Skip to content

[gemini]: fix loading checkpoints of models with tied parameters #6447

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We鈥檒l occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
LiRunGuo wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[gemini]: fix loading checkpoints of models with tied parameters #6447

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion colossalai/zero/gemini/gemini_ddp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -813,7 +813,9 @@ def load_parameter(chunk_slice, data):
for key in state_dict.keys():
if key.startswith(prefix) and key != extra_state_key:
input_name = key[len(prefix) :]
if input_name not in local_state:
# tied parameters are saved under every name they have (same as torch.nn.Module.state_dict),
# while local_state only holds the first one, so the other names are expected too
if input_name not in local_state and input_name not in self.name2param:
unexpected_keys.append(key)

def _init_chunks(self, param_order, strict_ddp_mode: bool, cpu_offload: bool, pin_memory: bool):
Expand Down
28 changes: 27 additions & 1 deletion tests/test_checkpoint_io/test_gemini_checkpoint_io.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@
import pytest
import torch
import torch.distributed as dist
from transformers import LlamaForCausalLM
from transformers import GPT2Config, GPT2LMHeadModel, LlamaForCausalLM
from utils import shared_tempdir

import colossalai
Expand Down Expand Up @@ -207,11 +207,37 @@ def exam_lazy_from_pretrained():
check_state_dict_equal(state_dict, orig_state_dict, ignore_dtype=True)


@clear_cache_before_run()
@parameterize("shard", [True, False])
def exam_tied_weights_state_dict(shard: bool):
# lm_head.weight and transformer.wte.weight are the same parameter, which is saved under both names
config = GPT2Config(n_layer=2, n_head=4, n_embd=64, vocab_size=512, n_positions=64)

def model_fn():
model = GPT2LMHeadModel(config)
assert model.lm_head.weight is model.transformer.wte.weight
return model

booster = Booster(plugin=GeminiPlugin(precision="fp16", initial_scale=(2**14)))
model, *_ = booster.boost(model_fn())
new_model, *_ = booster.boost(model_fn())

with shared_tempdir() as tempdir:
model_ckpt_path = f"{tempdir}/model"
booster.save_model(model, model_ckpt_path, shard=shard)
dist.barrier()
booster.load_model(new_model, model_ckpt_path)
state_dict = model.state_dict(only_rank_0=False)
assert "lm_head.weight" in state_dict and "transformer.wte.weight" in state_dict
check_state_dict_equal(state_dict, new_model.state_dict(only_rank_0=False), ignore_dtype=True)


def run_dist(rank, world_size, port):
colossalai.launch(rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
exam_state_dict()
exam_state_dict_with_origin()
exam_lazy_from_pretrained()
exam_tied_weights_state_dict()


@pytest.mark.dist
Expand Down