app.py

#!/usr/bin/env python3
"""
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
"""

# ── ZERO-GPU decorator must come BEFORE torch/CUDA ────────────────
import spaces                  # ① import first

# ── Standard libs ─────────────────────────────────────────────────
import os, sys, gc, time, warnings, tempfile, subprocess, random
import numpy as np, psutil

# ── Torch (CPU wheels already in container; tensors move to GPU) ──
import torch, torchaudio

# ── Other deps ────────────────────────────────────────────────────
import gradio as gr
from pydub import AudioSegment
from huggingface_hub import login
from torch.cuda.amp import autocast

# ── 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"
    )

warnings.filterwarnings("ignore")
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"

# ── 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: Please add HF_TOKEN secret in your Space.")
login(HF_TOKEN)

# ── 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())
musicgen.set_generation_params(duration=10, two_step_cfg=False)
SR = musicgen.sample_rate

# ── 4. prompt helpers (two shown; add the rest as before) ─────────
def _p(base,bpm,dr,syn,st,bass,gtr,db,dg,flow):
    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(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 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 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, "⚠️ 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)

    for i in range(n_chunks):
        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)
            parts.append(AudioSegment.from_wav(tmp.name))
        os.unlink(tmp.name)
        torch.cuda.empty_cache(); gc.collect(); log(f"chunk {i+1} done")

    track = parts[0]
    for seg in parts[1:]:
        track = track.append(seg, crossfade=xfade)
    track = track[:total_len*1000]
    track = apply_fade(apply_eq(track).normalize(headroom=-9.0))

    out_f = "output_cleaned.mp3"
    track.export(out_f, format="mp3", bitrate="128k",
                 tags={"title":"GhostAI Track","artist":"GhostAI"})
    return out_f, "✅ Done!"

def clear_inputs():
    return ("",3.0,250,0.9,1.0,30,10,1000,
            120,"none","none","none","none","none")

# ── 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,
            inputs=[gr.State(120),"none","none","none","none","none"],
            outputs=prompt)
        gr.Button("Nirvana 🎸").click(set_nirvana_grunge,
            inputs=[gr.State(120),"none","none","none","none","none"],
            outputs=prompt)

    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")
    steps = gr.Dropdown(["none","syncopated steps","steady steps","complex steps"],"none","Steps")
    bass  = gr.Dropdown(["none","slap bass","deep bass","melodic bass"],"none","Bass")
    gtr   = gr.Dropdown(["none","distorted","clean","jangle"],"none","Guitar")

    gen = gr.Button("Generate 🚀")
    clr = gr.Button("Clear 🧹")
    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)