Minimum training example for Qwen 1.5 72B about pai-megatron-patch

[Peter-Devine]

Yeah, no worries, I thought I'd just add my example for future reference by anyone that stumbles across this issue :)

[lwmlyy]

Hi, thanks for your interest.

For Qwen1.5 performance, please refer to this report
.

Please refer to convert for model converting and train for model training.

For finetuning Qwen-72b, you need 4 nodes of A100/A800-80G*8.

[Peter-Devine]

Hi @lwmlyy , thanks for the response.

Just in case anyone is interested, I was able to train Qwen 1.5 72B GPTQ 4bit with a LoRA adapter using this config with Axolotl using only 1 A100 (80GB) card.

base_model: Qwen/Qwen1.5-72B-Chat-GPTQ-Int4
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
gptq: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: ABCDEFG.json
    ds_type: json # see other options below
    type: sharegpt
    conversation: chatml
dataset_prepared_path: /workspace/ABCDEFG
val_set_size: 0.0002
output_dir: /workspace/ABCDEFG

sequence_len: 6000 
sample_packing: true
pad_to_sequence_len:

adapter: lora
lora_model_dir:
lora_r: 64
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

use_wandb: true
wandb_project: ABCDEFG
wandb_entity: ABCDEFG
wandb_name: ABCDEFG

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.00005

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention:

warmup_steps: 10
evals_per_epoch: 40
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 40
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

[lwmlyy]

Hi, the resource requirement I mentioned is for full-parameter training.

[jerryli1981]

您好，qwen1.5的32和72b现在都已经正常支持了，可以pull下最新的代码再试试