First, congratulations on the models, thank you very much for releasing them.

I haven't been able to get any model (9b - 35b) to reason with vllm:

TEMPLATE_URL="${TEMPLATE_URL:-https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B/resolve/main/chat_template.jinja}"

!uv pip install --system --force-reinstall -U --pre vllm \
  --extra-index-url https://wheels.vllm.ai/nightly/cu129 \
  --extra-index-url https://download.pytorch.org/whl/cu129 \
  --index-strategy unsafe-best-match

!python -c "import vllm; print(vllm.__version__)"
0.23.1rc1.dev578+g72f639927

code serve_ornith_vllm.sh

#!/usr/bin/env bash
# ---------------------------------------------------------------------------
# serve_ornith_vllm.sh  (versión DETACHED)
# Levanta Ornith-1.0 con vLLM en segundo plano (OpenAI-compatible).
#
# Igual que el serve de Gemma 4: nohup + disown -> el servidor sobrevive aunque
# cierres la shell o termine este script. Guarda el PID y espera a que el
# endpoint /health responda antes de devolverte el control.
#
# Ornith es un modelo de razonamiento (<think>...</think>) con tool-calling
# estilo Qwen3, así que activamos --reasoning-parser qwen3 y --tool-call-parser
# qwen3_xml para obtener reasoning_content y tool_calls.
#
# Uso:
#   chmod +x serve_ornith_vllm.sh stop_ornith.sh
#   ./serve_ornith_vllm.sh           # arranca en background y espera al health
#   ./stop_ornith.sh                 # lo apaga
#
# El número de GPUs (tensor-parallel-size) se DETECTA AUTOMÁTICAMENTE.
#
# Variables de entorno configurables (con sus valores por defecto):
#   MODEL          deepreinforce-ai/Ornith-1.0-9B
#   SERVED_NAME    Ornith-1.0-9B
#   TP_SIZE        (auto-detectado)  -> fuerza un valor para sobreescribir
#   PORT           8000
#   HOST           0.0.0.0
#   MAX_LEN        262144   (bajalo si te quedas sin VRAM/KV cache, p.ej. 16384)
#   GPU_UTIL       0.90
#   API_KEY        EMPTY    (token que exigira el server; cliente debe enviarlo)
#   LOG            ornith_vllm.log
#   PIDFILE        ornith_server.pid
#   SHOW_PROGRESS  1        (=1 muestra el log en vivo al cargar; =0 silencioso)
#
# Ejemplos:
#   ./serve_ornith_vllm.sh                       # usa TODAS las GPUs visibles
#   TP_SIZE=2 ./serve_ornith_vllm.sh             # fuerza solo 2 GPUs
#   MAX_LEN=16384 ./serve_ornith_vllm.sh         # recorta contexto (recomendado en 1 GPU)
#   SHOW_PROGRESS=0 ./serve_ornith_vllm.sh       # arranque silencioso
# ---------------------------------------------------------------------------
set -euo pipefail

# ---------------------------------------------------------------------------
# Deteccion automatica de GPUs
# Prioridad: 1) TP_SIZE manual  2) CUDA_VISIBLE_DEVICES  3) nvidia-smi
# ---------------------------------------------------------------------------
detect_gpus() {
    if [[ -n "${CUDA_VISIBLE_DEVICES:-}" ]]; then
        echo "${CUDA_VISIBLE_DEVICES}" | tr ',' '\n' | grep -cE '^[0-9]+$' || echo 0
        return
    fi
    if command -v nvidia-smi >/dev/null 2>&1; then
        nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | grep -c . || echo 0
        return
    fi
    echo 0
}

if [[ -n "${TP_SIZE:-}" ]]; then
    echo "[info] TP_SIZE forzado manualmente: ${TP_SIZE}"
else
    NUM_GPUS="$(detect_gpus)"
    if [[ "${NUM_GPUS}" -lt 1 ]]; then
        echo "[error] No se detecto ninguna GPU NVIDIA." >&2
        echo "        vLLM requiere GPU. Verifica los drivers con 'nvidia-smi'." >&2
        echo "        Si quieres forzar un valor de todos modos: TP_SIZE=1 ./serve_ornith_vllm.sh" >&2
        exit 1
    fi
    TP_SIZE="${NUM_GPUS}"
    echo "[info] GPUs detectadas: ${NUM_GPUS} -> tensor-parallel-size=${TP_SIZE}"

    if (( TP_SIZE > 1 )) && (( (TP_SIZE & (TP_SIZE - 1)) != 0 )); then
        echo "[warn] ${TP_SIZE} no es potencia de 2; si vLLM falla al cargar," >&2
        echo "       fuerza una potencia de 2 (p.ej. TP_SIZE=2 o TP_SIZE=4)." >&2
    fi
fi

# ---------------------------------------------------------------------------
# Configuracion del modelo / servidor
# ---------------------------------------------------------------------------
# MODEL="${MODEL:-deepreinforce-ai/Ornith-1.0-9B}"
# SERVED_NAME="${SERVED_NAME:-Ornith-1.0-9B}"
MODEL="${MODEL:-deepreinforce-ai/Ornith-1.0-35B}"
SERVED_NAME="${SERVED_NAME:-Ornith-1.0-35B}"
PORT="${PORT:-8000}"
HOST="${HOST:-0.0.0.0}"
MAX_LEN="${MAX_LEN:-262144}"
GPU_UTIL="${GPU_UTIL:-0.90}"
API_KEY="${API_KEY:-EMPTY}"
LOG="${LOG:-ornith_vllm.log}"
PIDFILE="${PIDFILE:-ornith_server.pid}"
SHOW_PROGRESS="${SHOW_PROGRESS:-1}"

# Evita arrancar dos veces sobre el mismo PIDFILE
if [ -f "$PIDFILE" ] && kill -0 "$(cat "$PIDFILE")" 2>/dev/null; then
    echo "[error] Ya hay un servidor corriendo (PID $(cat "$PIDFILE"), $PIDFILE)." >&2
    echo "        Apagalo primero con ./stop_ornith.sh" >&2
    exit 1
fi

echo "============================================================"
echo "  Sirviendo: ${MODEL}"
echo "  Nombre expuesto: ${SERVED_NAME}"
echo "  GPUs (TP): ${TP_SIZE}   |   Contexto max: ${MAX_LEN}"
echo "  Endpoint: http://${HOST}:${PORT}/v1   |   Log: ${LOG}"
echo "============================================================"

export VLLM_API_KEY="${API_KEY}"

# ─── Chat template oficial de Ornith (dispara el modo <think>) ───
TEMPLATE="${TEMPLATE:-ornith_chat_template.jinja}"
# TEMPLATE_URL="${TEMPLATE_URL:-https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B/resolve/main/chat_template.jinja}"
TEMPLATE_URL="${TEMPLATE_URL:-https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B/resolve/main/chat_template.jinja}"
if [ ! -f "$TEMPLATE" ]; then
  echo "Descargando chat template de Ornith..."
  curl -sfL -o "$TEMPLATE" "$TEMPLATE_URL" \
    || { echo "ERROR: no se pudo descargar el template"; exit 1; }
fi
TEMPLATE_PATH="$(pwd)/$TEMPLATE"



# ---------------------------------------------------------------------------
# Arranque DETACHED (sobrevive al cierre de la shell)
# ---------------------------------------------------------------------------
nohup vllm serve "${MODEL}" \
    --served-model-name "${SERVED_NAME}" \
    --tensor-parallel-size "${TP_SIZE}" \
    --host "${HOST}" --port "${PORT}" \
    --max-model-len "${MAX_LEN}" \
    --gpu-memory-utilization "${GPU_UTIL}" \
    --enable-prefix-caching \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_xml \
    --reasoning-parser qwen3 \
    --chat-template "${TEMPLATE_PATH}" \
    --default-chat-template-kwargs '{"enable_thinking": true}' \
    --trust-remote-code \
    --attention-backend TRITON_ATTN \
    # --speculative-config '{"method": "dflash", "model": "z-lab/Qwen3.5-9B-DFlash", "num_speculative_tokens": 8, "draft_tensor_parallel_size": 1}' \
    # --speculative-config '{"method": "dflash", "model": "z-lab/Qwen3.6-35B-A3B-DFlash", "num_speculative_tokens": 8, "draft_tensor_parallel_size": 1}' \
     --limit-mm-per-prompt '{"image":4,"video":0}' \
     --max-num-batched-tokens 4176 \
     --kv-cache-dtype bfloat16 \
     --max-num-seqs 10 \
    --api-key "${API_KEY}" \
    > "${LOG}" 2>&1 &

SERVER_PID=$!
echo "${SERVER_PID}" > "${PIDFILE}"
disown
echo "Servidor lanzado (PID ${SERVER_PID}, guardado en ${PIDFILE}). Logs: ${LOG}"

# ---------------------------------------------------------------------------
# Progreso de carga en vivo (opcional). tqdm de vLLM escribe en el log.
# ---------------------------------------------------------------------------
TAIL_PID=""
if [ "$SHOW_PROGRESS" = "1" ]; then
    echo "── Progreso de carga (SHOW_PROGRESS=1) ───────────────────────"
    tail -f "$LOG" &
    TAIL_PID=$!
fi
stop_tail() { [ -n "$TAIL_PID" ] && kill "$TAIL_PID" 2>/dev/null || true; TAIL_PID=""; }

# ---------------------------------------------------------------------------
# Espera a que cargue el modelo (/health no requiere API key)
# ---------------------------------------------------------------------------
[ "$SHOW_PROGRESS" = "1" ] || echo "Esperando a que el modelo cargue (puede tardar varios minutos)..."
until curl -sf "http://localhost:${PORT}/health" > /dev/null 2>&1; do
    if ! kill -0 "$SERVER_PID" 2>/dev/null; then
        stop_tail
        echo "ERROR: el servidor murió durante el arranque. Últimas líneas:"
        tail -n 40 "$LOG"
        rm -f "$PIDFILE"
        exit 1
    fi
    sleep 3
done

stop_tail
[ "$SHOW_PROGRESS" = "1" ] && echo "──────────────────────────────────────────────────────────────"

echo "OK: servidor listo en http://localhost:${PORT}/v1 (sigue corriendo en segundo plano)"
echo "  Modelo expuesto:    ${SERVED_NAME}"
echo "  Apaga el servidor:  ./stop_ornith.sh   (o: kill \$(cat ${PIDFILE}))"