tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

new lr

Browse Source
This commit is contained in:
Mads Henrichsen
2023-12-30 21:36:01 +01:00
parent e3716db386
commit c04df54b4b
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
examples/yayi2-30b/qlora.yml
View File

@@ -3,7 +3,7 @@ model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer tokenizer_type: AutoTokenizer
is_mistral_derived_model: false is_mistral_derived_model: false
trust_remote_code: true trust_remote_code: true
model_revision: refs/pr/5
load_in_8bit: false load_in_8bit: false
load_in_4bit: true load_in_4bit: true
@@ -19,7 +19,7 @@ output_dir: ./qlora-out
adapter: qlora adapter: qlora
lora_model_dir: lora_model_dir:
sequence_len: 8192 sequence_len: 2048 # Fits in 40gb VRAM. Can easily do 4096 in A100 80 or a A6000
sample_packing: false sample_packing: false
pad_to_sequence_len: true pad_to_sequence_len: true
@@ -37,7 +37,7 @@ lora_target_modules:
- k_proj - k_proj
- o_proj - o_proj
wandb_project: wandb_project: yayi2
wandb_entity: wandb_entity:
wandb_watch: wandb_watch:
wandb_name: wandb_name:
@@ -48,7 +48,7 @@ micro_batch_size: 1
num_epochs: 1 num_epochs: 1
optimizer: adamw_bnb_8bit optimizer: adamw_bnb_8bit
lr_scheduler: cosine lr_scheduler: cosine
learning_rate: 0.0002 learning_rate: 0.00002
train_on_inputs: false train_on_inputs: false
group_by_length: false group_by_length: false