scripts/rtx-docker/deepseek_only.sh

#!/bin/bash
# DeepSeek-OCR + DeepSeek-OCR-2 na RTX 6000 (NuMarkdown já fechou 6/6 no master).
# FIX: NGC 26.03 não traz torchaudio → inspeção da arquitetura falha. pip install torchaudio antes do serve.
set -uo pipefail
HF=$(grep -oE 'hf_[A-Za-z0-9]+' ~/.scherm_hf_token | head -1)
SIF="nvcr.io/nvidia/vllm:26.03.post1-py3"
GPU="NVIDIA_RTX_PRO_6000"; CACHE=~/.cache/huggingface
dk(){ docker "$@"; }
free_port(){ local c; c=$(dk ps -q --filter "publish=8000"); [ -n "$c" ] && { echo "  liberando 8000 ($c)"; dk rm -f $c >/dev/null 2>&1; sleep 3; }; }

run_ocr(){ # model, prompt_style, extra_serve_args, label
  local M="$1" STYLE="$2" EXTRA="$3" L="$4"
  echo "########## $L ($M) ##########"
  # download completo ANTES (1 processo só — sem deadlock de I/O)
  dk run --rm -v $CACHE:/root/.cache/huggingface -e HF_TOKEN="$HF" "$SIF" \
     python3 -c "from huggingface_hub import snapshot_download; snapshot_download('$M',token='$HF',max_workers=8)" >/dev/null 2>&1 || { echo "[dl FAIL] $M"; return; }
  free_port; dk rm -f vocr >/dev/null 2>&1; sleep 2
  # entrypoint custom: instala torchaudio (dep ausente no NGC) e SÓ então sobe o vllm serve
  dk run -d --name vocr --gpus all --ipc host --shm-size 32gb -p 8000:8000 \
    -v $CACHE:/root/.cache/huggingface -v ~/bench/pages:/pages -v ~/bench/scripts:/scripts -e HF_TOKEN="$HF" \
    --entrypoint bash "$SIF" -lc "pip install -q torchaudio 2>&1 | tail -1; exec vllm serve '$M' --served-model-name ocrmodel --gpu-memory-utilization 0.90 --max-model-len 16384 --trust-remote-code $EXTRA" >/dev/null 2>&1
  local ok=0
  for i in $(seq 1 150); do  # +tempo: pip install torchaudio (~150MB) + load
    curl -sf -m3 http://localhost:8000/v1/models 2>/dev/null | grep -q '"ocrmodel"' && { ok=1; break; }
    dk ps -q -f name=vocr | grep -q . || break
    sleep 10
  done
  if [ "$ok" != 1 ]; then echo "[!] $L não subiu"; dk logs --tail 15 vocr 2>&1 | tail -10; dk rm -f vocr >/dev/null 2>&1; return; fi
  dk exec vocr sh -c 'rm -f /pages/._*'
  dk exec vocr python3 /scripts/run_ocr_accuracy.py --served-model ocrmodel --base-url http://localhost:8000/v1 \
    --pages-dir /pages --doc-sizes 1 5 50 --concurrencies 1 4 --reps 3 --prompt-style "$STYLE" --model-label "$L" \
    --out /scripts/ocr_compare --tag "$GPU" --run-id "$L" 2>&1 | tail -10
  dk rm -f vocr >/dev/null 2>&1; sleep 3
}

DS_EXTRA="--logits_processors vllm.model_executor.models.deepseek_ocr:NGramPerReqLogitsProcessor --no-enable-prefix-caching --mm-processor-cache-gb 0"
run_ocr "deepseek-ai/DeepSeek-OCR"   "deepseek" "$DS_EXTRA" "deepseek-ocr"
run_ocr "deepseek-ai/DeepSeek-OCR-2" "deepseek" "$DS_EXTRA" "deepseek-ocr-2"
echo "===== DEEPSEEK OCR COMPLETO ====="