audio.cpp GB10 CUDA qualification

This record captures the technical qualification for exact music.generate claims for the two music cards below on the nvidia:sm-12.1 hardware class. The class describes GPU compute compatibility, not an individual Skulk node. Both cards passed human listening and production-signed activation on this exact build and hardware class. Other CUDA builds and hardware classes need their own qualification.

Contract	Qualified value
Engine source	audio.cpp v0.8.2, revision 4d88768fbcae4e6eb3352c6ab1422dabb7d90b58
Engine lane	audio_cpp-cuda, Linux arm64, GB10 SM 12.1
Wheel	skulk_audio_cpp_cuda-0.8.2.post1-py3-none-manylinux_2_35_aarch64.whl
Wheel SHA-256	39d9f4e2f037ac808ba3588119e3d11a2437e3eb2d85e8bf5eecf63cc3c4c772
Executable SHA-256	21db727c1f2ec030b93cabffbd09e3c522ceb77fa2ae1092cd96a0f4b0120c9c
Node build identity	audio.cpp@sha256:21db727c1f2ec030b93cabffbd09e3c522ceb77fa2ae1092cd96a0f4b0120c9c
Package provenance	Attested build and immutable publication

Qualification used an isolated Skulk process, a test-signed catalog, and real GB10 hardware while the normal Skulk service kept its existing workloads. The engine and model caches began empty. Each normal mount selected the published wheel, verified its executable and model specs, downloaded the card's immutable artifact bundle, and reached runner readiness on audio_cpp-cuda with the build identity above. This tested signed discovery, engine preparation, placement, model loading, and the /v1/music output path together.

Card	Immutable model revision	First delivered WAV
audio-cpp/ACE-Step1.5-Turbo-BF16	a776907b362419343f4b9996bdd899619efcf3f8	10.00 seconds, stereo 48 kHz, 1,920,044 bytes, SHA-256 a5573d1fecf4a22ee7d82d0c8b65f9067dd7708fa87f37c6adb42e17fb622273
audio-cpp/MiniMax-Music3-GGUF-Q4	9634a1e1364f94f1ac85a38c114ef105c678f824	9.98 seconds, stereo 44.1 kHz, 1,761,324 bytes, SHA-256 1e0c0021a95f61c1bac978c7ab133f62d4a2d6ee21f88b49a757996a0a8e8035

Both WAVs matched the API's digest and metadata, were below the 64 MiB limit, and contained non-silent PCM samples. A MiniMax request without lyrics returned HTTP 400. Active and queued cancellation reached terminal cancelled for both families, and their runners returned to ready. Parallel ACE and MiniMax requests completed with digest-matched WAV content while both models were mounted. These signal measurements did not establish subjective music quality; the separate listening set below closed that gate.

Both models then completed concurrent, paced stability runs on the same isolated node. Every result was a stereo WAV with the expected sample rate and a matching API digest. Sampled PCM16 amplitude was non-silent in every job.

Card	Completed jobs	Elapsed paced run	Job latency	Minimum sampled PCM16 RMS
ACE-Step 1.5 Turbo BF16	8	1,845.2 seconds	9.1–15.2 seconds; 9.1-second median	4,590
MiniMax Music 3 Q4	8	2,077.9 seconds	39.3–42.3 seconds; 39.3-second median	1,221

Each job was followed by a 220-second idle interval. These runs establish repeated generation over time but do not satisfy the release plan's continuous 30-minute generation gate. A subsequent continuous run closed that technical gate. Both models were mounted concurrently and admitted back-to-back work on the same isolated reader. There were no intentional pauses between requests; every job completed with valid, digest-matched, non-silent WAV output.

Card	Jobs / wall time	Busy fraction / longest request gap	Latency, min–median–max	Minimum sampled PCM16 RMS	Server RSS, beginning / high-water / ending
ACE-Step 1.5 Turbo BF16	93 / 1,808.78 seconds	99.93% / 0.029 seconds	11.05 / 20.08 / 34.12 seconds	2,159	284,508 / 1,458,764 / 1,082,408 KiB
MiniMax Music 3 Q4	31 / 1,804.79 seconds	99.97% / 0.033 seconds	49.19 / 58.20 / 63.21 seconds	1,373	269,736 / 2,520,672 / 920,608 KiB

The middle RSS value is the Linux process high-water mark; the sampled RSS peaks were 1,353,332 KiB for ACE-Step and 2,520,672 KiB for MiniMax. A separate per-process NVIDIA GPU-memory sampler observed 14,455–15,729 MiB for ACE-Step and 1,803–8,403 MiB for MiniMax across 325 successful five-second samples. Sampling began about three minutes after generation began, so these are observed ranges rather than lifetime GPU-memory peaks. Both processes remained alive through the run; the normal Skulk service retained its three existing placements.

In the earlier paced runs, the largest sampled resident set of either mounted audio.cpp server was 1,200,404 KiB during the ACE run and 1,259,520 KiB during the MiniMax run. Those peaks are not attributed to an individual model. Across 70 samples, aggregate CUDA memory used by the shared host peaked at 86.48 GiB, while host-available RAM stayed at or above 62.96 GiB. These aggregate readings include workloads in the unchanged normal Skulk service. GPU busy reached 0.95 during music generation.

After the paced runs, the isolated API restarted at the final reader code with the signed catalog server unavailable. Previously completed jobs for both cards retained their metadata and WAV bytes with identical SHA-256 digests. Both cards remained discoverable and remounted from the verified engine and model caches; the live node still reported the exact CUDA build and nvidia:sm-12.1 class. Both cards generated fresh digest-matched, non-silent WAVs with the catalog server unavailable. An additional restart with Skulk's --offline mode kept both cards and the verified engine ready, and both remounted models generated valid WAVs without a registry or package download. Killing the isolated API during an active MiniMax job and restarting offline marked that job failed with no content; previously completed WAVs remained available with their original digests. The three preexisting placements in the normal Skulk service remained present throughout qualification.

The GB10 CUDA memory pool is shared with host RAM. On this hardware NVML reported memory as unsupported, while CUDA measured a 130,595,991,552-byte pool and a live free figure. Skulk used the lower of CUDA free memory, host free memory less 16 GiB of OS headroom, and 75% of total host RAM for music admission. A failed CUDA memory query leaves the lane without a usable memory budget.

The exact claim requires the immutable card, music.generate, the build identity above, and nvidia:sm-12.1. A generic NVIDIA class or another audio.cpp executable does not inherit this result. The package needs host CUDA 12 runtime, cuBLAS, NCCL, and NVIDIA driver libraries; the binary probe reports missing libraries before advertising the lane.

Music quality and production-signed activation

Each card generated eight reviewed outputs: four different music prompts and lyric settings at two seeds, including longer requests. ACE-Step produced two 25-second results; MiniMax produced one 22.18-second result and one additional 24.96-second result with longer lyrics. One MiniMax request targeted 25 seconds but ended at 9.15 seconds, consistent with the API's generation-target contract. All reviewed outputs were accepted as music. The listener preferred ACE-Step's quality in this set; acceptance does not imply equal quality.

The registry then published exact supported claims for music.generate, audio_cpp-cuda, the executable build above, and nvidia:sm-12.1:

Card	Support claim
ACE-Step 1.5 Turbo BF16	support_y5ukbrwmdmmcl53pgwfqzaj5rpe2c7gzlgchg6csxd2nz4o56paq
MiniMax Music 3 Q4	support_lnm5obsbwkld4icrzsi7ccviwvxlj2l6vpa45dz4lyvmz2cvir2a

A stock TUF reader verified signed snapshot snapshot_141_309aa14cac11c76a45238f5c with both claims. The normal three-node Skulk service discovered both cards, prepared the pinned engine package on demand, advertised the matching live build, and mounted each card to RunnerReady on the GB10. A MiniMax placement explicitly selected this CUDA hardware class because unconstrained placement also had a supported Metal option. Both ordinary /v1/music requests returned non-silent stereo WAV content matching their job metadata and SHA-256 digests:

Card	Actual duration	Sample rate	Bytes	WAV SHA-256
ACE-Step 1.5 Turbo BF16	10.00 s	48 kHz	1,920,044	c6d0b4b0b07256e8444e25805034fb11e78dd492b0b0a1bda3e4f980698412a3
MiniMax Music 3 Q4	9.985 s	44.1 kHz	1,761,324	00680569c79bebef0f8d2a93dd61dd7d959336adea577e4661148b89ff731f1a

The normal-service MiniMax mount reused a verified model cache. A separate isolated Skulk process then repeated that exact production-signed path with empty engine and model caches. It read snapshot 141, prepared and verified the pinned CUDA package, downloaded the complete MiniMax bundle, reached DownloadCompleted and RunnerReady, and returned a 9.985-second stereo 44.1 kHz WAV of 1,761,324 bytes. The content matched the job's SHA-256 d9f6390c6294f0c05f0dbbaa5c92c37f7ca5eafceade18db0048ce7b5fa48e78 and contained non-silent PCM. This closes the signed clean-install gate for MiniMax on the exact GB10 class.

The test jobs and placements were removed. The three nodes finished healthy with their two original runners ready. This qualification is for these exact cards, build, and hardware class; it does not extend to every Linux arm64 CPU or NVIDIA CUDA combination.