ZeroGPU attempt: boot-safe loader + @spaces.GPU realtime handler.

- peitho_model.py: load base + overlay German v6 with NO import-time CUDA
warmup (warmup crashes ZeroGPU on boot, which has no GPU until a
@spaces.GPU call). Import spaces first so CUDA placement is deferred.
- app.py: load via peitho_model instead of liquid_audio.demo.model.
- chat.py: import from peitho_model; wrap the FastRTC streaming handler in
@spaces.GPU so the whole generation runs in one GPU-scoped slice.
- requirements: add spaces.

Experimental: realtime FastRTC on ZeroGPU is unverified locally; validate
on the live build with hardware set to ZeroGPU.

Co-authored-by: Cursor <cursoragent@cursor.com>

app.py

@@ -1,45 +1,29 @@
-"""HF Spaces entry point: Peitho 1.5B v6 chat demo.
+"""HF Spaces (ZeroGPU) entry point: Peitho 1.5B v6 German voice demo.
 
 Background: the stock LFM2.5-Audio-1.5B base model is English-only for
 speech-to-speech (see https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B —
-"Supported languages: English"). Our v5/v6 fine-tunes are what teach it
-German audio output. A short A/B with the unmodified base model produced
-garbled German + robotic audio — confirming the fine-tune is essential.
+"Supported languages: English"). Our v6 fine-tune is what teaches it German
+audio output. A short A/B with the unmodified base model produced garbled
+German + robotic audio — confirming the fine-tune is essential.
 
-This file loads stock LFM2.5-Audio-1.5B via `liquid_audio.demo.model`, then
-overlays the v6 fine-tune weights from `jempf/peitho-1.5b-v6` on top.
+Model loading + the German v6 overlay live in `peitho_model.py`, which is
+written to be boot-safe on ZeroGPU (no CUDA warmup at import). This entry
+point just imports the model, then launches the chat demo.
 """
 from __future__ import annotations
 
 import sys
 from pathlib import Path
 
-from accelerate import load_checkpoint_in_model
-from huggingface_hub import snapshot_download
-
 sys.path.insert(0, str(Path(__file__).parent))
 
-PEITHO_REPO = "jempf/peitho-1.5b-v6"
-
 
 def main() -> None:
-    print(f"Downloading Peitho v6 weights from {PEITHO_REPO}...")
-    weights_dir = snapshot_download(
-        repo_id=PEITHO_REPO,
-        allow_patterns=["model.safetensors", "config.json"],
-    )
-    print(f"Weights cached at {weights_dir}")
-
-    print("Loading base LFM2.5-Audio-1.5B...")
-    import liquid_audio.demo.model as model_mod
-
-    print(f"Overlaying v6 weights onto base model...")
-    load_checkpoint_in_model(model_mod.lfm2_audio, weights_dir)
-    model_mod.lfm2_audio.eval()
-    print("v6 ready.")
+    import peitho_model  # noqa: F401  # triggers load + v6 overlay (boot-safe)
 
     print("Importing chat demo and launching...")
     import chat as chat_mod
+
     chat_mod.demo.launch()
 
 

chat.py

@@ -21,6 +21,7 @@ from queue import Queue
 from threading import Thread
 from typing import Any
 
+import spaces
 import gradio as gr
 import httpx
 import numpy as np
@@ -39,7 +40,9 @@ from briefing import (
     format_schedule_html,
 )
 from liquid_audio import ChatState, LFMModality
-from liquid_audio.demo.model import lfm2_audio, mimi, proc
+from peitho_model import lfm2_audio, mimi, proc
+
+GPU_DURATION_SECONDS = 120
 
 AUDIO_EOS_TOKEN = 2048
 TURN_TTL_SECONDS = 600
@@ -143,6 +146,7 @@ def chat_producer(
     q.put(None)
 
 
+@spaces.GPU(duration=GPU_DURATION_SECONDS)
 def chat_response(
     audio: tuple[int, np.ndarray],
     _id: str,

peitho_model.py

@@ -0,0 +1,37 @@
+"""Boot-safe Peitho model loader for ZeroGPU.
+
+Unlike `liquid_audio.demo.model`, this module does NOT run a CUDA warmup at
+import time. On ZeroGPU there is no GPU outside an `@spaces.GPU` call, so any
+eager CUDA op at import would crash the Space on boot. We import `spaces`
+first (which patches torch so `.to("cuda")` is deferred until a GPU is
+actually allocated), load the base model, overlay the German v6 fine-tune,
+and request the CUDA placement — the real transfer happens lazily inside the
+first GPU-scoped call.
+"""
+from __future__ import annotations
+
+import spaces  # noqa: F401  # must precede torch CUDA usage to enable ZeroGPU
+import torch
+from accelerate import load_checkpoint_in_model
+from huggingface_hub import snapshot_download
+from liquid_audio import LFM2AudioModel, LFM2AudioProcessor
+
+BASE_REPO: str = "LiquidAI/LFM2.5-Audio-1.5B"
+PEITHO_REPO: str = "jempf/peitho-1.5b-v6"
+
+print("Loading processor + base LFM2.5-Audio-1.5B (no CUDA warmup)...")
+proc = LFM2AudioProcessor.from_pretrained(BASE_REPO).eval()
+lfm2_audio = LFM2AudioModel.from_pretrained(BASE_REPO).eval()
+mimi = proc.mimi.eval()
+
+print(f"Overlaying German v6 weights from {PEITHO_REPO}...")
+_weights_dir = snapshot_download(
+    repo_id=PEITHO_REPO,
+    allow_patterns=["model.safetensors", "config.json"],
+)
+load_checkpoint_in_model(lfm2_audio, _weights_dir)
+lfm2_audio.eval()
+
+lfm2_audio.to("cuda")
+mimi.to("cuda")
+print("Peitho v6 ready (CUDA placement deferred until first GPU call).")

requirements.txt

@@ -3,3 +3,4 @@ fastrtc[vad]>=0.0.30
 accelerate>=1.0.0
 huggingface_hub>=0.27.0
 httpx>=0.27.0
+spaces>=0.30.0