feat: modal integration

modal_wrapper.py

@@ -0,0 +1,41 @@
+from audiocraft.models.musicgen import MusicGen
+import modal
+
+
+MODEL_DIR = "/root/model/model_input"
+MODEL_ID = "facebook/musicgen-large"
+N_GPUS = 1
+GPU_CONFIG = modal.gpu.A100(count=N_GPUS)
+
+
+def download_model():
+    import torchaudio
+    from audiocraft.models.musicgen import MusicGen
+
+    MusicGen.get_pretrained(MODEL_ID)
+
+
+image = modal.Image.from_registry("python:3.9.19-slim-bookworm")
+
+image = (
+    image.apt_install("ffmpeg")
+    .env({"AUDIOCRAFT_CACHE_DIR": MODEL_DIR})
+    .pip_install("audiocraft==1.3.0", "torchaudio==2.1.0")
+    .run_function(download_model, timeout=20 * 60)
+)
+
+
+app = modal.App("infinifi", image=image)
+
+
+@app.cls(gpu=GPU_CONFIG, container_idle_timeout=15 * 60)
+class Model:
+    @modal.enter()
+    def load(self):
+        self.model = MusicGen.get_pretrained(MODEL_ID)
+        self.model.set_generation_params(duration=60)
+
+    @modal.method()
+    def generate(self, prompts):
+        wav = self.model.generate(prompts)
+        return [one_wav for i, one_wav in enumerate(wav)]

server.py

@@ -1,16 +1,28 @@
 import threading
 
-# from generate import generate
+import modal
 from contextlib import asynccontextmanager
 from fastapi import FastAPI
 from fastapi.responses import FileResponse
 from fastapi.staticfiles import StaticFiles
+from audiocraft.data.audio import audio_write
 
 # the index of the current audio track from 0 to 9
 current_index = -1
 # the timer that periodically advances the current audio track
 t = None
 
+prompts = [
+    "gentle, calming lo-fi beats that helps with studying and focusing",
+    "calm, piano lo-fi beats to help with studying and focusing",
+    "gentle lo-fi hip-hop to relax to",
+    "gentle, quiet synthwave lo-fi beats",
+    "morning lo-fi beats",
+]
+
+
+model = modal.Cls.lookup("infinifi", "Model")
+
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
@@ -20,20 +32,45 @@ async def lifespan(app: FastAPI):
         t.cancel()
 
 
+def generate_new_audio():
+    global current_index
+
+    offset = 0
+    wav = []
+    if current_index == 0:
+        offset = 5
+        wav = model.generate.remote(prompts)
+    elif current_index == 5:
+        offset = 0
+        wav = model.generate.remote(prompts)
+    else:
+        return
+
+    print("generating new audio...")
+
+    for idx, one_wav in enumerate(wav):
+        # Will save under {idx}.wav, with loudness normalization at -14 db LUFS.
+        audio_write(
+            f"{idx + offset}",
+            one_wav.cpu(),
+            model.sample_rate,
+            format="mp3",
+            strategy="loudness",
+            loudness_compressor=True,
+        )
+
+    print("audio generated.")
+
+
 def advance():
     global current_index, t
 
-    # if current_index == 0:
-    #   generate(offset=5)
-    # elif current_index == 5:
-    #   generate(offset=0)
-
     if current_index == 9:
         current_index = 0
     else:
         current_index = current_index + 1
 
-    print(f"advancing, current index {current_index}")
+    threading.Thread(target=generate_new_audio).start()
 
     t = threading.Timer(60, advance)
     t.start()