Multi-user concurrency: compose outside the GPU queue, parallel assists/examples/videos

Simple mode ran the composer LLM call inside the concurrency_id="gpu" (limit 1) event, so after clicking Generate the LLM request did not even START until every queued song from other users finished — the delay users are reporting. Also, Gradio's default concurrency limit is 1 per function, so one user's assist/example/video blocked everyone else's.

- Simple flow split: simple_compose (limit 8, starts immediately) -> .success -> studio_generate on the GPU queue. Status shows 'waiting for a GPU slot' between the two.
- compose_assist: limit 8. render_video: limit 4. load_example (cached files): unlimited.
- player.stop cancels the compose step too.

API note: the /simple_generate endpoint is replaced by /simple_compose + the chained sing step.

app.py

@@ -1533,19 +1533,21 @@ def compose_assist(raw_state, duration):
     return state, "Lyrics & structured prompt ready — review and tweak them, then press Generate."
 
 
-def simple_generate(state, duration, seed, randomize_seed, headroom, steps, guidance):
-    # One event so the output components engage (spinner) from the first click, through composing and singing.
+def simple_compose(state, duration):
+    # Composing is a plain LLM API call, so it runs OUTSIDE the GPU queue: it starts the moment the user
+    # clicks instead of waiting behind every queued song. The singing step chains after it on the GPU queue.
     state = _normalize_state(state)
     description = _composed_description(state)
     title = state["title"] or state["description"]
-    yield gr.skip(), "writing lyrics & structured prompt...", gr.skip(), gr.skip(), gr.skip(), title
+    yield gr.skip(), "writing lyrics & structured prompt...", gr.skip(), title
     lyr, gm, vd, arr = compose_song(description, duration)
     state.update(lyrics=lyr, global_meta=gm, vocals=vd, arrangement=arr)
-    yield {"cmd": "status", "text": "Lyrics ready — acquiring ZeroGPU..."}, "lyrics ready — acquiring ZeroGPU...", gr.skip(), gr.skip(), state, gr.skip()
-    for audio, status, wav, used_seed in generate(
-        description, lyr, gm, vd, arr, duration, seed, randomize_seed, headroom, steps, guidance
-    ):
-        yield audio, status, wav, used_seed, gr.skip(), gr.skip()
+    yield (
+        {"cmd": "status", "text": "Lyrics ready — waiting for a GPU slot..."},
+        "lyrics ready — waiting for a GPU slot...",
+        state,
+        gr.skip(),
+    )
 
 
 def studio_generate(state, duration, seed, randomize_seed, headroom, steps, guidance):
@@ -1618,12 +1620,17 @@ with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
 
     _knobs = [duration, seed, randomize_seed, headroom, steps, guidance]
 
-    ev_simple = composer.submit(lambda: ({"cmd": "reset", "loader": True, "status": "Writing lyrics & structured prompt with MiniMax-M3...", "nonce": random.random()}, None), None, [player, video_out]).then(
-        simple_generate, [composer] + _knobs,
-        [player, stats, wav_state, seed, composer, video_title],
+    ev_compose = composer.submit(lambda: ({"cmd": "reset", "loader": True, "status": "Writing lyrics & structured prompt with MiniMax-M3...", "nonce": random.random()}, None), None, [player, video_out]).then(
+        simple_compose, [composer, duration],
+        [player, stats, composer, video_title],
+        concurrency_limit=8, show_progress="minimal",
+    )
+    ev_simple = ev_compose.success(
+        studio_generate, [composer] + _knobs,
+        [player, stats, wav_state, seed, video_title],
         concurrency_limit=1, concurrency_id="gpu", show_progress="minimal",
     )
-    ev_simple.then(render_video, [wav_state, video_title], video_out).then(
+    ev_simple.then(render_video, [wav_state, video_title], video_out, concurrency_limit=4).then(
         lambda s: s.replace(" — rendering share video...", " — share video ready."), stats, stats
     )
 
@@ -1632,18 +1639,18 @@ with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
         [player, stats, wav_state, seed, video_title],
         concurrency_limit=1, concurrency_id="gpu", show_progress="minimal",
     )
-    ev_studio.then(render_video, [wav_state, video_title], video_out).then(
+    ev_studio.then(render_video, [wav_state, video_title], video_out, concurrency_limit=4).then(
         lambda s: s.replace(" — rendering share video...", " — share video ready."), stats, stats
     )
 
     player.stop(
         lambda: "stopped — kept the part that was already streamed.", None, stats,
-        cancels=[ev_simple, ev_studio], show_progress="hidden",
+        cancels=[ev_compose, ev_simple, ev_studio], show_progress="hidden",
     )
 
-    composer.edit(compose_assist, [composer, duration], [composer, stats], show_progress="minimal")
+    composer.edit(compose_assist, [composer, duration], [composer, stats], concurrency_limit=8, show_progress="minimal")
 
-    composer.apply(load_example, [composer], [composer, player, stats, video_out, wav_state, video_title], show_progress="minimal")
+    composer.apply(load_example, [composer], [composer, player, stats, video_out, wav_state, video_title], concurrency_limit=None, show_progress="minimal")
 
 if __name__ == "__main__":
     demo.launch()