LeVo2-GGUF

LeVo2 v2-medium and v2-large GGUF

This repository contains v2-medium and v2-large LeLM GGUFs, shared v0.2 renderer components, and their native low-bit catalog. Select one LeLM scale and one Flow precision independently. The Flow and VAE are deliberately shared between medium and large: SongGeneration v2-large changes only the LeLM scale, not the separate-token Flow checkpoint or Stable Audio 1920 decoder.

Each GGUF has a standard .sha256 sidecar and a deterministic .gguf.manifest.json conversion inventory. Verify every downloaded component before loading it.

License and permitted use

Research, academic, and education use only. Commercial and production use are prohibited. The converted weights and inference-enabling implementation remain subject to the upstream Tencent SongGeneration terms. Read LICENSE and THIRD_PARTY_NOTICES.md; this card grants no additional rights.

Native lyrics-to-WAV use

Build the matching source repository with CUDA:

git clone --recurse-submodules https://github.com/ckadirt/LeVo2.cpp.git
cd LeVo2.cpp
cmake -S . -B build-cuda -DCMAKE_BUILD_TYPE=Release \
  -DGGML_CUDA=ON -DGGML_CUDA_NCCL=OFF -DCMAKE_CUDA_ARCHITECTURES=89
cmake --build build-cuda -j

Generate tokens and render them without Python. This example uses the balanced v2-large Q6_K pairing; choose the matching v2-medium file instead when that smaller LeLM is preferred, and replace either selected precision tier when you want a different memory/quality trade-off:

./build-cuda/bin/levo-cli \
  --model LeVo2-v2-large-Q6_K.gguf \
  --lyrics lyrics.txt --prompt "female, pop" --duration 30 \
  --output tokens.npy --backend cuda --seed 1235

./build-cuda/bin/levo-render tokens.npy \
  --flow-model LeVo2-v2-flow-Q6_K.gguf \
  --vae-model LeVo2-v2-vae-F16.gguf \
  --output song.wav --backend cuda --steps 50 --cfg 1.5 --seed 1234

The renderer writes stereo 48 kHz IEEE-F32 WAV and crops output to exactly token_frames * 1920 samples per channel. The native loader verifies every quantized tensor type, policy revision, and Flow physical padded layout before running it. The F16 VAE has its own source-artifact lineage and is loaded only when its precision tags are present; decoder operators promote its stored F16 weights to F32 for execution.

Quantization policy and selection

The Q5_K_M and Q4_K_M files deliberately retain sensitive matrices at Q6_K: LeLM retains embeddings, output heads, conditioner tables, bridge weights, attention V/output, and FFN-down projections; Flow retains attention/FFN output projections. Flow controls, embeddings, normalizers, modulation tables, biases, and final output remain F32. This is a fixed, strict policy revision, not runtime dequantization or a name-only precision hint.

Use the F16 LeLM and F32 Flow files when fidelity is the priority. Q6_K is the best deterministic Flow match in the frozen comparison below. Q8_0, Q5_K_M, and Q4_K_M are available as explicit capacity/quality trade-offs; Q4_K_M is the smallest and has the largest measured deviation. The source repository's docs/quantization-implementation-report.md and docs/quantization-validation-matrix.json contain the complete policy and reproducible identities.

Provenance

The Flow conversion emits 231 runtime tensors / 663,310,785 parameters. The decoder-only VAE conversion folds weight normalization and emits 145 tensors / 84,395,776 parameters. Training-only encoders, discriminators, and maintenance state are classified and omitted in their manifests.

Validation

LeVo2.cpp v0.2 passes strict loaders, CPU/CUDA tests, stage-level Flow/VAE parity, and an eight-case token-to-WAV matrix against the pinned official Python oracle. Cases cover 50, 252, 750, and 1,250 token frames at one and fifty Euler steps, including the second 750-frame hop and 250-frame overlap.

The worst measured release values are:

A public-path 30-second smoke generated [3,750] tokens and a finite, non-silent stereo WAV with exactly 1,440,000 samples per channel. Full commands, fixture hashes, timings, GPU memory, deviations, and limitations are recorded in the source repository's docs/renderer-implementation-report.md and docs/renderer-release-matrix.json.

Each low-bit file passed its exact sidecar checksum, strict native loader, and real CUDA smoke. Every low-bit Flow file also rendered a finite two-second, 48 kHz stereo WAV through the native quantized graph on CUDA; Q6_K/Q5_K_M/ Q4_K_M additionally passed CPU renders (Q8_0 had already done so). Against the same F32 Flow output with fixed tokens, noise, one Euler step, and seed, the measured waveform SNR/correlation was:

Every v2-medium tier and every v2-large tier completed its strict CUDA load and greedy token smoke. Quantization changes near-tied token choices, so neither token IDs nor audio are claimed bit-identical to F16/F32. The v2-large source repository records its immutable inputs, 452-tensor inventory, all five artifact identities, CUDA timings, a Q6_K token-to-WAV gate, and a 120-second F16 song in docs/v2-large-validation-matrix.json. Verify file bytes with the supplied sidecars before loading.

The F16 VAE passed its independent 1,000-frame full-window storage comparison against a native F32 VAE reference on CPU and CUDA. CUDA measured 2.65128e-3 maximum error, 6.56119e-5 RMSE, and 0.999999914 correlation; CPU measured 2.65826e-3, 6.56703e-5, and 0.999999914. The decoder graph remains F32, so this artifact reduces stored-weight size rather than claiming an F16 activation/accumulation path.

Limitations

• The VAE F16 artifact is a storage optimization: its decoder graph is F32. Low-bit LeLM/Flow files are approximate inference tiers, not F16/F32 parity claims; quality depends on the lyric, prompt, seed, and sampler settings.
• Input is lyrics plus an optional style description; audio-prompt encoding and source separation are not implemented natively.
• PyTorch/CUDA and GGML can make different near-tied token choices for prompts outside the frozen LeLM parity fixture.
• Generated music can contain artifacts or inappropriate content. Respect copyright, privacy, and applicable law, and do not use it for high-stakes decisions.