Fix save_model and load_model

[james77777778]

Should fix #19921

The root cause is that #19852 will incorrectly duplicate *.weights.h5 in save_model.
And we should consider falling back to the original behavior if the system is read-only in load_model.

Sorry for the inconvenience!

[fchollet]

LGTM, thanks for the fix.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 73.91304% with 6 lines in your changes missing coverage. Please review.

Project coverage is 67.21%. Comparing base (558d38c) to head (80eccac).
Report is 2 commits behind head on master.

Files	Patch %	Lines
keras/src/saving/saving_lib.py	73.91%	3 Missing and 3 partials ⚠️

❗ There is a different number of reports uploaded between BASE (558d38c) and HEAD (80eccac). Click for more details.

HEAD has 2 uploads less than BASE

| Flag | BASE (558d38c) | HEAD (80eccac) | |------|------|------| |keras|4|2|

Additional details and impacted files

@@             Coverage Diff             @@
##           master   #19924       +/-   ##
===========================================
- Coverage   79.02%   67.21%   -11.81%     
===========================================
  Files         499      499               
  Lines       46436    46448       +12     
  Branches     8548     8550        +2     
===========================================
- Hits        36695    31222     -5473     
- Misses       8015    13567     +5552     
+ Partials     1726     1659       -67     

Flag	Coverage Δ
keras	67.14% <73.91%> (-11.74%)	⬇️
keras-jax	?
keras-numpy	57.25% <69.56%> (+0.02%)	⬆️
keras-tensorflow	?
keras-torch	62.40% <73.91%> (+0.02%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[fchollet]

What test can we add to prevent a similar issue in the future?

[james77777778]

What test can we add to prevent a similar issue in the future?

The reduant *.weights.h5 detection has been added in keras/src/saving/saving_lib_test.py
However, I don't have a good idea to detect OSError (for read-only system).

[james77777778]

I will try to add the tests for OSError tomorrow. (using os.chmod?!)

[WeichenXu123]

Hi @james77777778 Could you check #19921 (comment) ? seemingly the fix doesn't work.

[james77777778]

Hi @james77777778 Could you check #19921 (comment) ? seemingly the fix doesn't work.

Hey @WeichenXu123
I have proposed a fix for this in #19927

The issue may stem from a race condition in load_model.
Is your CI running load_model in multiple processes within a short period of time?

[WeichenXu123]

Hi @james77777778 Could you check #19921 (comment) ? seemingly the fix doesn't work.

Hey @WeichenXu123 I have proposed a fix for this in #19927

The issue may stem from a race condition in load_model. Is your CI running load_model in multiple processes within a short period of time?

Yes. It only occurs when multiple processes concurrently loading the model

[james77777778]

Yes. It only occurs when multiple processes concurrently loading the model

Thanks for the information. I have added a test to verify the concurrently loading in #19927 .