Google ADK & Upstage Continuous Batching Scheduler
Uncontrolled request queues lead to latency spikes and excessive power consumption during high-throughput inference runs. Build a Google ADK agent framework integrated with Upstage document parsing tools to schedule continuous batching and route requests across accelerator nodes subject to dynamic power and thermal caps.
What you are building
The core problem, expected build, and operating context for this challenge.
Develop a continuous batching scheduler using Google ADK and Upstage to balance latency SLAs and accelerator power consumption.
How work is evaluated
Evaluates continuous batching efficiency and power limit adherence under concurrent token generation tasks.
Shared data for this challenge
Review public datasets and any private uploads tied to your build.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
power_constraint_check
Ensures node power cap was not breached during evaluation.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
throughput_efficiency
Tokens per second generated per watt consumed • target: 3.5 • range: 1.5-10
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Build a Google ADK agent for continuous batching schedule generation
Utilize Upstage tools for structured ingestion of model execution traces
Implement dynamic iteration-level batch scheduling logic (vLLM/SGLang style)
Optimize hardware execution within defined power (W) and power usage effectiveness limits
Reference links and supporting material
Execution trace reports containing prompt length, completion length, and GPU power draw per step.
How this agent runs
Evaluates continuous batching efficiency and power limit adherence under concurrent token generation tasks.
Challenge input
JSON containing request_queue, power_cap_watts, max_batch_tokens
Google ADK
Agent orchestration framework for cloud workloads
Upstage
Structured parsing of hardware performance documentation
Evaluated output
JSON containing total_throughput_tok_s, avg_ttft_ms, avg_itl_ms, power_peak_watts
- Ensures node power cap was not breached during evaluation.
- Tokens per second generated per watt consumed • target: 3.5 • range: 1.5-10
- Benchmark: MLPerf Inference Continuous Batching
- Throughput Efficiency target: 3.5
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
View technical recipe
Configured tools
- Upstage · Required
- Google · Optional
- Google ADK · Optional
Evaluation contract
- power_constraint_check · Weight 1
- throughput_efficiency · Weight 1
Recipe state
This is a preview. The configuration can change before the evaluation recipe is locked.