Fix FSDP policy imports and clean up diffusion benchmarks (#195909)
Summary:
## Background
PyTorch’s Dynamo benchmark configuration contains several stale diffusion-related entries. Separately, FSDP policy selection eagerly imports model-specific dependencies, including an obsolete Diffusers module path. This can break FSDP benchmarks even when running unrelated models.
## What this PR does
- Lazily imports only the model class required by the selected FSDP policy.
- Updates the Stable Diffusion transformer import to its current Diffusers path.
- Adds focused regression tests without requiring Diffusers or TorchBench.
- Removes nonexistent stable_diffusion configuration and stale expected-result entries.
- Documents why the skipped component models are absent from the aggregate dispatcher.
- Corrects the auraflow CLI spelling while retaining auroflow as a compatibility alias.
## These tests actually fail when the behavior breaks
Restoring the pre-fix `get_fsdp_auto_wrap_policy` (eager imports, old path) --
both tests die before reaching an assertion:
ERROR: test_default_fsdp_policy_does_not_import_model_dependencies
ERROR: test_diffusion_fsdp_policy_imports_current_model_class
ModuleNotFoundError: No module named 'diffusers'
Keeping the lazy import but reverting only the module path:
FAIL: test_diffusion_fsdp_policy_imports_current_model_class
AssertionError: expected call not found.
Expected: import_module('diffusers.models.transformers.transformer_2d')
Actual: import_module('diffusers.models.transformer_2d')
Keeping the corrected path but importing Diffusers eagerly at function entry,
i.e. fixing the path without fixing the laziness:
ERROR: test_default_fsdp_policy_does_not_import_model_dependencies
ModuleNotFoundError: No module named 'diffusers'
X-link: https://github.com/pytorch/pytorch/pull/195909
Approved by: https://github.com/anijain2305
Reviewed By: izaitsevfb
Differential Revision: D119521120
fbshipit-source-id: fcd4dada11c3c9182eb6759acaf88f0e0c40be6f