ACE-Step validation

Use this evidence contract to qualify the native Swift/MLX ACE-Step 1.5 runtime. The contract separates source parity, local tests, installed-checkpoint execution, performance, and listening review.

Immutable sources

The implementation was compared with ACE-Step upstream commit 6d467e4b5081ccb0abf1ec1bf4fdf9051a2d34b0. Managed downloads use immutable Hugging Face revisions:

Component	Repository	Revision
Shared VAE/text/Turbo assets	ACE-Step/Ace-Step1.5	19671f406d603126926c1b7e2adc169acbcade22
1.7B planner	ACE-Step/Ace-Step1.5	19671f406d603126926c1b7e2adc169acbcade22
XL-Turbo	ACE-Step/acestep-v15-xl-turbo	d4a0b288b83ebb7e25a8c0b32c573c22e134e8ee
4B planner	ACE-Step/acestep-5Hz-lm-4B	0a3ec94b557aea7d508da38b31cfe7341f6ff737
XL-SFT	ACE-Step/acestep-v15-xl-sft	d06de46b4622f781cf07f4a013a67d591ca52819
XL-Base	ACE-Step/acestep-v15-xl-base	220c1166efbdd9583eafcb12eb160594bbfcb241

Generation recipes repeat the effective repository/revision set, adapter SHA-256 and scale, final effective conditioning metadata, complete inference configuration, candidate ranking, and input/output hashes. Recipe schema 5 adds the semantic-code CFG scale and negative prompt to schema 4's planner temperature, top-k/top-p, and repetition penalty and schema 3's resolved planner source, root, and immutable repository/revision provenance and schema 2's post-planning BPM, duration, key/scale, vocal language, and time signature. This protects old installs whose original manifest predates source provenance, distinguishes requested metadata from the values actually used, and records independently resolved DiT and planner components.

Parity and local gates

The deterministic test surface covers:

• Turbo and continuous timestep schedules, Euler and Heun updates, CFG, APG, ADG, guidance windows, and stabilization.
• Typed task and checkpoint capability routing and exact Base-only restrictions.
• source conditioning, repaint masks/injection/splice, retake endpoints and spherical interpolation, and flow-edit integration windows.
• prompt/condition encoder, audio-tokenizer, DCW, first-velocity, final-latent, and VAE parity dump paths.
• Python-contract metadata precedence: explicit duration and language survive conflicting planner proposals, 1.7B is discovered before 4B, and separately managed planners resolve without checkpoint-store symlinks.
• upstream two-phase LM prompting: metadata uses CFG 1.0; code generation continues from the planned CoT, guides conditional/unconditional logits at 2.0, and keeps the sampled token streams synchronized.
• stable seed fanout, candidate metrics/ranking, batch/session serialization, API decoding and security, WAV headers, LRC, recipes, and DAW bundle topology.
• PEFT LoRA and LyCORIS LoKr key mapping, numerical contributions, alpha/scale, decomposed Kronecker factors, and adapter stacking.

Pinned-upstream numerical probes isolate every core stage. On the installed XL-Turbo stack, Swift versus native MLX measured mean absolute errors of 9.74e-7 for DiT, 2.08e-7 for the condition encoder, 8.21e-6 for Qwen hidden states, 6.05e-7 for FSQ detokenization, and 6.32e-7 for VAE decode. The VAE also passes a real structured-audio round trip. This stage isolation identified the prior output failure as orchestration rather than checkpoint or kernel drift: Swift supplied a float chunk mask of 2.0, while upstream stores that assignment in a boolean tensor and therefore conditions the model with 1.0.

Run the normal repository contract:

./scripts/check.sh

Installed-checkpoint tests are opt-in because they require large local assets. The integration tests describe the required MERERUN_TEST_ACESTEP_* environment variables and skip truthfully when an asset is absent.

Installed-model evidence

All files in the installed-model evidence table were generated locally at 48 kHz stereo. Hashes make the specific validation artifacts identifiable without committing model output.

Path exercised	Configuration	Result
XL-SFT	50-step Heun, APG, final quality	979775ce2503c7c393b3fc435a45515ffb4be80a3658e26d90b541cc57a888a4
XL-Base extract	one-second source, one step, Drums	9c620ce47d89b67824c2eebd88c20c0a81b1c340355a2ddd0937c61db0f78851
XL-Base lego	one-second source, one step, Drums, 0.2–0.8 s mask	322d3c757da5477ec3a828d2d198abcb80c9a3ff14a20a3c2786119a6dacb171
XL-Base complete	one-second source, one step, Drums,Bass	1c8861285bcbf0952035b11aa1676c38f516a543c700c15a896105c0f9fdea35
XL-Turbo flow edit	one-second source, one step	ae758bdf4b940277fc09145f9340bdd852b1fd48915921ad231b86488353cb27
XL-Turbo + 4B recipe v2	one-second LM-planned vocal, one step	68b170220cf5849635f3f97f2b75d87d7a80045f2401edf9a707d677db2f2122
XL-Turbo + 4B upstream LM replay	85-second trailer prompt, seed 4747, 16-step Heun	65eaaac55db062189ab2d923ad5dcda0ecaa7287dcee00a7c2003af3b999ffad
Resident API WAV	warm one-second XL-Turbo request, one step	9c6f3ef3752f15636aa2f4eb7b74b3dcee556e78a62158e4bc4ac344f6eaee32
Float32 artifact and DAW workflow	recipe, LRC, candidates, bundle	d44d8386c6b479e209d169c74ff2aa69bd2b13dd980f83a708b9f381c985e629
LoRA train/save	one real backward and AdamW step, 256 layers	b7c9c40c3b74aca7fb8f12fb4999c0be1175db82bf9e1504bb9c19794a29c8dd
LoRA reload/generate	trained artifact applied to 256 layers	1ba7876df682c626b1f08da80a739ed0685becdf0182e492bb1c533d07d41449
LoKr train/save	one real backward and AdamW step, 256 layers	5175420753533a736b43afa608dfc957e48c8d216ac0ec0d6ac824030e1c739e
LoKr reload/generate	factored adapter applied to 256 layers	79a45e0d340ad0232bb81f4c68ff5c5cc57fc9541f11343f5171af0a319e1961
Provenance recipe	exact pins plus adapter hash and scale	6c3370f61c3ed3bea3fc48e99369432d1578696be2989d3713abbc094a68a2bb

The XL-SFT 50-step run completed in 15.52 seconds with 20.72 GB maximum RSS, 25.34 GB peak footprint, and no swap. The XL-Turbo flow edit completed in 4.98 seconds with 21.03 GB maximum RSS, 26.94 GB peak footprint, and no swap. The XL-Base extract, lego, and complete smokes loaded the immutable Base checkpoint and produced PCM24 output in 3.27, 2.31, and 4.87 seconds. Extract and lego peaked near 26.1 GB; complete peaked at 48.34 GB. All three completed without swap. The temporary Base installation was then removed through mere.run model remove, reclaiming 19,949,344,762 bytes while retaining shared assets still referenced by installed ACE-Step variants. The one-step LoRA train/save smoke took 6.52 seconds with 21.02 GB maximum RSS and 25.84 GB peak footprint. LoKr took 2.97 seconds with 21.18 GB maximum RSS and 26.15 GB peak footprint. Both adapter reload renders completed in roughly 2.3 seconds without swap. The real 4B-planned schema 2 recipe retained the effective BPM and explicit one-second duration while discarding unsupported planner language output.

The real resident API health probe reported XL-Turbo + 4B loaded and warm. A one-second PCM24 WAV request returned in 0.68 seconds. A heterogeneous batch returned independent candidate counts [1, 2], deterministic seeds [6201] and [6301, 6302], and exactly one selected candidate per item in 1.16 seconds. The response preserved per-item BPM, duration, key, and time-signature metadata. A real LM-planned API request returned effective BPM 108, G major, the explicit one-second duration, and omitted unsupported planner language. Wrong-model, wrong-content-type, and raw-WAV batch requests returned actionable HTTP 400 JSON errors.

The 85-second XL-Turbo + 4B replay exercised the full upstream-shaped LM path with one candidate: phase-one CFG 1.0, phase-two CFG 2.0, the NO USER INPUT unconditional prompt, temperature 0.85, top-p 0.9, and 425 semantic codes. Its schema 5 recipe records the requested seed 4747, explicit duration 85, language en, and the same effective values inside the CoT used for phase two. Diffusion used the explicitly requested 16-step Heun sampler with velocity normalization and EMA both disabled. The output is exactly 85 seconds of float32 stereo audio at 48 kHz; its recipe SHA-256 is 3530789814c6f0a98b03378fa989b7b0ebfe5419b7060817f83358a045672252.

Listening regression

Tests/MereRunCLITests/Fixtures/ACEStep/listening-regression.json freezes prompts, lyrics, seeds, duration, quality, candidate count, and listening criteria across transient electronic, acoustic, vocal, dense, and ambient material.

./scripts/acestep-listening-regression.sh

The runner writes the WAVs and exact recipes, SHA manifest, ordered M3U playlist, and a review CSV with structure, prompt adherence, audio quality, vocal alignment, regression, and notes fields. The automatic gate catches silence, clipping, DC, stationary broadband noise, missing time-varying spectral structure, and prematurely dead endings. The CSV deliberately retains a human listening decision for musical quality and prompt adherence.

The installed XL-Turbo + 4B run completed all five frozen cases as 48 kHz stereo PCM24. Each selected candidate used all 60 semantic audio codes expected for 12 seconds and passed temporalSpectralVariation >= 0.85 plus tailEnergyRatio >= 0.05:

Case	Score	Spectral variation	Tail ratio	SHA-256
transient electronic	86.7512	1.3137	0.0720	1a5c7c56e5373d69400f553256637199bb2b1e04877435abf8f2e2ebe6e555dc
acoustic space	90.2295	1.2200	0.5534	bb4a9a1e5a208bb6d594a83820ffd330c98e2d83aca1f932b7429572192dc3b7
vocal alignment	87.2707	0.9795	0.9391	781616e86a5947f46383f90744d69161dffb087d0f90e363e72b36fa62bb1b2c
dense arrangement	92.6983	1.3598	1.2525	91ca6ad435b7f5ee12a550aa0e19da3348bbfce0c805b110159ed0363680a704
ambient continuity	90.0086	1.1366	0.9754	356379ebeffdbeb7b50dd1e3f7062a00bfd742174d3a698cd4d74585a6a25391

Repeating the transient case with the same seed reproduced its planner metadata, candidate scores/code counts, and WAV byte-for-byte at SHA-256 1a5c7c56e5373d69400f553256637199bb2b1e04877435abf8f2e2ebe6e555dc.

The review CSV remains deliberately unscored until a person listens. The technical rank is evidence against the exact stationary-noise regression, but it is not a substitute for a musical-quality or lyric-alignment judgment.

For a repeatable timing capture:

./scripts/acestep-performance-proof.sh

The installed one-second, one-step XL-Turbo proof completed in 3.14 seconds, including process startup and checkpoint load, with 21.35 GB maximum RSS, 25.35 GB peak footprint, and zero swap. It produced PCM24 stereo at 48 kHz with SHA-256 a5d5a265488373d9b32beabd1ccdf5c7e99cc4d159b08a1a8aaa006ef12bdf1f and a matching schema 2 recipe.

Do not treat a source-level test, generated recipe, or unchecked listening playlist as installed-model proof. Each evidence layer answers a different question.