[BUG] RuntimeError: Error(s) in loading state_dict for T5ForConditionalGeneration:

Describe the bug
I'm running into this error while trying to load models under deepspeed initialization. It suggests "You may consider adding ignore_mismatched_sizes=True in the model from_pretrained method" which I don't want since it is saving the model with size 0.

model = AutoModelClass.from_pretrained(target) File "miniconda3/envs/mlc-llm/lib/python3.10/site-packages/transformers/models/auto/auto_factory.py", line 493, in from_pretrained return model_class.from_pretrained( File "miniconda3/envs/mlc-llm/lib/python3.10/site-packages/transformers/modeling_utils.py", line 2903, in from_pretrained ) = cls._load_pretrained_model( File "miniconda3/envs/mlc-llm/lib/python3.10/site-packages/transformers/modeling_utils.py", line 3310, in _load_pretrained_model raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}") RuntimeError: Error(s) in loading state_dict for T5ForConditionalGeneration: size mismatch for shared.weight: copying a param with shape torch.Size([32128, 768]) from checkpoint, the shape in current model is torch.Size([0]). ....

To Reproduce
target = 'google/flan-t5-base'

with deepspeed.zero.Init(): model = AutoModelClass.from_pretrained(target)

Expected behavior
I would expect the model is loaded as usual without deepspeed initialization and runs through training with no issues.

ds_report output
[2023-08-26 11:34:00,034] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (auto detect)
DeepSpeed C++/CUDA extension op report
JIT compiled ops requires ninja
ninja .................. [OKAY]
op name ................ installed .. compatible
[WARNING] async_io requires the dev libaio .so object and headers but these were not found.
[WARNING] async_io: please install the libaio-dev package with apt
[WARNING] If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [NO] ....... [NO]
fused_adam ............. [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
fused_lamb ............. [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
[WARNING] sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.0
[WARNING] using untested triton version (2.0.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
transformer_inference .. [NO] ....... [OKAY]

DeepSpeed general environment info:
torch install path ............... ['miniconda3/envs/mlc-llm/lib/python3.10/site-packages/torch']
torch version .................... 2.0.1+cu117
deepspeed install path ........... ['miniconda3/envs/mlc-llm/lib/python3.10/site-packages/deepspeed']
deepspeed info ................... 0.10.1, unknown, unknown
torch cuda version ............... 11.7
torch hip version ................ None
nvcc version ..................... 11.3
deepspeed wheel compiled w. ...... torch 2.0, cuda 11.7
shared memory (/dev/shm) size .... 503.89 GB

System info (please complete the following information):
OS:
Distributor ID: Ubuntu
Description: Ubuntu 20.04.3 LTS
Release: 20.04
Codename: focal

GPU count and types 1 NVIDIA A100
Python version - 3.10.12

Launcher context
Are you launching your experiment with the deepspeed launcher, MPI, or something else?
importing deepspeed within my code and initializing like above with deepspeed.zero.Init():

Here is my deepspeed config I use for training (although the error occurs before then) which I had to the Seq2SeqTrainingArguments

ds_config = {
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },

    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "betas": "auto",
            "eps": 1e-8,
            "weight_decay": "auto"
        }
    },

    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": 0,
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto"
        }
    },

    "zero_optimization": {
        "stage": 2,
        "allgather_partitions": True,
        "allgather_bucket_size": 2e8,
        "overlap_comm": True,
        "reduce_scatter": True,
        "reduce_bucket_size": 2e8,
        "contiguous_gradients": True
    },

    "steps_per_print": 2000,
    "wall_clock_breakdown": False,
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",

}

Start by reproducing the reported google/flan-t5-base load through AutoModelClass.from_pretrained inside deepspeed.zero.Init(). Trace the interaction between DeepSpeed initialization and the T5 shared.weight loading path, then verify that loading completes with the expected parameter shape and training can proceed without ignore_mismatched_sizes=True.