
+
+- Single-GPU A100 only (?)
+
+```shell
+python scripts/finetune.py examples/pythia-12b/config.yml
+```
+
+⚠️ Multiple-GPU A100 - Doesn't seem to work with multi-gpu without causing OOM! ⚠️
+

configs/gpt_neox_20b.yml → examples/pythia-12b/config.yml

@@ -1,11 +1,12 @@
-base_model: EleutherAI/gpt-neox-20b
-base_model_config: EleutherAI/gpt-neox-20b
+base_model: EleutherAI/pythia-12b-deduped
+base_model_config: EleutherAI/pythia-12b-deduped
 base_model_ignore_patterns: pytorch*  # prefer safetensors
 model_type: GPTNeoXForCausalLM
 tokenizer_type: AutoTokenizer
 load_in_8bit: false
-load_in_4bit: true
-load_4bit: false
+load_in_4bit: false
+gptq: false
+device_map: auto
 datasets:
   - path: vicgalle/alpaca-gpt4
     type: alpaca
@@ -21,16 +22,16 @@ lora_dropout: 0.0
 lora_target_modules:
 lora_target_linear: true
 lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
-wandb_project: gpt4all-neox-20b
+wandb_project: pythia-12b
 wandb_watch:
 wandb_run_id:
 wandb_log_model:
-output_dir: ./gpt4all-neox-20b
+output_dir: ./pythia-12b
 gradient_accumulation_steps: 1
-micro_batch_size: 2
+micro_batch_size: 1
 num_epochs: 5
 learning_rate: 0.00003
-optimizer: paged_adamw_32bit
+optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 train_on_inputs: false
 group_by_length: false
@@ -43,3 +44,6 @@ early_stopping_patience:
 resume_from_checkpoint:
 local_rank:
 gradient_checkpointing: true
+fsdp:
+fsdp_transformer_layer_cls_to_wrap:
+collator_pad_to_longest: true