Add bisect_model API under ct.models.utils

[jakesabathia2]

• This is our proposed solution of issue unet Chunking not supported in coremltools>7.1 and <7.1 does not support dynamic reshape #2264
• Basically, we moved / and did some slight modification of the code in ml-stable-diffusion repo, and put the utils of bisecting a mlpackage under scope of ct.models.utils.bisect_model
• After this fix is in, I am going to put another PR in ml-stable-diffusion, and make the chunk_mlprogram.py directly calling into this new API, so that any further implementation details changes in coremltools will not break the script itself.

Testing:
https://gitlab.com/coremltools1/coremltools/-/pipelines/1389709376
https://gitlab.com/coremltools1/coremltools/-/pipelines/1389735502
https://gitlab.com/coremltools1/coremltools/-/pipelines/1392495182

[1duo]

Nice! Would be great to include example usages / tests.

[jakesabathia2]

Nice! Would be great to include example usages / tests.

Thanks! Just added.

[aseemw]

Can you also add a small section describing the util in the docs, along with its use case (when to use it, e.g. SD model thats big) : https://apple.github.io/coremltools/docs-guides/source/mlmodel-utilities.html

[TobyRoseman]

• Please make sure to get a successful CI run before merging.

[YifanShenSZ]

Awesome utility!

[jakesabathia2]

Can you also add a small section describing the util in the docs, along with its use case (when to use it, e.g. SD model thats big) : https://apple.github.io/coremltools/docs-guides/source/mlmodel-utilities.html

Done!

[aseemw]

PR looks good.
I think it would be good to merge, after addressing two more comments from my side:

• a few changes in the documentation
• If we can add another unit test which uses constexpr ops that would be great (take the uncompressed model, palletize it and then chunk it for example). In the original chunking script, I remember fixing a bunch of issues which only came up when we had used constexpr ops.

[jakesabathia2]

If we can add another unit test which uses constexpr ops that would be great (take the uncompressed model, palletize it and then chunk it for example). In the original chunking script, I remember fixing a bunch of issues which only came up when we had used constexpr ops.

Done!

[jakesabathia2]

https://gitlab.com/coremltools1/coremltools/-/pipelines/1392495182
CI passes, except the test_pytorch_310 exceed time limit

[nighting0le01]

hi i think there is issue with this API when the model is quantized before chunking: #2320