Fix: use huggingface_hub to download models (xet compatible)

start.sh

@@ -26,6 +26,7 @@ download_if_missing() {
 
 HF_BASE="https://huggingface.co/lj1995/GPT-SoVITS/resolve/main"
 FASTTEXT_LID_URL="https://dl.fbaipublicfiles.com/fasttext/supervised-models/lid.176.bin"
+
 download_if_missing "$HF_BASE/chinese-roberta-wwm-ext-large/config.json" "$PM_DIR/chinese-roberta-wwm-ext-large/config.json"
 download_if_missing "$HF_BASE/chinese-roberta-wwm-ext-large/pytorch_model.bin" "$PM_DIR/chinese-roberta-wwm-ext-large/pytorch_model.bin"
 download_if_missing "$HF_BASE/chinese-roberta-wwm-ext-large/tokenizer.json" "$PM_DIR/chinese-roberta-wwm-ext-large/tokenizer.json"
@@ -37,16 +38,27 @@ download_if_missing "$HF_BASE/v2Pro/s2Gv2ProPlus.pth" "$PM_DIR/v2Pro/s2Gv2ProPlu
 download_if_missing "$HF_BASE/sv/pretrained_eres2netv2w24s4ep4.ckpt" "$PM_DIR/sv/pretrained_eres2netv2w24s4ep4.ckpt"
 download_if_missing "$FASTTEXT_LID_URL" "$PM_DIR/fast_langdetect/lid.176.bin"
 
-# Download Aemeath (爱弥斯) custom character models
+# Download Aemeath custom character models using huggingface_hub (handles xet storage)
 CHAR_DIR="/data/models"
 mkdir -p "$CHAR_DIR"
 
-CHAR_BASE="https://huggingface.co/spaces/huanx/gpt-sovits/resolve/main"
-download_if_missing "$CHAR_BASE/aimisi-e16.ckpt" "$CHAR_DIR/aimisi-e16.ckpt"
-download_if_missing "$CHAR_BASE/aimisi_e10_s180.pth" "$CHAR_DIR/aimisi_e10_s180.pth"
+echo "Downloading character models via huggingface_hub..."
+python3 -c "
+from huggingface_hub import hf_hub_download
+import os
+
+char_dir = '$CHAR_DIR'
+repo = 'huanx/gpt-sovits'
 
-REF_BASE="https://huggingface.co/spaces/huanx/gpt-sovits/resolve/main"
-download_if_missing "$REF_BASE/zh_vo_MAIN_YHX_2_12.wav" "$CHAR_DIR/zh_vo_MAIN_YHX_2_12.wav"
+for fname in ['aimisi-e16.ckpt', 'aimisi_e10_s180.pth', 'zh_vo_MAIN_YHX_2_12.wav']:
+    dest = os.path.join(char_dir, fname)
+    if not os.path.exists(dest) or os.path.getsize(dest) < 1000:
+        print(f'Downloading {fname}...')
+        hf_hub_download(repo_id=repo, filename=fname, local_dir=char_dir, repo_type='space')
+        print(f'  -> {os.path.getsize(dest)} bytes')
+    else:
+        print(f'{fname} already exists ({os.path.getsize(dest)} bytes)')
+"
 
 echo "Pretrained: $(find "$PM_DIR" -type f | wc -l) files"
 echo "Character: $(find "$CHAR_DIR" -type f | wc -l) files"
@@ -77,47 +89,25 @@ VITS_WEIGHTS="$PM_DIR/v2Pro/s2Gv2ProPlus.pth"
 if [ -s "$CHAR_DIR/aimisi-e16.ckpt" ] && [ -s "$CHAR_DIR/aimisi_e10_s180.pth" ]; then
     T2S_WEIGHTS="$CHAR_DIR/aimisi-e16.ckpt"
     VITS_WEIGHTS="$CHAR_DIR/aimisi_e10_s180.pth"
+    echo "Using Aemeath custom models"
+else
+    echo "WARN: Aemeath models not found, using default pretrained"
 fi
 
-cat > "$CFG_PATH" <<EOF
+cat > "$CFG_PATH" << EOFCFG
 custom:
-  bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
-  cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
+  bert_base_path: $PM_DIR/chinese-roberta-wwm-ext-large
+  cnhuhbert_base_path: $PM_DIR/chinese-hubert-base
   device: cpu
   is_half: false
   t2s_weights_path: $T2S_WEIGHTS
   version: v2ProPlus
   vits_weights_path: $VITS_WEIGHTS
-EOF
-
-echo "Using GPT weights: $T2S_WEIGHTS"
-echo "Using SoVITS weights: $VITS_WEIGHTS"
-
-# Force CPU
-export CUDA_VISIBLE_DEVICES=""
-export DEVICE="cpu"
-export FTLANG_CACHE="$PM_DIR/fast_langdetect"
-export SPACE_TTS_CONFIG="GPT_SoVITS/configs/tts_infer_cpu.yaml"
-export NLTK_DATA="/usr/local/nltk_data"
-
-mkdir -p "$NLTK_DATA"
-python - <<'PY'
-import nltk
-packages = {
-    "averaged_perceptron_tagger_eng": "taggers/averaged_perceptron_tagger_eng",
-    "averaged_perceptron_tagger": "taggers/averaged_perceptron_tagger",
-    "cmudict": "corpora/cmudict",
-}
-for pkg, resource_path in packages.items():
-    try:
-        nltk.data.find(resource_path)
-    except LookupError:
-        try:
-            nltk.download(pkg, download_dir="/usr/local/nltk_data", quiet=True)
-        except Exception as exc:
-            print(f"WARN: failed to download NLTK package {pkg}: {exc}")
-PY
-
-# Start Space UI + API
-echo "Starting Space UI on 0.0.0.0:7860..."
-exec python -m uvicorn ui:app --host 0.0.0.0 --port 7860
+EOFCFG
+
+echo "Config written to $CFG_PATH"
+cat "$CFG_PATH"
+
+# Launch UI
+echo "Starting GPT-SoVITS UI on port 7860..."
+exec python3 ui.py