[Feature] Load partially instantiated state-dict

[vmoens]

Summary:

Allows to load state_dicts from disk when the saved copy contains more elements that the pre-populated copy.

Test plan: TODO

Fixes #{issue number} Closes #101

[codecov[bot]]

Codecov Report

Merging #103 (5501d83) into main (a11364c) will decrease coverage by 6.18%. The diff coverage is 80.85%.

@@            Coverage Diff             @@
##             main     #103      +/-   ##
==========================================
- Coverage   90.40%   84.22%   -6.19%     
==========================================
  Files          25       25              
  Lines        2346     2384      +38     
==========================================
- Hits         2121     2008     -113     
- Misses        225      376     +151     

Impacted Files	Coverage Δ
torchsnapshot/flatten.py	86.73% <66.66%> (-3.63%)	:arrow_down:
torchsnapshot/io_preparer.py	92.55% <72.72%> (-1.07%)	:arrow_down:
torchsnapshot/snapshot.py	92.75% <95.23%> (-3.03%)	:arrow_down:
torchsnapshot/storage_plugins/gcs.py	0.00% <0.00%> (-78.90%)	:arrow_down:
torchsnapshot/storage_plugins/s3.py	25.00% <0.00%> (-65.00%)	:arrow_down:
torchsnapshot/storage_plugin.py	43.75% <0.00%> (-21.88%)	:arrow_down:
torchsnapshot/memoryview_stream.py	63.33% <0.00%> (-6.67%)	:arrow_down:
torchsnapshot/io_types.py	79.68% <0.00%> (-3.13%)	:arrow_down:
torchsnapshot/scheduler.py	91.11% <0.00%> (-2.23%)	:arrow_down:
... and 2 more

:mega: We’re building smart automated test selection to slash your CI/CD build times. Learn more

[yifuwang]

Hi @vmoens, thank you so much for raising the issue and drafting a proposal!

Previously, in order to reduce the memory footprint on load, we would find the tensors in the target module's state dict, in-place restore them, and use them to build a state dict for .load_state_dict().

Your proposal made us realize that this is not the right approach. Instead, we should always rebuild the original state dict and reuse the tensor in the target's module's state dict in addition. This way we can achieve the same memory saving while making the behavior much more general.