Update start.sh

start.sh

@@ -9,14 +9,19 @@ echo "======================================================="
 export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"
 export TORCH_DTYPE="${TORCH_DTYPE:-bfloat16}"
 
-# SDPA/FA toggles (FA pode ficar desativado; manter SDPA/mem-eff)
+# SDPA/FA toggles
 export ENABLE_FLASH_SDP="${ENABLE_FLASH_SDP:-1}"
 export ENABLE_MEMORY_EFFICIENT_SDP="${ENABLE_MEMORY_EFFICIENT_SDP:-1}"
 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
+# Triton.ops legacy off / BNB off
+export BNB_DISABLE=1
+export BITSANDBYTES_DISABLE=1
+export LOAD_IN_8BIT=0
+
+# CUDA/NCCL/perf — single-node robust
 export CUDA_MODULE_LOADING="${CUDA_MODULE_LOADING:-LAZY}"
 export CUDA_DEVICE_MAX_CONNECTIONS="${CUDA_DEVICE_MAX_CONNECTIONS:-32}"
 export CUDA_DEVICE_ORDER="${CUDA_DEVICE_ORDER:-PCI_BUS_ID}"
@@ -27,91 +32,22 @@ export NCCL_DEBUG="${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="${NCCL_IB_DISABLE:-1}"
-export NCCL_MIN_NCHANNELS="${NCCL_MIN_NCHANNELS:-8}"
-export NCCL_NTHREADS="${NCCL_NTHREADS:-256}"
+export NCCL_SOCKET_IFNAME="${NCCL_SOCKET_IFNAME:-lo}"
+export NCCL_BLOCKING_WAIT=1
+export TORCH_NCCL_BLOCKING_WAIT=1
+export NCCL_TIMEOUT="${NCCL_TIMEOUT:-600}"
 
-# Caches HF
-export APP_HOME="${APP_HOME:-/app}"
+# HF caches
 export HF_HOME="${HF_HOME:-/app/.cache/huggingface}"
-export HF_DATASETS_CACHE="${HF_DATASETS_CACHE:-/app/.cache/datasets}"
-export TRANSFORMERS_CACHE="${TRANSFORMERS_CACHE:-/app/.cache/transformers}"
-export DIFFUSERS_CACHE="${DIFFUSERS_CACHE:-/app/.cache/diffusers}"
-export TORCH_HOME="${TORCH_HOME:-/app/.cache/torch}"
-export HF_HUB_ENABLE_HF_TRANSFER="${HF_HUB_ENABLE_HF_TRANSFER:-1}"
-export TOKENIZERS_PARALLELISM="${TOKENIZERS_PARALLELISM:-false}"
-export MODELS_DIR="${MODELS_DIR:-/app/models}"
-
-mkdir -p /app/.cache/torch /app/.cache/huggingface /app/.cache/datasets /app/.cache/transformers /app/.cache/diffusers || true
-mkdir -p /home/user/.cache/models || true
-[ -L /app/models ] || ln -sf /home/user/.cache/models /app/models || true
-
-# ---------------------- Diagnóstico ----------------------
-nvidia-smi || true
-python - <<'PY' || true
-import torch, os
-print("Torch:", torch.__version__)
-print("CUDA available:", torch.cuda.is_available())
-print("Num GPUs:", torch.cuda.device_count())
-print("MemEff SDP:", os.getenv("ENABLE_MEMORY_EFFICIENT_SDP"))
-print("Math SDP:", os.getenv("ENABLE_MATH_SDP"))
-print("FA disable:", os.getenv("FLASH_ATTENTION_DISABLE"))
-print("Dtype:", os.getenv("TORCH_DTYPE"))
-PY
-
-# ---------------------- Builder (Apex + Q8) ----------------------
-# Executa com GPU disponível; busca wheels no HF e compila se necessário (sem FlashAttention)
-if nvidia-smi >/dev/null 2>&1; then
-  if [ "${DISABLE_BUILDER:-0}" = "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=1           # publica wheels geradas
-    BUILDER_TIMEOUT_SEC=6000000      # tempo limite
-    #Q8_REPO / Q8_COMMIT          # pin do LTX Q8
-    ( timeout ${BUILDER_TIMEOUT_SEC:-60000} bash -lc "/app/builder.sh" ) || {
-      echo "⚠️ Builder excedeu tempo/retornou erro; prosseguindo com a aplicação."
-    }
-  else
-    echo "ℹ️ Builder desabilitado por DISABLE_BUILDER=1"
-  fi
-else
-  echo "⚠️ GPU não visível; pulando builder (Apex/Q8)."
-fi
-
-# ---------------------- Verificação dos modelos ----------------------
-echo "🔍 Verificando a integridade dos modelos..."
-#[ -x "./setup/vince.sh" ] && ./setup/vince.sh || bash ./setup/vince.sh || true
-#[ -x "./setup/seedvr.sh" ] && ./setup/seedvr.sh || bash ./setup/seedvr.sh || true
-#[ -x "./setup/ltx.sh" ] && ./setup/ltx.sh || bash ./setup/ltx.sh || true
-echo "✅ Modelos verificados."
-
-# ---------------------- Preferências Diffusers ----------------------
-export DIFFUSERS_ATTENTION_IMPLEMENTATION="${DIFFUSERS_ATTENTION_IMPLEMENTATION:-sdpa}"   # flash opcional
-export DIFFUSERS_USE_TORCH_SDPA="${DIFFUSERS_USE_TORCH_SDPA:-1}"
-export DIFFUSERS_ENABLE_CPU_OFFLOAD="${DIFFUSERS_ENABLE_CPU_OFFLOAD:-0}"
-export DIFFUSERS_ENABLE_TORCH_COMPILE="${DIFFUSERS_ENABLE_TORCH_COMPILE:-0}"
-export DIFFUSERS_SDP_KERNEL_FALLBACK="${DIFFUSERS_SDP_KERNEL_FALLBACK:-sdpa}"
-
-# ---------------------- Lançamento da aplicação ----------------------
-PORT="${PORT:-7860}"
-OPTS="--server.port $PORT --server.name 0.0.0.0 --theme Default --allowed-paths /app/outputs,/app/ckpt"
-
-
-chmod +x /app/info.sh || true
-bash -lc "/app/info.sh"
+unset TRANSFORMERS_CACHE
 
+# ---------------------- Banner ----------------------
+./info.sh || true
 
-export APP_SELECT="vince"
+echo "🚀 Subindo serviços..."
 
+# Exemplo: subir UI mínima SD Img2Img (ajuste conforme seu app)
+# python app/app_animatediff_min.py
 
-case "${APP_SELECT:-seed}" in
-  seed)
-    echo "🚀 Subindo SeedVR Refine..."
-    python app_seedvr.py $OPTS
-    ;;
-  vince|*)
-    echo "🚀 Subindo VINCIE..."
-    python app_vince.py $OPTS
-    ;;
-esac
+# Ou subir VINCIE UI se for o caso
+python app/app_vince.py