Can't finetune Mixtral model

hahmad2008 commented 7 months ago

Please check that this issue hasn't been reported before.

[X] I searched previous Bug Reports didn't find any similar reports.

Expected Behavior

I am trying to finetune mistralai/Mixtral-8x7B-v0.1 using the examples in the repo, but I got CUDA out of memory.

I am using 4 A10 GPUs each with 20G memory. I am finetuning using qlora with deepspeed zero2

Current behaviour

Out of memory error while loading the model on GPUs.

Command: accelerate launch scripts/finetune.py config.yaml --deepspeed deepspeed_configs/zero2.json

Steps to reproduce

Command: accelerate launch scripts/finetune.py config.yaml --deepspeed deepspeed_configs/zero2.json

using this commit: c67fb7158312e47e3326f077f74485cf0a23b51a

Config yaml

base_model: mistralai/Mixtral-8x7B-v0.1
model_type: AutoModelForCausalLM
tokenizer_type: LlamaTokenizer
trust_remote_code: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: mhenrichsen/alpaca_2k_test
    type: alpaca
dataset_prepared_path: prepared-dataset
val_set_size: 0.0
output_dir: finetuned-model

## You can optionally freeze the entire model and unfreeze a subset of parameters
unfrozen_parameters:
#  - lm_head.*
#  - model.embed_tokens.*
#  - model.layers.2[0-9]+.block_sparse_moe.gate.*
#  - model.layers.2[0-9]+.block_sparse_moe.experts.*
#  - model.layers.3[0-9]+.block_sparse_moe.gate.*
#  - model.layers.3[0-9]+.block_sparse_moe.experts.*

model_config:
  output_router_logits: true

adapter: qlora
lora_model_dir:

sequence_len: 256
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
#lora_target_modules:
#  - gate
#  - q_proj
#  - k_proj
#  - v_proj
#  - o_proj
#  - w1
#  - w2
#  - w3

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_table_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero2.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

Possible solution

No response

Which Operating Systems are you using?

[X] Linux
[ ] macOS
[ ] Windows

Python Version

3.10

axolotl branch-commit

c67fb7158312e47e3326f077f74485cf0a23b51a

Acknowledgements

[X] My issue title is concise, descriptive, and in title casing.
[X] I have searched the existing issues to make sure this bug has not been reported yet.
[X] I am using the latest version of axolotl.
[X] I have provided enough information for the maintainers to reproduce and diagnose the issue.

winglian commented 7 months ago

You need to use fsdp+qlora for it to work across gpus. Otherwise it tries to load the entire model on each gpu.

hahmad2008 commented 7 months ago

@winglian , I used FSDP with qlora and the model still loaded as copied to the GPUs.

I tried it with passing accelerate config and without and having the same behavior:

accelerate launch --config_file acc-config.yaml scripts/finetune.py config.yaml
accelerate launch scripts/finetune.py config.yaml

acc-config.yaml

compute_environment: LOCAL_MACHINE
distributed_type: FSDP
downcast_bf16: 'no'
gpu_ids: all
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

config.yaml

base_model: mistralai/Mixtral-8x7B-v0.1
model_type: AutoModelForCausalLM
tokenizer_type: LlamaTokenizer
trust_remote_code: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: mhenrichsen/alpaca_2k_test
    type: alpaca
dataset_prepared_path: prepared-dataset
val_set_size: 0.0
output_dir: finetuned-model

## You can optionally freeze the entire model and unfreeze a subset of parameters
unfrozen_parameters:
#  - lm_head.*
#  - model.embed_tokens.*
#  - model.layers.2[0-9]+.block_sparse_moe.gate.*
#  - model.layers.2[0-9]+.block_sparse_moe.experts.*
#  - model.layers.3[0-9]+.block_sparse_moe.gate.*
#  - model.layers.3[0-9]+.block_sparse_moe.experts.*

model_config:
  output_router_logits: true

adapter: qlora
lora_model_dir:

sequence_len: 256
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
#lora_target_modules:
#  - gate
#  - q_proj
#  - k_proj
#  - v_proj
#  - o_proj
#  - w1
#  - w2
#  - w3

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false

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

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_table_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_offload_params: false
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
special_tokens:

hahmad2008 commented 7 months ago

seems I need to enable this fsdp_offload_params: true

axolotl-ai-cloud / axolotl