utensil commited on
Commit
fb3d40f
1 Parent(s): 72bf8aa

falcon + qlora + xformer mbs 40 gas 2 on A6000

Browse files
examples/falcon/config-7b-qlora.yml CHANGED
@@ -18,7 +18,7 @@ val_set_size: 0.01
18
  adapter: qlora
19
  lora_model_dir:
20
  sequence_len: 2048
21
- max_packed_sequence_len: 2048
22
  lora_r: 64
23
  lora_alpha: 16
24
  lora_dropout: 0.05
@@ -30,8 +30,8 @@ wandb_watch:
30
  wandb_run_id:
31
  wandb_log_model:
32
  output_dir: ./qlora-out
33
- batch_size: 8
34
- micro_batch_size: 4
35
  num_epochs: 3
36
  optimizer: paged_adamw_32bit
37
  torchdistx_path:
@@ -50,7 +50,7 @@ resume_from_checkpoint:
50
  auto_resume_from_checkpoints: true
51
  local_rank:
52
  logging_steps: 1
53
- xformers_attention: false
54
  flash_attention:
55
  gptq_groupsize:
56
  gptq_model_v1:
 
18
  adapter: qlora
19
  lora_model_dir:
20
  sequence_len: 2048
21
+ max_packed_sequence_len:
22
  lora_r: 64
23
  lora_alpha: 16
24
  lora_dropout: 0.05
 
30
  wandb_run_id:
31
  wandb_log_model:
32
  output_dir: ./qlora-out
33
+ micro_batch_size: 40
34
+ gradient_accumulation_steps: 2
35
  num_epochs: 3
36
  optimizer: paged_adamw_32bit
37
  torchdistx_path:
 
50
  auto_resume_from_checkpoints: true
51
  local_rank:
52
  logging_steps: 1
53
+ xformers_attention: true
54
  flash_attention:
55
  gptq_groupsize:
56
  gptq_model_v1: