Unify model and input loading via tt-forge-models repository

[chandrasekaranpradeep]

Summary

Standardize all model and input loading in tt-forge-onnx to use the tt-forge-models repository (e.g. third_party/tt_forge_models), remove PyTorch and JAX model tests from forge/test/models, and align ONNX, PaddlePaddle, and TensorFlow tests with a single source of truth for models and inputs.

Context

• tt-forge-onnx focuses on frameworks not covered by tt-xla: ONNX, PaddlePaddle, and TensorFlow.
• tt-forge-models is the shared model repository used across TT-Forge frontend projects and provides a consistent ModelLoader (and related) interface for loading models and inputs.
• Currently, loading strategies are inconsistent:
  • ONNX: Some tests use tt-forge-models (e.g. regnet, dla, googlenet, fuyu, clip, roberta, centernet); many others load PyTorch models in-test and use local utils or URLs/checkpoints for weights and inputs.
  • PaddlePaddle: Models and inputs are loaded inside the test suite via framework APIs (e.g. resnet18(pretrained=True), paddle.rand(...)) and local utils, not from tt-forge-models.
  • TensorFlow: Models and inputs are loaded in-test (e.g. Keras ResNet50(weights="imagenet")) and local model_utils (e.g. get_sample_inputs()), not from tt-forge-models.
• forge/test/models still contains PyTorch and JAX model tests, which are out of scope for tt-forge-onnx (PyTorch/JAX are covered by tt-xla).

Goals

1. ONNX

   • Load all models and inputs from tt-forge-models (via its loaders), instead of:
     • Loading PyTorch models in-test.
     • Downloading .pt/.pth from external URLs.
     • Using local utils (e.g. test.models.pytorch.vision.mnist.model_utils.utils) or repo-local files (e.g. forge/test/models/files/samples/audio/).
       Note: We do not currently have loaders for the ONNX framework, so for now load PyTorch models from the tt-forge-models repository, convert them to ONNX using the torch.onnx.export API, and then compile and execute them.

2. PaddlePaddle

   • Load models and inputs from tt-forge-models (using its loaders/utils) instead of loading models and constructing inputs entirely within the test suite (e.g. eval(variant)(pretrained=True), paddle.rand(...), or local utils).

3. TensorFlow

   • Load models and inputs from tt-forge-models instead of using framework APIs and local model_utils (e.g. Keras ResNet50(weights="imagenet") and get_sample_inputs() in test.models.tensorflow.vision.resnet.model_utils.image_utils).

4. Cleanup

   • Remove PyTorch model tests under forge/test/models/pytorch/.
   • Remove JAX model tests under forge/test/models/jax/.
   • Rely on tt-forge-models (and, where applicable, ONNX tests that use it) as the single source for model definitions, weights, and inputs for the frameworks in scope.

cc @nvukobratTT