fix: fix G4 runtime errors

serve-vllm.sh

#!/bin/bash

# Ensure working directory is /content
cd /content

## G4 runtime fixes
pip uninstall flashinfer
export VLLM_USE_FLASHINFER=0
export VLLM_USE_FLASHINFER_SAMPLER=0

# 1. Clean up old processes and conflicting dependencies
pkill -9 -f "vllm.entrypoints" 2>/dev/null || true
pkill -9 -f "cloudflared" 2>/dev/null || true
pip uninstall -y -q torchaudio 2>/dev/null || true
rm -f /content/cloudflare.log /content/vllm.log

# 2. Check and install Cloudflare tunnel binary if not present
if ! command -v cloudflared &> /dev/null; then
    wget -q -nc https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64.deb
    dpkg -i cloudflared-linux-amd64.deb > /dev/null 2>&1
fi

# 3. Start vLLM server in daemon mode (262k Context)
echo "1. Starting vLLM daemon (Qwen/Qwen3.8-27B, 262k context)..."
nohup python3 -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen3.8-27B \
  --port 8000 \
  --dtype bfloat16 \
  --max-model-len 262144 \
  --gpu-memory-utilization 0.92 \
  --enforce-eager \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser deepseek_r1 \
  --attention-backend flash_attn \
  --trust-remote-code > /content/vllm.log 2>&1 &

VLLM_PID=$!

# 4. Check for immediate startup/argument errors (within first 2 seconds)
sleep 2
if ! kill -0 $VLLM_PID 2>/dev/null; then
    echo -e "\n❌ vLLM failed to start due to argument/configuration errors! Logs:"
    cat /content/vllm.log
    exit 1
fi

# 5. Wait for model loading and health check
echo -n "⏳ Loading model weights into GPU memory, please wait"
READY=0
for i in {1..60}; do
  if ! kill -0 $VLLM_PID 2>/dev/null; then
    echo -e "\n❌ vLLM process terminated unexpectedly! Error logs:"
    tail -n 35 /content/vllm.log
    exit 1
  fi

  if curl -s -f http://localhost:8000/health > /dev/null 2>&1; then
    echo -e "\n✅ vLLM server is healthy and ready!"
    READY=1
    break
  fi

  echo -n "."
  sleep 3
done

if [ "$READY" -ne 1 ]; then
  echo -e "\n❌ Timeout: vLLM server failed to become ready in time."
  tail -n 35 /content/vllm.log
  exit 1
fi

# 6. Start Cloudflare tunnel in the background
echo -e "\n2. Establishing Cloudflare tunnel..."
nohup cloudflared tunnel --url http://127.0.0.1:8000 > /content/cloudflare.log 2>&1 &

# 7. Extract the public tunnel URL
TUNNEL_URL=""
for i in {1..30}; do
  if [ -f /content/cloudflare.log ]; then
    TUNNEL_URL=$(grep -o 'https://[a-zA-Z0-9-]*\.trycloudflare\.com' /content/cloudflare.log | head -n 1)
    if [ -n "$TUNNEL_URL" ]; then
      break
    fi
  fi
  sleep 1
done

# 8. Print connection details
echo "================================================================="
if [ -n "$TUNNEL_URL" ]; then
  echo "🚀 PUBLIC API URL  : $TUNNEL_URL"
  echo "👉 OPENAI BASE URL : $TUNNEL_URL/v1"
else
  echo "❌ Failed to retrieve Cloudflare URL. cloudflare.log contents:"
  cat /content/cloudflare.log
fi
echo "🤖 MODEL           : Qwen/Qwen3.8-27B"
echo "🧠 CONTEXT SIZE    : 262,144 tokens"
echo "🛠️ TOOL CALLING    : Enabled (qwen3_coder parser)"
echo "================================================================="