Repository navigation
Releases
sa3.cpp v0.2.4
sa3.cpp v0.2.4
Pre-release for gary4local v0.7.3.
Fixed
sa3-trainpre-encodes tracks longer than 600 seconds to their end. Pre-encode laid every file on the PyTorch reference's 600 s grid and dropped the rest without a message, so a 613 s track trained as599.9s -> 6459 frames. A longer file now gets a grid of its own length, and each step's random crop is drawn from anywhere in it. There is no need to split long files. (#97)- Tracks that fit 600 s are encoded bit-identically and their cached latents are reused.
- A cached entry for a longer track is re-encoded once: the cache key now also checks the frame count, which is what tells a cropped entry from a whole one.
seconds_totalfor a long track is its real length. The seconds conditioner clamps it to its range, as it already did.
Changed
sa3-serverwithout--train-portexplains an empty training tab./training/answers 503 with how to start the training server, where it used to 404.
Validated: Windows, CUDA, medium Q4_K_M base, a 612.97 s file with 20 s crops, against unmodified v0.2.3 sources on fresh caches; CI on Linux, macOS and the CUDA compile.
Not yet validated: Vulkan, Metal, and a listening comparison of an adapter trained on a long track.
The C ABI is unchanged. The ggml pin is unchanged (e3db5210), so this release packages the same attested ggml-cuda.dll as v0.2.3. Packaging is the same as v0.2.3: Windows has the split core/CUDA/Vulkan and standalone ZIPs, and macOS the signed, notarized universal package.
sa3 v0.2.3 Linux Vulkan preview
Experimental Linux Vulkan packages for sa3 v0.2.3, preview 1.
Promoted byte-for-byte from the successful native x64/ARM64 CI run.
Source: 65ec932feda13b986d6c3d2ea03b7f7e0c1e506f; pinned ggml: e3db5210fe3ecdaa4d65b4c93d515f2cf15524b7. See LINUX-PREVIEW.json and each core's
BUILD-INFO.json for exact toolchains and ELF requirements. Extract core plus the
matching Vulkan overlay. Verify SHA256SUMS before use.
Requires a supported Vulkan GPU, system loader and driver. CPU generation is not
a validation target. Model weights and CUDA user libraries are not included.
The GB10-specific CUDA smoke builds are not portable CUDA releases. Hosted x64
checks establish packaging, not x64 GPU inference. This preview does not change
existing Windows/macOS assets or mark a new latest release.
sa3.cpp v0.2.3
sa3.cpp v0.2.3
Pre-release for gary4local validation. It becomes a full release once that passes.
Fixed
- Envelope points that share a time are a jump, in the order given. The sampler sorted
(time, amount)pairs by both values, so two points at one time came out lowest amount first. A drawn drop, peak then 0 at the same moment, became a dip to 0 followed by a ramp down from the peak to the next point, and the LoRA kept playing for bars after it was drawn to stop. Envelopes are now sorted by time only, and at a tied time the curve reads the later point. This applies tolora_envelopeand everylora_lanesenvelope, whether it arrives through the C ABI, the server's JSON or the CLI; the web UI's lane editor reads tied points the same way. Renders whose envelopes contain a vertical jump sound different; every other envelope renders as before. (#92)
Downstream
The sa3-iplug2-demo and the sa3 Ableton extension should be rebuilt against this release to pick up the fix. gary4juce already works around it by sending tied points 1 ms apart, and is unaffected either way.
The C ABI is unchanged. The ggml pin is unchanged (e3db5210), so this release packages the same attested ggml-cuda.dll as v0.2.2. Packaging is the same as v0.2.2: Windows has the split core/CUDA/Vulkan and standalone ZIPs, and macOS the signed, notarized universal package.
sa3.cpp v0.2.2
sa3.cpp v0.2.2
Pre-release for gary4local validation. It becomes a full release once that passes.
Fixed
- Negative prompts under CFG. stable_audio_3 gives the negative prompt's masked tokens the null embedding in the unconditioned pass; sa3.cpp gave them the padding embedding, so with a short negative prompt like
"low quality"nearly the whole unconditioned context was wrong. Those rows are now zeroed, only when a negative prompt is given. Renders withcfg_scale != 1and a negative prompt sound different; at cfg 1, or without a negative prompt, nothing changes. Against the PyTorch reference, the final latent's error drops from 3.4% to 0.24%. (#91) - Linux builds.
sa3-lib-sat-smokenow links libm on non-Apple Unix, so./build.shcompletes on glibc. (#91)
Added
lora_lanescapability./health.capabilitiesand--control-inforeport"lora_lanes": true, so a host can tell a lanes-aware server from one that would silently ignore automation. gary4juce's remote SA3 reports the same flag. (#91)
Changed (since v0.2.0)
- ggml moves to upstream GGML 0.29. The submodule pins
e3db5210(ggml#22), up fromb77ae790(0.27). The fork's training kernels and backward ops carry over. (#90, first in the v0.2.1 pre-release) - CUDA and Vulkan renders change; Metal's don't. On CUDA, the same seed gives a slightly different take: one upstream commit sums RMS norm in a different order, so it is rounding, not a different computation. In a blind A/B no difference was audible. On Metal, every render and training output is byte-identical to v0.2.0.
- Plugin builds trim source paths with
SA3_PRIVATE_GGML(#89). This is off for these release packages, so they are unaffected. - The Windows CUDA backend is built once per ggml pin (#88). This release packages the
ggml-cuda.dllthat dry run 42518070848 compiled and attested;BUILD-INFO.jsonnames that run.
The C ABI is unchanged. Packaging is the same as v0.2.0: Windows has the split core/CUDA/Vulkan and standalone ZIPs, and macOS the signed, notarized universal package.
sa3.cpp v0.2.1
sa3.cpp v0.2.1
Pre-release for gary4local validation of the GGML 0.29 refresh. It becomes a full release once that passes.
- ggml moves to upstream GGML 0.29. The submodule pins
e3db5210(ggml#22), up fromb77ae790(0.27). The fork's training kernels and backward ops carry over. It also adds CUDA, Vulkan and Metal fixes for matmuls that pin an F32 input, and Metal cross-entropy. (#90) - CUDA and Vulkan renders change; Metal's don't. On CUDA, the same seed now gives a slightly different take. All of the change comes from one upstream commit that sums RMS norm in a different order, so it is rounding, not a different computation. In a blind A/B no difference was audible. Vulkan takes change too. On Metal, every render and training output is byte-identical to v0.2.0.
- Training. F16 DoRA, resume, medium Q4 DoRA, LoRA and libtrain all run on CUDA, Vulkan and Metal.
- Plugin builds trim source paths. With
SA3_PRIVATE_GGML, builds now strip the build folder from the paths ggml's asserts compile in (#89). This is off for these release packages, so they are unaffected.
The C ABI is unchanged. Packaging is the same as v0.2.0: Windows has the split core/CUDA/Vulkan and standalone ZIPs, and macOS the signed, notarized universal package.
sa3.cpp v0.2.0
sa3.cpp v0.2.0
- Automation lanes in the C ABI.
sa3_lane_v1,lane_initandadapter_infolet embedded hosts draw lanes for Generate, Transform and Continue on SA3 variants. SAT variants returnINVALID_ARGUMENT. No fields moved andsizeof(sa3_api_v1)is unchanged, but a library that predates lanes ignores the new request fields, so hosts should check thatlane_initisn't null.request_initsets the recommended blend. The runtime version islibsa3 9. - Automation tab. Import a LoRA and its dice captions from the tab. It now defaults to the shared-structure blend, and the Advanced text explains which steps each setting covers. Server and CLI defaults don't change, so API callers get the same audio.
- Decoder adapters survive frugal renders. Only the first Generate on a reused pipeline got its autoencoder adapter, and later ones silently rendered without it. If you use a decoder adapter in frugal mode, including in the studio, takes after the first now sound different because the correction is applied.
- Non-ASCII paths on Windows. Everything reachable through the V1 ABI now takes UTF-8 paths, so models and adapters load from folders like
C:\Users\José\.... Training paths, debug dumps and the CLI tools still use narrow paths.
Packaging is the same as v0.1.8: Windows keeps the split core/CUDA/Vulkan and standalone ZIPs, and macOS keeps the signed, notarized universal package.
sa3.cpp v0.1.8
sa3.cpp v0.1.8
- Parity with stable_audio_3. The DiT now runs its local-cond block on zeros, as stable_audio_3 does for plain generation, and adapters' DoRA on the
seconds_totalconditioner is converted and applied. With the same seed and adapter, a clip matches the PyTorch service: gary4juce's remote take reproduces locally at waveform cos 0.968 (CUDA) and 0.963 (Vulkan on an Intel iGPU). - torch-cuda noise is the default, the noise PyTorch draws on the DGX Spark that serves gary4juce's remote SA3. Every seed's output changes from v0.1.7.
--rng mt19937(or"rng": "mt19937") keeps the old noise. - Re-convert adapters made before this release to pick up the seconds-conditioner adapter. The local-cond fix applies without re-converting.
- Adapter names resolve to the file
/loraslists. Previouslykoancould silently runkoan-27-2. - LoRA strength automation: per-frame strength envelopes, lanes with their own prompts, the Studio's Automation tab (beta), and lane isolation and shared-structure options.
- Training: warm-start from a saved adapter with
--init-adapter. - GGML refreshed to upstream 0.27 (shared fork
b77ae790). Each build reproduces its own outputs exactly; the changed arithmetic gives a different take at a given seed, with no audible change in a headphone A/B. Medium Q4 LoRA training on Vulkan works at this pin.
Windows keeps the split core/CUDA/Vulkan and standalone ZIPs, with the Visual C++ runtime in the core; macOS keeps the signed, notarized universal package. In a standalone with both GPU backends, SA3_DEVICE only chooses between CPU and GPU: select a GPU with SA3_GPU (an index or part of the device name, e.g. SA3_GPU=nvidia).
sa3.cpp v0.1.7
sa3.cpp v0.1.7
- Jerry Grunge comes from its step-320 checkpoint. SAOS's
jerry-grungevariant is nowjerry_encoded_bs128_epoch=106-step=320.ckpt(batch size 128, learning rate 1e-4), our favorite of thepatch/jerry_grunge's checkpoints.- The DiT is
finetunes/jerry-grunge/jerry-grunge-bs128-step320-dit-0.3B-v1.0-<tier>.ggufat F16, Q8_0, Q5_K_M and Q4_K_M, in thepatch/stable-audio-open-small-GGUF. - sat-server's jerry mode selects it by
finetune_repo: "thepatch/jerry_grunge"andfinetune_checkpoint: "jerry_encoded_bs128_epoch=106-step=320.ckpt". - The downloaders (
tools/download_models.py,models.cmd,models.sh) fetch it with--saos-variant jerry-grunge.
- The DiT is
- The previous Jerry Grunge conversion is withdrawn. It came from an overfit checkpoint, and its files are no longer in the GGUF repository, so v0.1.6 and earlier can no longer download Jerry Grunge.
- Quantization: Q8_0 tracks F16 closely. At some seeds, Q5_K_M and Q4_K_M take a different take on a prompt than F16. See STABLE_AUDIO_OPEN_SMALL.md.
Everything else is unchanged from v0.1.6: ARC, KickBass, Foundation-1, SAO 1.0, SA3 and sat-server's routes. Windows keeps the split core/CUDA/Vulkan and standalone ZIPs; macOS keeps the signed, notarized universal package.
sa3.cpp v0.1.6
sa3.cpp v0.1.6
- One SAOS
sat-serverserves the base model and its finetunes.--model saosserves ARC and every published finetune (kickbass, jerry-grunge) installed at its encoding.- One DiT stays resident. The variants share the T5 and autoencoder, so switching swaps only the DiT, which cost about 0.2 s on an RTX 5070.
- SAOS routes keep gary4local's Python stable-audio ("jerry") contract, which gary4juce's jerry tab drives:
- A request selects a model with
model_type,finetune_repoandfinetune_checkpoint. /generateand/generate/loopanswer synchronously with{success, audio_base64, metadata}, or with the WAV itself./models/statuslists the installed variants under Python's cache keys./models/promptsserves a finetune'sprompts.jsonprompt bank.- The finetunes honor the
euler/dpmppsampler choice, which the Python service ignored.
- A request selects a model with
- Jerry Grunge's prompt bank is now published beside its DiT in thepatch/stable-audio-open-small-GGUF.
Changed: SAOS no longer uses the asynchronous session_id//poll_status shape, and seconds is now seconds_total. --model arc, kickbass and jerry-grunge still start a SAOS server, with that variant active.
Foundation-1, SAO 1.0, SA3 and the model files are unchanged. Windows keeps the split core/CUDA/Vulkan and standalone ZIPs; macOS keeps the signed, notarized universal package. See SAT_SERVER.md and RUNTIME_RELEASE.md.
sa3.cpp v0.1.5
sa3.cpp v0.1.5
- Foundation-1
POST /randomizereturns a whole preset split into plugin
controls: family, timbre knobs, tags, effects and melody behavior. It keeps
gary4local's Python request and response shape and needs no model loaded.
Existing prompts for a given seed are unchanged. - Foundation-1
POST /audio2audiotransforms input audio toward a Foundation
prompt, with the Python service's request contract and host-tempo stretch.- It needs the new
stable-audio-open-oobleck-encoder-v1.0-F16.gguffrom
thepatch/foundation-1-GGUF.
Without it, generation is unaffected and the route answers 503. - The encoder matches stable-audio-tools' with cosine 1.0000.
- The whole transformation range works: with the gary profile, Python's service
returned NaN audio at transformations of 0.5 and below.
- It needs the new
- Windows packaging reads the Visual C++ runtime and compiler from the build's
CMake configuration, and proves the bundled runtime DLLs are the ones that load.
Existing models, adapters and APIs are unchanged. Windows retains split
core/CUDA/Vulkan and standalone ZIPs; macOS retains the signed, notarized
universal package. See SAT_SERVER.md and
RUNTIME_RELEASE.md.