fix(modeling): deepspeed checkpoint loading #482
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
jon-tow
requested changes
May 23, 2023
There was a problem hiding this comment.
Thanks for working on this, Max!
Checkpointing works properly for models that do not freeze any layers (at least under ZeRO-2) but not otherwise. E.g. for a model with num_layers_unfrozen=2 I get the following error.
RuntimeError: Error(s) in loading state_dict for AutoModelForCausalLMWithHydraValueHead:
Missing key(s) in state_dict: "frozen_head.decoder_blocks.0.ln_1.weight",
"frozen_head.decoder_blocks.0.ln_1.bias", "frozen_head.decoder_blocks.0.attn.bias",
"frozen_head.decoder_blocks.0.attn.masked_bias", "frozen_head.decoder_blocks.0.attn.c_attn.weight",
"frozen_head.decoder_blocks.0.attn.c_attn.bias", "frozen_head.decoder_blocks.0.attn.c_proj.weight",
"frozen_head.decoder_blocks.0.attn.c_proj.bias", "frozen_head.decoder_blocks.0.ln_2.weight",
"frozen_head.decoder_blocks.0.ln_2.bias", "frozen_head.decoder_blocks.0.mlp.c_fc.weight",
"frozen_head.decoder_blocks.0.mlp.c_fc.bias", "frozen_head.decoder_blocks.0.mlp.c_proj.weight",
"frozen_head.decoder_blocks.0.mlp.c_proj.bias", "frozen_head.decoder_blocks.1.ln_1.weight",
"frozen_head.decoder_blocks.1.ln_1.bias", "frozen_head.decoder_blocks.1.attn.bias",
"frozen_head.decoder_blocks.1.attn.masked_bias", "frozen_head.decoder_blocks.1.attn.c_attn.weight",
"frozen_head.decoder_blocks.1.attn.c_attn.bias", "frozen_head.decoder_blocks.1.attn.c_proj.weight",
"frozen_head.decoder_blocks.1.attn.c_proj.bias", "frozen_head.decoder_blocks.1.ln_2.weight",
"frozen_head.decoder_blocks.1.ln_2.bias", "frozen_head.decoder_blocks.1.mlp.c_fc.weight",
"frozen_head.decoder_blocks.1.mlp.c_fc.bias", "frozen_head.decoder_blocks.1.mlp.c_proj.weight",
"frozen_head.decoder_blocks.1.mlp.c_proj.bias", "frozen_head.final_norm.weight", "frozen_head.final_norm.bias",
"frozen_head.lm_head.weight". We'll probably need to filter frozen_head from state_dict when checkpointing. What do you think?
trlx/models/modeling_ppo.py
Outdated
| for k, v in v_head_state_dict.items(): | ||
| base_model_state_dict[f"v_head.{k}"] = v | ||
| return base_model_state_dict | ||
| base_model_state_dict = self.base_model.state_dict(*args, **dict(prefix="base_model.", **kwargs)) |
There was a problem hiding this comment.
Same as for ILQL - need to add support for Seq2Seq.
jon-tow
reviewed
Jun 7, 2023
|
@maxreciprocate Do you want to resolve the conflicts and then we will merge? |
|
@maxreciprocate Bump on this |
|
Looks good, merging |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR adds an option to load and resume training from previously saved deepspeed checkpoints
Example of successive training:
accelerate launch --config_file configs/accelerate/zero3.yaml examples/ppo_sentiments.py accelerate launch --config_file configs/accelerate/zero3.yaml examples/ppo_sentiments.py '{"train": {"resume_from_checkpoint": "ckpts/best_checkpoint"}}'