Harden composer + share video against the failure modes in production logs

app.py

@@ -496,7 +496,7 @@ Given a sound instruction and/or lyrics, produce global_metadata, vocal_details
 Answer with ONLY a JSON object with keys: global_metadata, vocal_details, arrangement."""
 
 
-def _llm_json(system, user):
+def _llm_json(system, user, required=()):
     import json as _json
 
     from openai import OpenAI
@@ -506,25 +506,35 @@ def _llm_json(system, user):
     last_error = None
     # Three DISTINCT providers, all verified enabled for this account (bare/":fastest" can route to
     # together, which 403s here and killed the fallbacks). Timeouts sized to measured composer latency.
-    for model, timeout in (
-        ("deepseek-ai/DeepSeek-V4-Flash-0731:baseten", 45),
-        ("deepseek-ai/DeepSeek-V4-Flash-0731:deepinfra", 75),
-        ("deepseek-ai/DeepSeek-V4-Flash-0731:novita", 100),
-    ):
-        try:
-            completion = client.with_options(timeout=timeout).chat.completions.create(
-                model=model,
-                messages=[{"role": "system", "content": system}, {"role": "user", "content": user}],
-            )
-            text = completion.choices[0].message.content or ""
-            # Tolerate fences/preambles and reject truncated replies: parse the outermost {...} span.
-            start, end = text.find("{"), text.rfind("}")
-            if start == -1 or end <= start:
-                raise ValueError(f"no JSON object in composer reply (finish_reason={completion.choices[0].finish_reason})")
-            return _json.loads(text[start : end + 1])
-        except Exception as e:
-            print(f"composer attempt failed ({model}): {type(e).__name__}: {e}", flush=True)
-            last_error = e
+    # Two passes over the chain: under load every provider can 429 transiently, and a second pass a few
+    # seconds later usually lands.
+    for attempt in range(2):
+        for model, timeout in (
+            ("deepseek-ai/DeepSeek-V4-Flash-0731:baseten", 45),
+            ("deepseek-ai/DeepSeek-V4-Flash-0731:deepinfra", 75),
+            ("deepseek-ai/DeepSeek-V4-Flash-0731:novita", 100),
+        ):
+            try:
+                completion = client.with_options(timeout=timeout).chat.completions.create(
+                    model=model,
+                    messages=[{"role": "system", "content": system}, {"role": "user", "content": user}],
+                )
+                text = completion.choices[0].message.content or ""
+                # Tolerate fences/preambles and reject truncated replies: parse the outermost {...} span.
+                start, end = text.find("{"), text.rfind("}")
+                if start == -1 or end <= start:
+                    raise ValueError(f"no JSON object in composer reply (finish_reason={completion.choices[0].finish_reason})")
+                data = _json.loads(text[start : end + 1])
+                # Valid JSON with the wrong shape must retry too, not KeyError later.
+                missing = [key for key in required if key not in data]
+                if missing:
+                    raise ValueError(f"composer reply missing keys: {missing}")
+                return data
+            except Exception as e:
+                print(f"composer attempt failed ({model}, pass {attempt + 1}): {type(e).__name__}: {e}", flush=True)
+                last_error = e
+        if attempt == 0:
+            time.sleep(3)
     raise gr.Error(
         "The composer model is overloaded right now — try again in a moment, "
         "or write the lyrics and structured prompt directly in the Studio tab."
@@ -534,7 +544,11 @@ def _llm_json(system, user):
 def compose_song(description, duration):
     if not description.strip():
         raise gr.Error("Describe the song you want first.")
-    data = _llm_json(_COMPOSER_SYSTEM, f"Song description: {description}\nTarget duration: {int(duration)} seconds.")
+    data = _llm_json(
+        _COMPOSER_SYSTEM,
+        f"Song description: {description}\nTarget duration: {int(duration)} seconds.",
+        required=("lyrics", "global_metadata", "vocal_details", "arrangement"),
+    )
     return data["lyrics"], data["global_metadata"], data["vocal_details"], data["arrangement"]
 
 
@@ -977,7 +991,11 @@ def render_video(wav_path, title):
 
     import scipy.io.wavfile
 
-    sr, wave = scipy.io.wavfile.read(wav_path)
+    title = " ".join((title or "").split())[:96] or "MiniMax Music 3"
+    try:
+        sr, wave = scipy.io.wavfile.read(wav_path)
+    except (FileNotFoundError, OSError):
+        return gr.skip()  # the visitor left and gradio cleaned the cached wav
     mono = wave.astype(np.float32).mean(axis=1) / 32768.0
     fps, size, bars = 24, 720, 56
     total_frames = int(len(mono) / sr * fps)
@@ -1053,7 +1071,10 @@ def render_video(wav_path, title):
         if len(chunk) < window:
             chunk = np.pad(chunk, (0, window - len(chunk)))
         spectrum = np.abs(np.fft.rfft(chunk * np.hanning(window)))
-        levels = np.array([spectrum[(freqs >= band_edges[b]) & (freqs < band_edges[b + 1])].mean() for b in range(bars)])
+        levels = np.array([
+            spectrum[m].mean() if (m := (freqs >= band_edges[b]) & (freqs < band_edges[b + 1])).any() else 0.0
+            for b in range(bars)
+        ])
         levels = np.log1p(12 * np.nan_to_num(levels))
         smooth = np.maximum(levels, smooth * 0.85)
         frame = base.copy()
@@ -1524,6 +1545,7 @@ def compose_assist(raw_state, duration):
             f"Current structured prompt, keep the lyrics coherent with it:\n"
             f"Global metadata: {state['global_meta']}\nVocal details: {state['vocals']}\n"
             f"Arrangement: {state['arrangement']}\nTarget duration: {int(duration)} seconds.",
+            required=("lyrics",),
         )
         state["lyrics"] = data["lyrics"]
         return state, "Lyrics written — tweak them, or press Generate."
@@ -1532,6 +1554,7 @@ def compose_assist(raw_state, duration):
             _PROMPT_SYSTEM,
             f"Sound instruction: {instruction or description or '(none — describe a sound that fits the lyrics)'}\n"
             f"Current lyrics, keep the structured prompt coherent with them:\n{state['lyrics']}",
+            required=("global_metadata", "vocal_details", "arrangement"),
         )
         state.update(global_meta=data["global_metadata"], vocals=data["vocal_details"], arrangement=data["arrangement"])
         return state, "Structured prompt written — tweak it, or press Generate."