Skip to content
GitLab
Open
Issue created by Administrator@rootOwner

No longer able to load provided OPT checkpoint after recent changes

Created by: EIFY

🐛 Bug

No longer able to load provided OPT checkpoint after recent changes

To Reproduce

Edit metaseq/service/constants.py as before, in my case:

MAX_SEQ_LEN = 2048
BATCH_SIZE = 2048  # silly high bc we dynamically batch by MAX_BATCH_TOKENS
MAX_BATCH_TOKENS = 3072
DEFAULT_PORT = 6010
MODEL_PARALLEL = 1
TOTAL_WORLD_SIZE = 1
MAX_BEAM = 16

try:
    # internal logic denoting where checkpoints are in meta infrastructure
    from metaseq_internal.constants import CHECKPOINT_FOLDER
except ImportError:
    CHECKPOINT_FOLDER = "/home/jason_chou/redspot_home/350m/"
(...)

where

$ pwd
/home/jason_chou/redspot_home
$ ls 350m/
dict.txt  gpt2-merges.txt  gpt2-vocab.json  reshard.pt

and then run metaseq-api-local, but it no longer works:

$ metaseq-api-local
2022-10-05 22:19:25 | INFO | metaseq.hub_utils | loading model(s) from /home/jason_chou/redspot_home/350m/reshard.pt
2022-10-05 22:19:26 | INFO | metaseq.checkpoint_utils | Done reading from disk
Traceback (most recent call last):
  File "/home/jason_chou/.conda/envs/user/bin/metaseq-api-local", line 8, in <module>
    sys.exit(cli_main())
  File "/home/default_user/metaseq/metaseq_cli/interactive_hosted.py", line 370, in cli_main
    distributed_utils.call_main(cfg, worker_main, namespace_args=args)
  File "/home/default_user/metaseq/metaseq/distributed/utils.py", line 279, in call_main
    return main(cfg, **kwargs)
  File "/home/default_user/metaseq/metaseq_cli/interactive_hosted.py", line 176, in worker_main
    models = generator.load_model()  # noqa: F841
  File "/home/default_user/metaseq/metaseq/hub_utils.py", line 565, in load_model
    models, _model_args, _task = _load_checkpoint()
  File "/home/default_user/metaseq/metaseq/hub_utils.py", line 548, in _load_checkpoint
    return checkpoint_utils.load_model_ensemble_and_task(
  File "/home/default_user/metaseq/metaseq/checkpoint_utils.py", line 482, in load_model_ensemble_and_task
    model = build_model_hook(cfg, task)
  File "/home/default_user/metaseq/metaseq/hub_utils.py", line 538, in _build_model
    setattr(cfg["model"], "inference", True)
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/omegaconf/dictconfig.py", line 337, in __setattr__
    raise e
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/omegaconf/dictconfig.py", line 334, in __setattr__
    self.__set_impl(key, value)
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/omegaconf/dictconfig.py", line 318, in __set_impl
    self._set_item_impl(key, value)
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/omegaconf/basecontainer.py", line 511, in _set_item_impl
    self._validate_set(key, value)
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/omegaconf/dictconfig.py", line 180, in _validate_set
    target = self._get_node(key) if key is not None else self
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/omegaconf/dictconfig.py", line 465, in _get_node
    self._validate_get(key)
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/omegaconf/dictconfig.py", line 166, in _validate_get
    self._format_and_raise(
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/omegaconf/base.py", line 190, in _format_and_raise
    format_and_raise(
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/omegaconf/_utils.py", line 821, in format_and_raise
    _raise(ex, cause)
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/omegaconf/_utils.py", line 719, in _raise
    raise ex.with_traceback(sys.exc_info()[2])  # set end OC_CAUSE=1 for full backtrace
omegaconf.errors.ConfigAttributeError: Key 'inference' is not in struct
    full_key: model.inference
    object_type=dict

Apparently this can be traced back to when setattr(cfg["model"], "inference", True) was added (https://github.com/facebookresearch/metaseq/pull/356). However, another issue surfaced even with that line commented out:

$ metaseq-api-local
2022-10-05 22:23:31 | INFO | metaseq.hub_utils | loading model(s) from /home/jason_chou/redspot_home/350m/reshard.pt
2022-10-05 22:23:31 | INFO | metaseq.checkpoint_utils | Done reading from disk
Traceback (most recent call last):
  File "/home/jason_chou/.conda/envs/user/bin/metaseq-api-local", line 8, in <module>
    sys.exit(cli_main())
  File "/home/default_user/metaseq/metaseq_cli/interactive_hosted.py", line 370, in cli_main
    distributed_utils.call_main(cfg, worker_main, namespace_args=args)
  File "/home/default_user/metaseq/metaseq/distributed/utils.py", line 279, in call_main
    return main(cfg, **kwargs)
  File "/home/default_user/metaseq/metaseq_cli/interactive_hosted.py", line 176, in worker_main
    models = generator.load_model()  # noqa: F841
  File "/home/default_user/metaseq/metaseq/hub_utils.py", line 565, in load_model
    models, _model_args, _task = _load_checkpoint()
  File "/home/default_user/metaseq/metaseq/hub_utils.py", line 548, in _load_checkpoint
    return checkpoint_utils.load_model_ensemble_and_task(
  File "/home/default_user/metaseq/metaseq/checkpoint_utils.py", line 487, in load_model_ensemble_and_task
    model.load_state_dict(state["model"], strict=strict)
  File "/home/default_user/.conda/envs/user/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1604, in load_state_dict
    raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for TransformerLanguageModel:
        Missing key(s) in state_dict: "decoder.layer_norm.weight", "decoder.layer_norm.bias".

which seems to be due to recent cleanup PRs (https://github.com/facebookresearch/metaseq/pull/366, https://github.com/facebookresearch/metaseq/pull/380, https://github.com/facebookresearch/metaseq/pull/381).

Expected behavior

metaseq-api-local up & running

Environment

  • metaseq Version: latest main (7828d728)
  • PyTorch Version: 1.12.1+cu113
  • OS (e.g., Linux): Ubuntu 18.04.6 LTS
  • How you installed metaseq: pip
  • Python version: 3.10.4
  • CUDA/cuDNN version: CUDA 11.8
  • GPU models and configuration: 1 x T4