ML Flow Checkpointing

[wanderingweights]

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 have the below config and I was hoping to have the model checkpoints saving as artifacts but I only get the metrics and config saving.

Is this expected to work or am I missing something?

Thanks for looking over!

Current behaviour

No model checkpoints.

Steps to reproduce

I use the docker image:

"winglian/axolotl:main-latest"

hf login

then

accelerate launch -m axolotl.cli.train theconfig.yml

Config yaml

base_model: NousResearch/Meta-Llama-3-8B-Instruct
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: true
load_in_4bit: false
strict: false

chat_template: llama3
datasets:
  - path: base.yml
    type: alpaca
    ds_type: json

dataset_prepared_path:
val_set_size: 0.05
output_dir: attempt_1

sequence_len: 4096
sample_packing: false
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

mlflow_tracking_uri: ml_flow_url
mlflow_experiment_name: test_1
hf_mlflow_log_artifacts: true

gradient_accumulation_steps: 4
micro_batch_size: 2
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
s2_attention:

hub_model_id: some_repo
hub_strategy: checkpoint
hub_token: 

save_steps: 5
save_strategy: steps
warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
max_steps: 10
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
   pad_token: " "

Possible solution

No response

Which Operating Systems are you using?

• [X] Linux
• [ ] macOS
• [ ] Windows

Python Version

3.10

axolotl branch-commit

winglian/axolotl:main-latest

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.

[NanoCode012]

Hey, thanks for reporting this. I noticed that our callback doesn't inherit MLflowCallback, so it's not properly reading that env.

At least, do you notice that the yaml is saved to mlflow?

https://huggingface.co/docs/transformers/v4.45.2/en/main_classes/callback#transformers.integrations.MLflowCallback

[winglian]

I think @awhazell added support for mlflow and might be able to help

[awhazell]

I can confirm the config yaml is saved as an artifact, but also that the checkpoints/end model are not.

A fix could be as simple as inheriting the callback @NanoCode012 linked (or adding it to the callbacks list separately) but would need to make sure it doesn't conflict with any of the setup from HFCausalTrainerBuilder.build

[NanoCode012]

Hey @awhazell , what potential conflicts were you thinking of?

My only concern may be duplicate logs due to report_to config and MLflowCallback logging.

Regarding the change needed, I believe we can just import the callback and append to the callbacks variable. Would you be interested in working on this PR?

[awhazell]

I was thinking about whether setting mlflow options in both the trainer kwargs and env variables could cause issues- but I think you're right and it shouldn't be an issue, they should always be consistent anyway

Opened a PR here https://github.com/axolotl-ai-cloud/axolotl/pull/1976