[Bug] ZeroRedundancyOptimizer cannot support param-wise settings with torch.__version__ < 1.12.0

[nijkah]

Prerequisite

• I have searched Issues and Discussions but cannot get the expected help.
• The bug has not been fixed in the latest version(https://github.com/open-mmlab/mmengine).

Environment

OrderedDict([('sys.platform', 'linux'), ('Python', '3.7.2 (default, May 11 2021, 10:20:27) [GCC 7.5.0]'), ('CUDA available', True), ('numpy_random_seed', 2147483648), ('GPU 0', 'NVIDIA A100-SXM-80GB MIG 2g.20gb'), ('CUDA_HOME', '/usr/local/cuda'), ('NVCC', 'Cuda compilation tools, release 11.1, V11.1.105'), ('GCC', 'gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0'), ('PyTorch', '1.9.1+cu111'), ('PyTorch compiling details', 'PyTorch built with:\n - GCC 7.3\n - C++ Version: 201402\n - Intel(R) Math Kernel Library Version 2020.0.0 Product Build 20191122 for Intel(R) 64 architecture applications\n - Intel(R) MKL-DNN v2.1.2 (Git Hash 98be7e8afa711dc9b66c8ff3504129cb82013cdb)\n - OpenMP 201511 (a.k.a. OpenMP 4.5)\n - NNPACK is enabled\n - CPU capability usage: AVX2\n - CUDA Runtime 11.1\n - NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86\n - CuDNN 8.0.5\n - Magma 2.5.2\n - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=11.1, CUDNN_VERSION=8.0.5, CXX_COMPILER=/opt/rh/devtoolset-7/root/usr/bin/c++, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_KINETO -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_VERSION=1.9.1, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, \n'), ('TorchVision', '0.10.1+cu111'), ('OpenCV', '4.5.1'), ('MMEngine', '0.3.1')])

Reproduces the problem - code sample

configs/swin/mask-rcnn_swin-t-p4-w7_fpn_1x_coco.py

# configs/swin/mask-rcnn_swin-t-p4-w7_fpn_1x_coco.py
# optimizer config is edited like
optim_wrapper = dict(
    type='OptimWrapper',
    paramwise_cfg=dict(
        custom_keys={
            'absolute_pos_embed': dict(decay_mult=0.),
            'relative_position_bias_table': dict(decay_mult=0.),
            'norm': dict(decay_mult=0.)
        }),
    optimizer=dict(
        _delete_=True,
        type='ZeroRedundancyOptimizer',
        optimizer_type='AdamW',
        lr=0.0001,
        betas=(0.9, 0.999),
        weight_decay=0.05))

Reproduces the problem - command or script

bash tools/dist_train.sh configs/swin/mask-rcnn_swin-t-p4-w7_fpn_1x_coco.py 1 

Reproduces the problem - error message

11/30 10:45:04 - mmengine - INFO - paramwise_options -- roi_head.mask_head.conv_logits.bias:lr=0.0001
11/30 10:45:04 - mmengine - INFO - paramwise_options -- roi_head.mask_head.conv_logits.bias:weight_decay=0.05
Traceback (most recent call last):
  File "/nas/k8s/dev/mlops/hakjinlee/workspace/sandbox/mmengine/mmengine/registry/build_functions.py", line 121, in build_from_cfg
    obj = obj_cls(**args)  # type: ignore
  File "/nas/k8s/dev/mlops/hakjinlee/workspace/sandbox/mmengine/mmengine/optim/optimizer/zero_optimizer.py", line 60, in __init__
    super().__init__(params, optimizer_class, **kwargs)
  File "/usr/local/lib/python3.7/site-packages/torch/distributed/optim/zero_redundancy_optimizer.py", line 166, in __init__
    self._reference_is_trainable_mask = list(map(_is_trainable, self._all_params))
  File "/usr/local/lib/python3.7/site-packages/torch/distributed/optim/zero_redundancy_optimizer.py", line 49, in _is_trainable
    return param.requires_grad
AttributeError: 'dict' object has no attribute 'requires_grad'

Additional information

Related to #716

Some configurations such as mask-rcnn_swin-t-p4-w7_fpn_1x_coco.py requires paramwise_cfg to give different parameters to some modules.

However, torch.distributed.optim.ZeroRedundancyOptimizer only recently started supporting the feature for multiple param groups. (Link) (It supports in torch.__version__ >= 1.12.0).

The key implementation is

• Save parameters as List[torch.Tensor] in ZeroRedundancyOptimizer._all_params
• Give original parameters to the inner optimizer(optimizer_class) as Optimizer.__init__(self, params, defaults)

We can copy-and-paste its __init__ logic to support param-wise settings when torch.__version__ < 1.12.0.
However, I think this kind of solution is quite messy and vulnerable to change in torch.distributed.optim.ZeroRedundancyOptimizer.

I kindly ask for any good ideas to handle this!

[C1rN09]

I'll take a look ✋

[C1rN09]

I got it. So this feature requires an intrusive change in ZeroRedundancyOptimizer.

We can copy-and-paste its __init__ logic to support param-wise settings when torch.__version__ < 1.12.0. However, I think this kind of solution is quite messy and vulnerable to change in torch.distributed.optim.ZeroRedundancyOptimizer.

I quite agree with your opinion! We cannot do this copy-and-paste forever. As an alternative, what about do some checks (e.g. all_dict) in our __init__ and then give some more instructive information in assertion? We can explain the reason and suggest upgrading to torch>=1.12 😸

[C1rN09]

Hi, @nijkah ! Can you take a look? #818