Serveurperso
/

YuE2-GGUF

@@ -107,7 +107,7 @@ VAE, Oobleck decoder               1920x upsample -> 48 kHz stereo
 One backbone GGUF holds the two halves of a single Qwen3 transformer: the same 28 layers with two sets of attention projections and MLPs, one to write tokens, one to paint latents, sharing the embeddings and the final norm. The AR half works like a language model: token by token, it first writes the ABC score, a symbolic plan in plain text you can read and edit, then the semantic codes, one per 40 ms frame, and every token it processes lands in the KV cache. The NAR half is the same network used the other way round: it starts from Gaussian noise for every frame of the song, attends on the cache the AR half just left, and refines all the frames together with a midpoint flow matching solver, 32 steps of two evaluations, from noise to latents. The VAE turns the latents into sound, 1920 samples per frame.
-Only one module is in VRAM at a time. The AR half is evicted once the codes are written, the NAR half loads, is evicted in turn, and the VAE loads; the KV cache stays through all of it, so the halves trade places around it and nothing is recomputed. `--keep-loaded` keeps everything resident on a card with the budget.
 The ABC score is the white box of the pipeline: read it, edit a chord or a bar, hand it back, and the model realizes your version instead of writing a new one. A track also comes back with its semantic codes, and feeding them back re-renders it without the autoregressive stage, so the flow matching steps, the noise seed or the output format can be iterated for a fraction of the cost.

 One backbone GGUF holds the two halves of a single Qwen3 transformer: the same 28 layers with two sets of attention projections and MLPs, one to write tokens, one to paint latents, sharing the embeddings and the final norm. The AR half works like a language model: token by token, it first writes the ABC score, a symbolic plan in plain text you can read and edit, then the semantic codes, one per 40 ms frame, and every token it processes lands in the KV cache. The NAR half is the same network used the other way round: it starts from Gaussian noise for every frame of the song, attends on the cache the AR half just left, and refines all the frames together with a midpoint flow matching solver, 32 steps of two evaluations, from noise to latents. The VAE turns the latents into sound, 1920 samples per frame.
+Only one module is in VRAM at a time. The AR half is evicted once the codes are written, the NAR half loads, is evicted in turn, and the VAE loads; the KV cache stays through all of it, so the halves trade places around it and nothing is recomputed, and once the track is out the cache goes too, nothing stays on the GPU between two requests. `--keep-loaded` keeps everything resident on a card with the budget.
 The ABC score is the white box of the pipeline: read it, edit a chord or a bar, hand it back, and the model realizes your version instead of writing a new one. A track also comes back with its semantic codes, and feeding them back re-renders it without the autoregressive stage, so the flow matching steps, the noise seed or the output format can be iterated for a fraction of the cost.