YuE-s2-1B-general — SceneWorks rehost (bf16 / q8 / q4)
A redistribution mirror of m-a-p/YuE-s2-1B-general at revision 9dfa90b7013f6b5e7eb5eb2991620dca33058a0e, the YuE
stage-2 (1B codebook upsampler) from HKUST / M-A-P, published by SceneWorks so the weights resolve by an immutable
commit SHA for SceneWorks Inference's candle YuE engine. It is not an official M-A-P
distribution.
Layout
| Dir | Contents |
|---|---|
bf16/ |
the upstream safetensors snapshot, unmodified, plus a derived tokenizer.json |
q8/ |
pre-quantized Q8 tier (2.41 GB): projections stored as GGML Q8_0 ×224 blocks |
q4/ |
pre-quantized Q4 tier (1.60 GB): projections stored as GGML Q4_0 ×32 + Q4_K ×192 blocks |
Each tier directory is self-contained (config, tokenizer, weights) — download only the tier you
use. The q8/q4 tiers are produced from bf16/ by candle-llm's snapshot preparer
(prepare_snapshot): every attention/MLP projection is quantized once and stored as its raw GGML
blocks — a U8 tensor [rows, blocks_per_row, block_bytes], the block type given by
block_bytes (18 = Q4_0, 34 = Q8_0, 144 = Q4_K) — which the candle loader rebuilds directly
(no dequantize/re-quantize). config.json carries quantization: {bits, storage: "ggml"}.
Embeddings, the LM head and norms stay bf16 in every tier. The xcodec codec + Vocos decoders are not tiered
(approved carve-out); every tier pairs with
SceneWorks/xcodec-mini-infer.
tokenizer.model is the upstream mm SentencePiece model (the source of truth). tokenizer.json
is derived from it (byte-fallback BPE + the mm special tokens at their ids) and verified id-for-id
against the upstream _MMSentencePieceTokenizer over 3010 cases
(0 mismatches).
Vocabulary width 83840, context 8192.
Provenance
SOURCE_REVISION.json records the upstream repo + revision, the YuE code commit, and the sha256
of every file. Prepared by scripts/audio/prepare_yue_assets.py in the SceneWorks Inference
repository.
License
Apache-2.0, © 2025 Ruibin Yuan and core contributors from M-A-P and HKUST. See LICENSE and
NOTICE (retained per Section 4(d) of the Apache License). Upstream project:
https://github.com/multimodal-art-projection/YuE