Repository navigation
Conversation
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
|
Self-review (Codex-assisted): checked the posted diff at |
|
Closing as superseded by merged #8047. Revalidated on current main |
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
|
Self-review (Codex-assisted): checked the posted diff at |
|
Closing as superseded by merged #8047. Revalidated on current main |
Problem and reproduction
AMD CI reports six failures in
tests/diffusion/distributed/test_wan_vae_fastpath_install.py. Reproduced locally on main75570ee3a6572a7dc7a9463ada35a9a11c367993:Before: 6 failed, 73 passed (norm wrapper rejection, install reversibility, pending bias, dtype checks, and stale import). Running the installer file alone gives 1 failed, 64 passed.
Root cause
Importing Wan2.2 replaces diffusers'
WanRMS_normprocess-wide with OmniRMSNormVAE. These tests specifically require diffusers' original numerical semantics; the fastpath correctly rejects the substitute. The substitute test also imports the class from its old location.Change
Load a private copy of the installed diffusers module once to obtain its original normalization class. A function-scoped monkeypatch binds that class during these tests and restores the previous binding afterward. Import
RMSNormVAEfrom its current Wan2.2 module before fixtures execute. Production code, test assertions and tolerances are unchanged; no tests are skipped.Validation
Too many argumentsdiagnostic atoriginal_setattr(module, name, value)(main line 207, branch line 229). Baseline and branch each report only this diagnostic.git diff --check: passed.Tests use CPU tensors and require the repository test dependencies, no model weights. Environment: Python 3.12, torch 2.12.0+git6bbd260 / HIP 7.2.53211, diffusers 0.40.0, transformers 5.14.1, vllm 0.29.0+rocm723.
AI assistance: Codex investigated the CI failures, reproduced them locally, implemented the test isolation, ran validation and drafted this PR. Human review is pending.