Switch to musicgen-medium for cleaner continuations

musicgen-small's audio continuations degrade into noise after a chunk or two: text-only chunks stay clean (spectral flatness ~0.1) but every audio-seeded continuation is ~3x noisier (~0.3) regardless of guidance_scale or seed normalization. Swap to musicgen-medium, which continues more cleanly. Model id is now the LOFINITY_MUSICGEN env var (default facebook/musicgen-medium). Widen the @spaces.GPU duration budget to 40 + 40*chunks and bump the progress estimate, since medium renders slower and an overrun gets the brew KILLED mid-render.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app.py

@@ -70,6 +70,10 @@ print(
 )
 
 ENGINE = os.getenv("LOFINITY_ENGINE", "musicgen")
+# musicgen-medium continues from an audio seed more cleanly than -small, whose
+# continuations degrade into noise after a chunk or two. Override via the env to
+# fall back to facebook/musicgen-small (faster, smaller) if needed.
+MUSICGEN_MODEL = os.getenv("LOFINITY_MUSICGEN", "facebook/musicgen-medium")
 # 30s is musicgen-small's single-shot max (1500 tokens). Longer tapes are
 # stitched from 30s chunks: each one re-seeds the model with the last OVERLAP_S
 # of the track so it keeps playing from there. musicgen-small's context is 2048
@@ -86,7 +90,7 @@ OLLAMA_MODEL = os.getenv("OLLAMA_MODEL", "llama3.2:3b")
 # read real per-chunk progress from it; it reports a time-based estimate from this
 # rough per-length budget instead (tunable; only affects the cosmetic brew bar).
 GPU_WARMUP_S = 5.0  # enrichment + cold-start/queue allowance before audio flows
-GPU_SECS_PER_CHUNK = 10.0  # rough GPU render time per 30s chunk
+GPU_SECS_PER_CHUNK = 25.0  # rough GPU render time per 30s chunk (musicgen-medium)
 
 app = Server(title="LoFinity")
 
@@ -267,7 +271,8 @@ _musicgen_lock = threading.Lock()
 
 
 def load_musicgen():
-    """Lazy-load MusicGen once; first call downloads ~2.3 GB from the Hub."""
+    """Lazy-load MusicGen once; the first call downloads the model from the Hub
+    (musicgen-medium is ~3.5 GB)."""
     global _musicgen
     with _musicgen_lock:
         if _musicgen is None:
@@ -284,11 +289,9 @@ def load_musicgen():
                 device = "mps"
             else:
                 device = "cpu"
-            print(f"[lofinity] loading musicgen-small on {device}…")
-            processor = AutoProcessor.from_pretrained("facebook/musicgen-small")
-            model = MusicgenForConditionalGeneration.from_pretrained(
-                "facebook/musicgen-small"
-            )
+            print(f"[lofinity] loading {MUSICGEN_MODEL} on {device}…")
+            processor = AutoProcessor.from_pretrained(MUSICGEN_MODEL)
+            model = MusicgenForConditionalGeneration.from_pretrained(MUSICGEN_MODEL)
             model.to(device)
             model.eval()
             _musicgen = (processor, model, device)
@@ -422,7 +425,10 @@ def _gpu_budget(prompt: str, seconds: int = CHUNK_S) -> int:
     better queue priority; the signature must mirror gpu_brew so ZeroGPU can pass
     it the same args."""
     chunks = max(1, round(int(seconds) / CHUNK_S))
-    return 25 + 25 * chunks  # enrichment + 30s->50, 60s->75, 90s->100
+    # musicgen-medium renders slower than -small, so budget generously — a brew
+    # that overruns its @spaces.GPU duration is KILLED mid-render; over-budgeting
+    # only costs quota, under-budgeting loses the tape.
+    return 40 + 40 * chunks  # 30s->80, 60s->120, 90s->160
 
 
 @spaces.GPU(duration=_gpu_budget)