Add load and run tests for checkpoints that we want to have BC

[jerryzh168]

Stacked PRs:

• Refactor TorchAOBaseTensor for better BC support #2793
• ->Add load and run tests for checkpoints that we want to have BC #2792

---

Add load and run tests for checkpoints that we want to have BC

Summary:
Added load and run tests to make sure previously saved checkpoints can continue to load and run.

includes FP8, INT4 and INT4 + preshuffled checkpoints since these might reach larger audience

Test Plan:
python test/integration/test_load_and_run_checkpoint.py

Reviewers:

Subscribers:

Tasks:

Tags:

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2792

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure, 1 Cancelled Job, 9 Pending

As of commit 483898d with merge base fee314b ():

NEW FAILURE - The following job has failed:

• Run TorchAO Experimental Tests / test-cpu-ops (macos-14) (gh)
  Process completed with exit code 1.

CANCELLED JOB - The following job was cancelled. Please retry:

• Run TorchAO Experimental Tests / test-cpu-ops (linux.arm64.2xlarge) (gh)
  ##[error]The operation was canceled.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[vkuzo]

should we ensure we have CI with sm90+ running this test?

[jerryzh168]

yeah it's tested here:

ao/.github/workflows/1xH100_tests.yml

Line 50 in 083361b

pytest test/integration --verbose -s

added this skip because it fails in this test:

ao/.github/workflows/1xL4_tests.yml

Line 50 in 083361b

pytest test/integration --verbose -s

(SM89)

[vkuzo]

nit: specify what here is torchao version and what is config version? torchao-testing/model:single_linear-config:Float8DynamicActivationFloat8WeightConfig-config_version:1-torchao:0.13-dev?

[jerryzh168]

this seems a bit long, would changing 0.13-dev to 0.13.dev be clearer that this is a torchao version?

also just found that : is not supported in hf model naming, only -, . and _ are supported.

can also move around the 0.13.dev to make it clearer as well, please let me know

[vkuzo]

this seems a bit long, would changing 0.13-dev to 0.13.dev be clearer that this is a torchao version?

I think we should go for clarity here, if it's long that's ok. Can we pick one of -, ., _ as a delimiter, and make the name be formatted with key1.val1,key2.val2,..., with whatever delimiters between pairs and key:val that we can use

[jerryzh168]

does this have to be in the name, how about adding these in model card descriptions? there are some restrictions to the naming, including max length of 96:

Only regular characters and '-', '_', '.' are accepted. '--' and '..' are forbidden. '-' and '.' cannot start or end the name. The name cannot end with ".git" or ".ipynb". Max length is 96.

tried Float8 and it's 111: model_single-linear-config_Float8DynamicActivationFloat8WeightConfig-config-version_v2-torchao-version_0.13.dev

[vkuzo]

would be good to make the naming exact

[jerryzh168]

updated model card to include exact descriptions due to naming length restrictions