Upload start.sh

start.sh

@@ -1,18 +1,40 @@
-#!/bin/bash
+#!/usr/bin/env bash
+set -euo pipefail
 
-set -e  # Para em erros críticos
-set -u  # Trata variáveis não definidas como erro
+echo "======================================================="
+echo " ADUC-SDR — Start (VINCIE/SeedVR, 8× L40S)"
+echo "======================================================="
 
-# Diretórios persistentes
-REPOS_DIR="/data/repos"
-MODELS_DIR="/data/models"
+# ---------------------- Env base ----------------------
+export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"
+export TORCH_DTYPE="${TORCH_DTYPE:-bfloat16}"
 
+# 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}"
 
-# Cria diretórios persistentes se não existirem
-mkdir -p "$REPOS_DIR" "$MODELS_DIR" || true
-chown -R appuser:appuser "$REPOS_DIR" "$MODELS_DIR"  || true
-chmod -R 777 "$REPOS_DIR" "$MODELS_DIR" || true
+# 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="${CUDA_DEVICE_ORDER:-PCI_BUS_ID}"
+export PYTORCH_CUDA_ALLOC_CONF="${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="${NCCL_SOCKET_IFNAME:-lo}"
+export NCCL_BLOCKING_WAIT=1
+export TORCH_NCCL_BLOCKING_WAIT=1
+export NCCL_TIMEOUT="${NCCL_TIMEOUT:-600}"
 
+# HF caches
+export HF_HOME="${HF_HOME:-/app/.cache/huggingface}"
+unset TRANSFORMERS_CACHE
 
 
 # ---------------------- Builder (Apex + Q8) ----------------------
@@ -37,27 +59,17 @@ else
 fi
 
 
+# ---------------------- Banner ----------------------
+./info.sh || true
 
-# Inicializa VincieService de forma persistente
-log "Inicializando VincieService com persistência em /data..."
-python -c "
-import os
-from services.vincie import VincieService
 
-# Configura caminhos persistentes via ENV (ajuste se necessário)
-os.environ['REPOS_DIR'] = '$REPOS_DIR'
-os.environ['MODELS_DIR'] = '$MODELS_DIR'
+ls -la /app && ls -R /app | head -n 2000
 
-svc = VincieService(repo_dir='$REPOS_DIR/VINCIE', ckpt_dir='$MODELS_DIR')
-svc.ensure_repo()  # Garante repositório persistente
-svc.ensure_model() # Garante modelo persistente
-log('VincieService inicializado com sucesso.')
-" || log "Erro ao inicializar VincieService"
 
-# Inicia o app principal (Gradio ou similar)
-log "Iniciando aplicação principal..."
-if [ "${1:-gradio}" = "gradio" ]; then
-    python /app/app_vince.py
-else
-    exec "$@"
-fi
+echo "🚀 Subindo serviços..."
+
+# Exemplo: subir UI mínima SD Img2Img (ajuste conforme seu app)
+# python app_animatediff_min.py
+
+# Ou subir VINCIE UI se for o caso
+python app_vince.py