[fix] update integration test values#47146
Conversation
|
run-slow: dac, moonshine_streaming |
|
This comment contains models: ["models/dac", "models/moonshine_streaming"] |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
There was a problem hiding this comment.
Hmm so we dont have a repro here, should we start including one here?
There was a problem hiding this comment.
this was a day0, transformers is the only integration 🤗
There was a problem hiding this comment.
Super nit but might be worth to refactor into fixtures structure, so many values :D
|
[For maintainers] Suggested jobs to run (before merge) run-slow: dac, moonshine_streaming |
CI recapDashboard: View test results in Grafana |
What does this PR do?
Fixes
torch 2.11 CI bump in #45243 and #45248 set
cudnn.conv.fp32_precision="ieee", forcing full-precision conv and shifting DAC/moonshine outputs off their TF32-recorded expected values. This was confirmed by running tests twice: once with default cuDNN settings (matches the previously recorded test values) and once withtorch.backends.cuda.matmul.allow_tf32 = False+torch.backends.cudnn.allow_tf32 = False(equivalent to what conftest.py's enable_tf32(False) + torch 2.11's new cudnn.conv.fp32_precision="ieee" produce). First run passed, meaning numerical drift.To update the expected: