updating gptj-config #109
updating gptj-config #109
Conversation
examples/ppo_sentiments.py
Outdated
| @@ -17,7 +17,7 @@ def get_positive_score(scores): | |||
| return dict(map(lambda x: tuple(x.values()), scores))["POSITIVE"] | |||
|
|
|||
|
|
|||
| default_config = yaml.safe_load(open("configs/ppo_config.yml")) | |||
| default_config = yaml.safe_load(open("configs/ppo_gptj.yml")) | |||
There was a problem hiding this comment.
are you sure you want to change the default to gpt-j? I wouldn't mind the second script which imports main from here and changes config as the most simplest option right now, until we switch to hydra or something else
There was a problem hiding this comment.
Agreed, this is a bad idea.
There was a problem hiding this comment.
Yeah I just forgot to change this back.
|
Added distributed config logging to wandb |
configs/ppo_config.yml
Outdated
| @@ -21,6 +21,7 @@ train: | |||
|
|
|||
| pipeline: "PromptPipeline" # prompt pipeline to load | |||
| orchestrator: "PPOOrchestrator" # orchestrator to load | |||
| entity_name: "dahoas" # put your wandb login here | |||
There was a problem hiding this comment.
this will give somewhat cryptic error if you're not logged in as dahoas, maybe we can make this as optional enviroment variable instead?
There was a problem hiding this comment.
Lets fix this in a future PR
There was a problem hiding this comment.
Have it automatically determine what W&B account is logged in. cc @ayulockin
There was a problem hiding this comment.
it already does so, this option is optional and was added in #78
There was a problem hiding this comment.
otherwise it forces people to wandb disable as in #106
trlx/utils/__init__.py
Outdated
| return { | ||
| "mixed_precision": accelerate_config.mixed_precision, | ||
| "num_gpus": accelerate_config.num_processes, | ||
| "gradient_accumulation_steps": ds_plugin.gradient_accumulation_steps, |
There was a problem hiding this comment.
I think if ds is not used then ds_plugin is None, giving attribute error here
https://github.com/huggingface/accelerate/blob/e4e5611e5d4270a846caf42cba3388e54b83f074/src/accelerate/state.py#L62
maybe some processing of repr(accelerator.state) (which is outputed in accelerate env) would be equivalent here
There was a problem hiding this comment.
I guess it's true in some cases a user may not be using deepspeed so a check should be performed. However with the accelerator.state object there are some items I don't care about (such as the local rank).
Updating the gptj config and verifying gptj runs on 8 A100s with zero2
https://wandb.ai/dahoas/trlx/runs/5vc7xsx8?workspace=user-dahoas