Helper function to reload a quantized state_dict

[ManoBharathi93]

issue number: #162

[dacorvo]

@ManoBharathi93 your pull-request is not exactly aligned with the contributions guidelines.
In particular, please try to use conventional commits.
Regarding your tests, you simply duplicated the existing quantize tests without calling the newly introduced function, hence not verifying it works.

What I suggest for requantize tests is that you simply extract test_serialize_quantized_mlp from test_quantize_mlp.py and put it in a new test_requantize_mlp.py file.
Then, change the two lines at 123-124 to use your new helper.

[ManoBharathi93]

@dacorvo I have loaded quantized state_dict through requantize helper to model. but fails when it run in cuda.

Can i skip cuda using pytest.mark.skip_device ? coz it's passes the test case on cpu.
``

================================================================================================ short test summary info ================================================================================================
FAILED test/model/test_requantize_mlp.py::test_serialize_quantized_mlp[cuda-safetensors-fp16-w-qint8] - RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument other in method wrapper_CUDA__equal)
FAILED test/model/test_requantize_mlp.py::test_serialize_quantized_mlp[cuda-safetensors-fp32-w-qint8] - RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument other in method wrapper_CUDA__equal)
================================================================================== 2 failed, 47 passed, 7 skipped, 3 warnings in 2.71s ==================================================================================

``

[ManoBharathi93]

@dacorvo , can you check my changes please.

[dacorvo]

Please rebase your branch, then make sure it follows the contributing guidelines: the CI fails even before running the tests.
If you have trouble rebasing your branch, please look at the instructions in the guidenies and let me know if something is not clear.

[dacorvo]

Also, as I said before please do not cc all the tests from test_quantize_mlp.py in your new test file. This increases the codebase without any benefit.

[ManoBharathi93]

Also, as I said before please do not cc all the tests from test_quantize_mlp.py in your new test file. This increases the codebase without any benefit.

Okay I will modify in the new commit

[ManoBharathi93]

@dacorvo, Added test for MLP and serialize_quantized_mlp only and removed rest of the tests

[dacorvo]

Commit messages:
🚩 added_requantize
🚩 Add: Helper to reload a quantized state_dict
🚩 add:helper to requantize
✅ fix:style corrections
Error: 🚫 According to the conventional-commits specification, some of the commit messages are not valid.

[dacorvo]

If you just want to reword your commits, rebase iteratively from the root of your branch:

$ git rebase -i c75cd06

In the editor, select r or reword in front of the guilty commits, save and exit.
Now, rewrite each message when prompted.
Once done:

$ git push -f origin reload_quantized_state_dict

[ManoBharathi93]

Commit messages: 🚩 added_requantize 🚩 Add: Helper to reload a quantized state_dict 🚩 add:helper to requantize ✅ fix:style corrections Error: 🚫 According to the conventional-commits specification, some of the commit messages are not valid.

It means, red flagged commits needs to rename as fix: added_requantize,... ?

[dacorvo]

They should all start with an allowed prefix, whose list is here:
416db04

[dacorvo]

Alternatively, you can squash all your commits into one clean commit an rephrase it to feat: add requantize helper or something like that (just select s on the left while rebasing).

[ManoBharathi93]

They should all start with an allowed prefix, whose list is here: 416db04

Then in my case all should be prefix with fix: ..

[dacorvo]

I see this may be the first time you use this git feature. As I said, you could just squash them all (select s), and use a single commit message. Anyway, with reword what you edited on the first page will be ignored unfortunately, and git will ask you again for each commit (because on the first page it is just the commit titles).

[ManoBharathi93]

feat: add requantize helpe

Yes I was my first time with this git features. I was squash as you said, but it deletes all my commits and my head points to this commit c75cd now and

[dacorvo]

$ git rebase --continue

[ManoBharathi93]

git rebase --continue

git rebase --continue
error: cannot 'squash' without a previous commit
error: please fix this using 'git rebase --edit-todo'.

[dacorvo]

ah yes, the first one must stay if you squash (p for the first one, s for the others)
You can alwaus git rebase --abort and restart if you want

[ManoBharathi93]

@dacorvo , Thanks for your patience and helping me, it takes some time to understand for me.. Now kindly check this PR.

[ManoBharathi93]

@dacorvo do I missed something ? I have been waiting for your comments..test cases passed

[dacorvo]

@dacorvo do I missed something ? I have been waiting for your comments..test cases passed

Patience is the key ... you must understand that:

• this is an open-source project,
• I have many other things to do.

Also, please credit @calmitchell617 in your commit message because you basically cced his branch.

[ManoBharathi93]

Credits:@calmitchell617 with help of his work I made this PR Thanks 😊 .

[ManoBharathi93]

@calmitchell617 work has to be merged since I reuse his code with some extra test case I am closing this pr.

[dacorvo]

@ManoBharathi93 thank you for acknowledging this. I have other issues that require help if you want to contribute.

[ManoBharathi93]

@ManoBharathi93 thank you for acknowledging this. I have other issues that require help if you want to contribute.

😊 absolutely.