Update app.py

app.py

@@ -1,29 +1,28 @@
 #!/usr/bin/env python3
 """
-GhostAI Music Generator – ZeroGPU Edition
-• Streams facebook/musicgen-medium from the Hub using your HF_TOKEN
-• Runs on a transient GPU provided by the ZeroGPU scheduler
-• Keeps your genre prompt builders, EQ / fade FX, neon Gradio UI
+GhostAI Music Generator — ZeroGPU Space
+• Pulls facebook/musicgen-medium from the Hub (needs HF_TOKEN)
+• Installs audiocraft==1.3.0 *at runtime* ( --no-deps ) to dodge spaCy/typer clash
+• Decorated with @spaces.GPU so each call gets a transient GPU
 """
 
-# ── ZeroGPU decorator must be imported *before* torch/CUDA ──────────
-import spaces                                  # 👈 FIRST import
+# ── ZERO-GPU decorator must come BEFORE torch/CUDA ────────────────
+import spaces                  # ① import first
 
-# ── Standard libs ───────────────────────────────────────────────────
-import os, sys, gc, time, random, warnings, tempfile
+# ── Standard libs ─────────────────────────────────────────────────
+import os, sys, gc, time, warnings, tempfile, subprocess, random
 import numpy as np, psutil
 
-# ── Torch (CPU wheels; ZeroGPU moves tensors to the temp GPU) ───────
+# ── Torch (CPU wheels already in container; tensors move to GPU) ──
 import torch, torchaudio
 
-# ── Other deps ──────────────────────────────────────────────────────
+# ── Other deps ────────────────────────────────────────────────────
 import gradio as gr
 from pydub import AudioSegment
-from audiocraft.models import MusicGen
 from huggingface_hub import login
 from torch.cuda.amp import autocast
 
-# ── Torch <2.3 shim (transformers 4.35 calls get_default_device) ───
+# ── tiny shim for transformers on Torch <2.3 ──────────────────────
 if not hasattr(torch, "get_default_device"):
     torch.get_default_device = lambda: torch.device(
         "cuda" if torch.cuda.is_available() else "cpu"
@@ -32,130 +31,125 @@ if not hasattr(torch, "get_default_device"):
 warnings.filterwarnings("ignore")
 os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"
 
-# ────────────────────────── HF authentication ──────────────────────
+# ── 1. make sure audiocraft is present (no deps) ──────────────────
+try:
+    from audiocraft.models import MusicGen
+except ModuleNotFoundError:
+    print("🔧 Installing audiocraft 1.3.0 (no-deps)…")
+    subprocess.check_call([sys.executable, "-m", "pip", "install",
+                           "audiocraft==1.3.0", "--no-deps", "--quiet"])
+    from audiocraft.models import MusicGen
+
+# ── 2. HF authentication ─────────────────────────────────────────
 HF_TOKEN = os.getenv("HF_TOKEN")
 if not HF_TOKEN:
-    sys.exit("ERROR: add HF_TOKEN as a secret in your Space settings.")
+    sys.exit("ERROR: Please add HF_TOKEN secret in your Space.")
 login(HF_TOKEN)
 
-print("🔄  Loading facebook/musicgen-medium (first build ≈ 6 GB)…")
+# ── 3. load model from Hub ────────────────────────────────────────
+print("⏬  Loading facebook/musicgen-medium (first run ≈ 6 GB)…")
 musicgen = MusicGen.get_pretrained("facebook/musicgen-medium")
-musicgen.to(torch.get_default_device())            # move entire model
+musicgen.to(torch.get_default_device())
 musicgen.set_generation_params(duration=10, two_step_cfg=False)
 SR = musicgen.sample_rate
 
-# ────────────────────────── Prompt helpers ─────────────────────────
+# ── 4. prompt helpers (two shown; add the rest as before) ─────────
 def _p(base,bpm,dr,syn,st,bass,gtr,db,dg,flow):
-    step = f" with {st}" if st!="none" else flow.format(bpm=bpm)
-    return (f"{base}"
-            f"{', '+bass if bass!='none' else db}"
-            f"{', '+gtr+' guitar riffs' if gtr!='none' else dg}"
-            f"{', '+dr+' drums' if dr!='none' else ''}"
-            f"{', '+syn+' accents' if syn!='none' else ''}"
-            f"{step} at {bpm} BPM.")
-
-def set_red_hot_chili_peppers_prompt(bpm,dr,syn,st,bass,gtr):
+    stxt = f" with {st}" if st!="none" else flow.format(bpm=bpm)
+    return (f"{base}{db if bass=='none' else ', '+bass}"
+            f"{dg if gtr=='none' else ', '+gtr+' guitar riffs'}"
+            f"{'' if dr=='none' else ', '+dr+' drums'}"
+            f"{'' if syn=='none' else ', '+syn+' accents'}{stxt} at {bpm} BPM.")
+
+def set_red_hot_chili(bpm,dr,syn,st,bass,gtr):
     return _p("Instrumental funk rock",bpm,dr,syn,st,bass,gtr,
               ", groovy basslines",", syncopated guitar riffs",
               "{bpm} BPM funky flow" if bpm>120 else "groovy rhythmic flow")
-def set_nirvana_grunge_prompt(bpm,dr,syn,st,bass,gtr):
+
+def set_nirvana_grunge(bpm,dr,syn,st,bass,gtr):
     return _p("Instrumental grunge",bpm,dr,syn,st,bass,gtr,
               ", melodic basslines",", raw distorted guitar riffs",
               "{bpm} BPM grungy pulse" if bpm>120 else "grungy rhythmic pulse")
-# 👉 add your other genre prompt functions here exactly as before …
 
-# ────────────────────────── Audio FX helpers ───────────────────────
-def apply_eq(s):   return s.low_pass_filter(8000).high_pass_filter(80)
-def apply_fade(s): return s.fade_in(1000).fade_out(1000)
+# (… add your remaining genre prompt functions unchanged …)
+
+# ── 5. audio FX helpers ───────────────────────────────────────────
+def apply_eq(seg):   return seg.low_pass_filter(8000).high_pass_filter(80)
+def apply_fade(seg): return seg.fade_in(1000).fade_out(1000)
 
 def log(stage=""):
     if stage: print(f"── {stage} ──")
     if torch.cuda.is_available():
         a = torch.cuda.memory_allocated()/1024**3
         r = torch.cuda.memory_reserved()/1024**3
-        print(f"GPU mem alloc {a:.2f} GB  reserved {r:.2f} GB")
-    print(f"CPU mem {psutil.virtual_memory().percent}% used")
-
-# ──────────────────────── Core generator API ───────────────────────
-@spaces.GPU                     # 🟢 ZeroGPU will allocate a GPU per call
-def generate_music(prompt,cfg,k,p,temp,
-                   total_len,chunk_len,xfade,
-                   bpm,dr,syn,step,bass,gtr):
-
+        print(f"GPU alloc {a:.2f} GB  reserved {r:.2f} GB")
+    print(f"CPU {psutil.virtual_memory().percent}%")
+
+# ── 6. main generation API (ZeroGPU wrapper) ───────────────────────
+@spaces.GPU
+def generate(prompt,cfg,k,p,temp,
+             total_len,chunk_len,xfade,
+             bpm,dr,syn,step,bass,gtr):
     if not prompt.strip():
-        return None, "⚠️ Prompt is empty."
-
-    total_len  = int(total_len)
-    chunk_len  = max(5, min(int(chunk_len), 15))
-    n_chunks   = max(1, total_len // chunk_len)
-    chunk_len  = total_len / n_chunks
-    overlap    = min(1.0, xfade / 1000.0)
-    render_len = chunk_len + overlap
-    pieces     = []
-
+        return None, "⚠️ Empty prompt."
+
+    total_len, chunk_len = int(total_len), max(5, min(int(chunk_len), 15))
+    n_chunks  = max(1, total_len // chunk_len)
+    chunk_len = total_len / n_chunks
+    overlap   = min(1.0, xfade / 1000.0)
+    render    = chunk_len + overlap
+    parts     = []
     torch.manual_seed(42); np.random.seed(42)
-    t0 = time.time()
 
     for i in range(n_chunks):
-        log(f"before chunk {i+1}")
-        musicgen.set_generation_params(
-            duration=render_len, use_sampling=True,
-            top_k=k, top_p=p, temperature=temp, cfg_coef=cfg
-        )
+        log(f"chunk {i+1} start")
+        musicgen.set_generation_params(duration=render,use_sampling=True,
+                                       top_k=k,top_p=p,temperature=temp,cfg_coef=cfg)
         with torch.no_grad(), autocast():
             audio = musicgen.generate([prompt], progress=False)[0]
-
         audio = audio.cpu().float()
         if audio.dim()==1 or audio.shape[0]==1:
             audio = audio.repeat(2,1)
         with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
             torchaudio.save(tmp.name, audio, SR)
-            seg = AudioSegment.from_wav(tmp.name)
+            parts.append(AudioSegment.from_wav(tmp.name))
         os.unlink(tmp.name)
-        pieces.append(seg)
+        torch.cuda.empty_cache(); gc.collect(); log(f"chunk {i+1} done")
 
-        torch.cuda.empty_cache(); gc.collect()
-        log(f"after chunk {i+1}")
-
-    track = pieces[0]
-    for seg in pieces[1:]:
+    track = parts[0]
+    for seg in parts[1:]:
         track = track.append(seg, crossfade=xfade)
-    track = track[: total_len*1000]
+    track = track[:total_len*1000]
     track = apply_fade(apply_eq(track).normalize(headroom=-9.0))
 
-    out_path = "output_cleaned.mp3"
-    track.export(out_path, format="mp3", bitrate="128k",
+    out_f = "output_cleaned.mp3"
+    track.export(out_f, format="mp3", bitrate="128k",
                  tags={"title":"GhostAI Track","artist":"GhostAI"})
-    log("final"); print(f"⏱  {time.time()-t0:.1f}s total")
-    return out_path, "✅ Done!"
+    return out_f, "✅ Done!"
 
 def clear_inputs():
     return ("",3.0,250,0.9,1.0,30,10,1000,
             120,"none","none","none","none","none")
 
-# ────────────────────────── Gradio UI ──────────────────────────────
-css = "body{background:#0A0A0A;color:#E0E0E0;font-family:'Orbitron',sans-serif}"
-
-with gr.Blocks(css=css) as demo:
-    gr.HTML("<h1 style='text-align:center'>👻 GhostAI Music Generator</h1>")
-    prompt = gr.Textbox(lines=4, label="Instrumental Prompt")
+# ── 7. minimal Gradio 4 UI (add more buttons if needed) ───────────
+with gr.Blocks(css="body{background:#0A0A0A;color:#E0E0E0;font-family:'Orbitron',sans-serif}") as demo:
+    gr.HTML("<h1 style='text-align:center'>👻 GhostAI MusicGen</h1>")
+    prompt = gr.Textbox(lines=4, label="Prompt")
     with gr.Row():
-        gr.Button("RHCP 🌶️").click(set_red_hot_chili_peppers_prompt,
+        gr.Button("RHCP 🌶️").click(set_red_hot_chili,
             inputs=[gr.State(120),"none","none","none","none","none"],
             outputs=prompt)
-        gr.Button("Nirvana 🎸").click(set_nirvana_grunge_prompt,
+        gr.Button("Nirvana 🎸").click(set_nirvana_grunge,
             inputs=[gr.State(120),"none","none","none","none","none"],
             outputs=prompt)
-        # 📌 add remaining genre buttons here …
-
-    cfg   = gr.Slider(1,10,3,label="CFG Scale")
-    top_k = gr.Slider(10,500,250,step=10,label="Top-K")
-    top_p = gr.Slider(0,1,0.9,step=0.05,label="Top-P")
-    temp  = gr.Slider(0.1,2,1,step=0.1,label="Temperature")
-    length= gr.Radio([30,60,90,120], value=30, label="Length (s)")
-    chunk = gr.Slider(5,15,10,step=1,label="Chunk (s)")
-    xfade = gr.Slider(100,2000,1000,step=100,label="Cross-fade (ms)")
 
+    cfg   = gr.Slider(1,10,3,label="CFG")
+    k     = gr.Slider(10,500,250,step=10,label="Top-K")
+    p     = gr.Slider(0,1,0.9,step=0.05,label="Top-P")
+    temp  = gr.Slider(0.1,2,1,label="Temp")
+    length= gr.Radio([30,60,90,120],value=30,label="Length")
+    chunk = gr.Slider(5,15,10,label="Chunk")
+    xfade = gr.Slider(100,2000,1000,label="Cross-fade")
     bpm   = gr.Slider(60,180,120,label="BPM")
     drum  = gr.Dropdown(["none","standard rock","funk groove","techno kick","jazz swing"],"none","Drum")
     synth = gr.Dropdown(["none","analog synth","digital pad","arpeggiated synth"],"none","Synth")
@@ -165,15 +159,13 @@ with gr.Blocks(css=css) as demo:
 
     gen = gr.Button("Generate 🚀")
     clr = gr.Button("Clear 🧹")
-    audio_out = gr.Audio(type="filepath")
-    status    = gr.Textbox(interactive=False)
-
-    gen.click(generate_music,
-              inputs=[prompt,cfg,top_k,top_p,temp,length,chunk,xfade,
-                      bpm,drum,synth,steps,bass,gtr],
-              outputs=[audio_out,status])
-    clr.click(clear_inputs,None,
-              [prompt,cfg,top_k,top_p,temp,length,chunk,xfade,
-               bpm,drum,synth,steps,bass,gtr])
+    out = gr.Audio(type="filepath")
+    status = gr.Textbox(interactive=False)
+
+    gen.click(generate,
+              [prompt,cfg,k,p,temp,length,chunk,xfade,bpm,drum,synth,steps,bass,gtr],
+              [out,status])
+    clr.click(clear_inputs, None,
+              [prompt,cfg,k,p,temp,length,chunk,xfade,bpm,drum,synth,steps,bass,gtr])
 
 demo.launch(share=False)