SGLang RadixAttention cluster live on RTX 5090 (32GB VRAM)
Infrastructure / Dedicated 8h Runtime

Rent a high-throughput GPU container. Split the cost with a friend.

One dedicated RTX 5090 instance issues 2 distinct API keys with simultaneous, unthrottled token streaming. Runs in volatile RAM (tmpfs) with zero prompt logging. Hard termination at T+8h. Flat ₹757 via UPI.

• Qwen 2.5 Coder 32B (AWQ 4-bit)
• 180–350 tok/s (RadixAttention)
• Air-Gapped (Outbound DROP)
• ₹378 per person (2 devs)
sglang-cluster://127.0.0.1:8000 — Dual Concurrent Streams
● 2 ACTIVE SESSIONS | MEMORY: 26.4 / 32 GB
[BEARER: sk-friend1-a77] 242 tok/s • TTFT 38ms
Prompt: "Write a high-performance LRU cache in Go"
// Thread-safe concurrent LRU with DoublyLinkedList type LRUCache struct { capacity int mutex sync.RWMutex items map[string]*Node list *DoublyLinkedList } func (c *LRUCache) Get(key string) (interface{}, bool)
[BEARER: sk-friend2-b88] 229 tok/s • TTFT 41ms
Prompt: "Refactor FastAPI router to async generator streaming"
async def stream_generator(request_body: dict): async with httpx.AsyncClient() as client: async for chunk in client.aiter_bytes(): yield f"data: {chunk}\n\n" return StreamingResponse(stream_generator(body))
SGLang RadixAttention: Shared prefix KV cached in VRAM • Zero cross-talk leakage Memory Overwrite on Exit: 100%
[01 / ARCHITECTURE]

How the Dual-Key Sandbox Works

A hardened container supervisor that bridges multi-tenant developer teams to an isolated local inference daemon.

01 / DUAL BEARER AUTH

Independent API Keys

The reverse proxy validates Authorization: Bearer sk-friend1-... and sk-friend2-... independently. Neither engineer sees the other's request history or prompt state.

02 / AIR-GAPPED RAM

Zero NVMe Retention

Container mounts temporary directories onto volatile tmpfs (system RAM). Post-boot iptables rules drop all outbound traffic: the model cannot call external APIs or send telemetry.

03 / HARD TERMINATION

Watchdog Destruction

A background daemon enforces strict lifetime limits. At T+8.0 hours, container initiates self-termination (SIGTERM 1) and fires a hard instance delete to the GPU cloud provider.

[02 / SPECIFICATION & PRICING]

Configure Your Instance

Select duration and team size. One prepaid UPI transfer. No credit card required.

Hardware Allocation Spec
Accelerator NVIDIA RTX 5090 (32GB GDDR7)
Engine SGLang v0.4 RadixAttention
Model Qwen 2.5 Coder 32B
Max Context 32,768 tokens (AWQ 4-bit)
Temporary Storage tmpfs volatile RAM (Zero Disk)
Network Boundary iptables AIR-GAP (0 outbound)
Billing Model Prepaid Flat Rate via UPI
SESSION DURATION: 8 Hours (Standard Sprint)
DEVELOPER KEYS (SEATS): 2 Keys (Pair Programming)
FOUNDATION MODEL: Qwen 2.5 Coder 32B
Prepaid Total
₹757 / 8h runtime
Split Per Developer
₹378 each
2 distinct API keys issued
Deploy Session via Telegram Bot (@Nimeshpcbot) →
Order reference auto-encoded into dynamic UPI intent • Zero KYC required
[03 / CLIENT CONFIGURATION]

Plug into Your Local IDE

Full drop-in compatibility with any client speaking the OpenAI /v1/chat/completions format.

// Cursor: Settings -> Models -> Override OpenAI Base URL
{
  "models": [
    {
      "title": "Dedicated Qwen 2.5 Coder 32B",
      "provider": "openai",
      "model": "default",
      "apiBase": "https://sandbox.pkai.in/v1",
      "apiKey": "sk-friend1-xxxxxxx" // Or sk-friend2-xxxxxxx for your friend
    }
  ]
}
[04 / TECHNICAL VERIFICATION]

Frequently Asked Questions