Repository navigation
test: pin RoFormer forward-pass budget + GPU per-pass timing guard (v0.47.1) - #313
Merged
Merged
Conversation
…0.47.1) v0.47.0 (#299) moved MDXC/RoFormer overlap to MSST num_overlap semantics. That's correct, but it silently made the MelBand karaoke models (YAML num_overlap 4/8) run 4-6x more forward passes than 0.46.x: karaoke-gen saw 25->75s, 26->149s and 27->157s per track on an L4 and had to pin 0.44.2. - tests/unit/test_mdxc_roformer_pass_budget.py pins the forward passes for a 3-minute track for five real model geometries, at the YAML overlap and at overlap=2, so any future chunking change that moves compute fails on every PR. Mutation-checked: restoring the 0.44 step formula fails all six tests. - tests/integration/test_roformer_inference_timing.py measures per-pass GPU time for the aufr33 karaoke model (CUDA only; added to the core-models job). - README: performance note on the 0.47 overlap semantics and how to opt back into speed with overlap=2. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jw21syo4HizJJjYw2mg2Vp
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jw21syo4HizJJjYw2mg2Vp
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
@coderabbitai ignore
Why
v0.47.0 (#299) changed MDXC/RoFormer
overlapto MSSTnum_overlapsemantics (step = chunk / overlap, defaulted from the model YAML). That change is correct and stays. It did silently make the MelBand karaoke models, whose YAMLs setnum_overlap4/8, run 4–6× more forward passes than 0.46.x. 0.44.2–0.46.0 used a fixed 8 s step, which in practice meant almost no overlap. karaoke-gen measured 25→75s (aufr33), 26→149s (gabox_v2) and 27→157s (becruily) per 3-min track on an L4 and pinned 0.44.2. Those slowdowns match the pass-count ratios (3.9×/5.7×/5.7×). The instrumental models (num_overlap 2) were unaffected. #298 is not the cause: 0.45.0 and 0.46.0 still use the old step formula.What
tests/unit/test_mdxc_roformer_pass_budget.py: pins forward passes for a 3-min track for five real model geometries, at the YAML overlap and at overlap=2. Any future chunking/step/tail change that moves compute now fails on every PR. Mutation-checked: restoring the 0.44 step formula fails all 6 tests.tests/integration/test_roformer_inference_timing.py(CUDA only, added tocore-models): per-forward-pass GPU time for the aufr33 karaoke model at YAML overlap and overlap=2. It prints the timings and has a generous ceiling (ROFORMER_MAX_SECONDS_PER_PASS, default 2.0s).overlap=2.Testing
test_mps_native_fp16failures (scatter(): Yet not supported for complex) also fail on untouchedmainon this Mac's MPS torch, so they're unrelated. CI runs Linux.🤖 Generated with Claude Code
https://claude.ai/code/session_01Jw21syo4HizJJjYw2mg2Vp