Update start.sh

start.sh

@@ -5,6 +5,14 @@ echo "======================================================="
 echo " ADUC-SDR — Start (VINCIE/SeedVR, 8× L40S)"
 echo "======================================================="
 
+ls -la /app ls -R /app | head -n 2000
+ls -la /data ls -R /data | head -n 2000
+ls -la /app/.cache/huggingface -R /app/.cache/huggingface | head -n 2000
+ls -la /app/.cache -R /app/.cache | head -n 2000
+
+
+
+
 # ---------------------- Env base ----------------------
 export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"
 export TORCH_DTYPE="${TORCH_DTYPE:-bfloat16}"
@@ -16,61 +24,75 @@ export ENABLE_MATH_SDP="${ENABLE_MATH_SDP:-0}"
 export FLASH_ATTENTION_DISABLE="${FLASH_ATTENTION_DISABLE:-0}"
 export XFORMERS_FORCE_DISABLE="${XFORMERS_FORCE_DISABLE:-1}"
 
-# CUDA/NCCL/perf — single-node robust
-export CUDA_MODULE_LOADING="${CUDA_MODULE_LOADING:-LAZY}"
+# CUDA/NCCL/perf single-node robust
+export CUDA_MODULE_LOADING="LAZY"
 export CUDA_DEVICE_MAX_CONNECTIONS="${CUDA_DEVICE_MAX_CONNECTIONS:-32}"
-export CUDA_DEVICE_ORDER="${CUDA_DEVICE_ORDER:-PCI_BUS_ID}"
-export PYTORCH_CUDA_ALLOC_CONF="${PYTORCH_CUDA_ALLOC_CONF:-max_split_size_mb:512,garbage_collection_threshold:0.8}"
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+export PYTORCH_CUDA_ALLOC_CONF="max_split_size_mb:512,garbage_collection_threshold:0.8"
 export OMP_NUM_THREADS="${OMP_NUM_THREADS:-8}"
 export MKL_NUM_THREADS="${MKL_NUM_THREADS:-8}"
-export NCCL_DEBUG="${NCCL_DEBUG:-INFO}"
-export NCCL_ASYNC_ERROR_HANDLING="${NCCL_ASYNC_ERROR_HANDLING:-1}"
-export NCCL_P2P_DISABLE="${NCCL_P2P_DISABLE:-0}"
-export NCCL_IB_DISABLE="${NCCL_IB_DISABLE:-1}"
-export NCCL_SOCKET_IFNAME="${NCCL_SOCKET_IFNAME:-lo}"
+export NCCL_DEBUG="INFO"
+export NCCL_ASYNC_ERROR_HANDLING=1
+export NCCL_P2P_DISABLE=0
+export NCCL_IB_DISABLE=1
+export NCCL_SOCKET_IFNAME="lo"
 export NCCL_BLOCKING_WAIT=1
 export TORCH_NCCL_BLOCKING_WAIT=1
-export NCCL_TIMEOUT="${NCCL_TIMEOUT:-600}"
+export NCCL_TIMEOUT=600
 
 # HF caches
-export HF_HOME="${HF_HOME:-/app/.cache/huggingface}"
+export HF_HOME=/app/.cache/huggingface
 unset TRANSFORMERS_CACHE
 
+# ---------------------- Baixar o modelo antes de iniciar ----------------------
+echo "Baixando o modelo ByteDance-Seed/VINCIE-3B para cache persistente..."
+
+# Use Python para baixar o modelo via huggingface_hub
+python -c "
+import os
+from huggingface_hub import snapshot_download
+
+os.environ['HF_HOME'] = '/app/.cache/huggingface'
+repo_id = 'ByteDance-Seed/VINCIE-3B'
+snapshot_download(repo_id=repo_id, cache_dir=os.path.join(os.environ['HF_HOME'], 'hub'))
+" || { echo "Erro ao baixar o modelo. Continuando..."; }
+
+# Verificar se o cache foi criado
+ls -la $HF_HOME/hub || echo "Cache não encontrado após download."
+
 
-# ---------------------- Builder (Apex + Q8) ----------------------
-# Executa com GPU disponível; busca wheels no HF e compila se necessário (sem FlashAttention)
+
+
+
+# ---------------------- Builder Apex/Q8 ----------------------
 if nvidia-smi >/dev/null 2>&1; then
-  if [ "${DISABLE_BUILDER:-0}" = "0" ]; then
-    echo "🔧 Executando builder (Apex + Q8)..."
+  if [ "${DISABLE_BUILDER:-0}" -eq 0 ]; then
+    echo "Executando builder Apex/Q8..."
     chmod +x /app/builder.sh || true
-    # Configuráveis:
-    SELF_HF_REPO_ID="carlex3321/aduc-sdr"    # repo de wheels no HF
-    HF_UPLOAD_WHEELS=0         # publica wheels geradas
-    BUILDER_TIMEOUT_SEC=6000000      # tempo limite
-    #Q8_REPO / Q8_COMMIT          # pin do LTX Q8
-    ( timeout ${BUILDER_TIMEOUT_SEC:-60000} bash -lc "/app/builder.sh" ) || {
-      echo "⚠️ Builder excedeu tempo/retornou erro; prosseguindo com a aplicação."
-    }
+    # Executa com GPU disponível, busca wheels no HF e compila se necessário (sem FlashAttention)
+    SELF_HF_REPO_ID=carlex3321/aduc-sdr  # repo de wheels no HF
+    HF_UPLOAD_WHEELS=0  # publica wheels geradas
+    BUILDER_TIMEOUT_SEC=6000000  # tempo limite
+    Q8_REPO="" Q8_COMMIT=""  # pin do LTX Q8
+    timeout ${BUILDER_TIMEOUT_SEC:-60000} bash -lc /app/builder.sh || echo "Builder excedeu tempo/retornou erro, prosseguindo com a aplicação."
   else
-    echo "ℹ️ Builder desabilitado por DISABLE_BUILDER=1"
+    echo "Builder desabilitado por DISABLE_BUILDER=1"
   fi
 else
-  echo "⚠️ GPU não visível; pulando builder (Apex/Q8)."
+  echo "GPU não visível, pulando builder Apex/Q8."
 fi
 
-
-# ---------------------- Banner ----------------------
+# Configuráveis
 ./info.sh || true
+ls -la /app ls -R /app | head -n 2000
+ls -la /data ls -R /data | head -n 2000
+ls -la /app/.cache/huggingface -R /app/.cache/huggingface | head -n 2000
+ls -la /app/.cache -R /app/.cache | head -n 2000
 
 
-ls -la /app && ls -R /app | head -n 2000
-ls -la /data && ls -R /data | head -n 2000
-
 
-echo "🚀 Subindo serviços..."
-
-# Exemplo: subir UI mínima SD Img2Img (ajuste conforme seu app)
-# python app_animatediff_min.py
 
+echo "Subindo serviços..."
+# ---------------------- Banner ----------------------
+python /app/vince.py
 # Ou subir VINCIE UI se for o caso
-python app_vince.py