Skip to content

Instantly share code, notes, and snippets.

@jphme
Last active September 12, 2023 13:33
  • Star 0 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
Star You must be signed in to star a gist
Save jphme/d9e09b9d285e2ec03b21c81a98f8dc37 to your computer and use it in GitHub Desktop.
Phi 1.5 multi-gpu config (Axolotl)
#only changed self.vocab_size = vocab_size in configuration_mixformer_sequential.py, as suggested by @NanoBit
base_model: /workspace/models/phi-1_5
base_model_config: /workspace/models/phi-1_5
trust_remote_code: true
base_model_ignore_patterns:
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
is_llama_derived_model: false
load_in_8bit: false
load_in_4bit: false
strict: false
datasets:
- path: ehartford/wizard_vicuna_70k_unfiltered
type: sharegpt:chat
dataset_prepared_path: /workspace/tmp/last_run_prepared
val_set_size: 0.03
output_dir: /workspace/models/phi_wizard_vicuna_uncensored
resume_from_checkpoint:
adapter:
lora_model_dir:
sequence_len: 2048
sample_packing: true
lora_r:
lora_alpha:
lora_dropout:
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:
relora_steps:
relora_warmup_steps:
relora_cpu_offload:
wandb_project: phi_wizard_vicuna
wandb_entity:
wandb_watch:
wandb_run_id: test2
wandb_log_model:
gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 3
optimizer: adamw_bnb_8bit
#paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.00005
train_on_inputs: true
group_by_length: false
bf16: full
fp16: false
tf32: false
gradient_checkpointing: false
early_stopping_patience:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
warmup_steps: 10
eval_steps: 125
save_steps: 250
save_total_limit: 3
debug:
#standard zero2.json from axolotl/deepspeed without any changes
deepspeed: zero2.json
weight_decay:
fsdp:
fsdp_config:
special_tokens:
pad_token: "<|endoftext|>"
#500 schon done
resize_token_embeddings_to_32x: true
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment