Skip to main content
Back to challenges
Challenge

LangChain & AutoGen KV-Cache Multi-Tenant Capacity Planner

Build a multi-tenant KV-cache memory capacity manager using LangChain and AutoGen multi-agent simulation to analyze memory fragmentation under continuous batching. Implement dynamic cache reallocation algorithms that prevent out-of-memory errors and maintain sequence throughput above 850 tokens per second.

EngineeringHosted by Vera
Sign in free to start
Challenge brief

What you are building

The core problem, expected build, and operating context for this challenge.

Orchestrate dynamic KV-cache allocation under continuous batching using a hybrid LangChain and AutoGen multi-agent simulation framework.

Delivery guide

How work is evaluated

Evaluation

Evaluates KV-cache planner accuracy under simulated bursty tenant workloads.

Datasets

Challenge datasets

Public datasets, plus the private datasets you have access to.

Evaluation rubric

How submissions are scored

These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.

Dimensions
2 scoring checks
Binary
2 pass or fail dimensions
Ordinal
0 scaled dimensions
Dimension 1

zero_oom_policy

Ensures no OOM events occurred during peak multi-tenant contention.

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Dimension 2

throughput_tok_per_sec

Measures continuous token generation throughput. • target: 850 • range: 0-2000

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Learning goals

What you should walk away with

  • Build a LangChain-based KV-cache capacity planner for continuous batching runtimes

  • Simulate multi-tenant memory contention using AutoGen collaborative agent negotiation

  • Implement dynamic block-paged memory re-allocation algorithms to eliminate internal fragmentation

  • Achieve zero Out-Of-Memory (OOM) preemptions while sustaining 850+ tok/s output

Resources and assets

Reference links and supporting material

Dataset notes

Synthetic telemetry trace dataset representing multi-tenant LLM requests with dynamic sequence expansion and variable context lengths.

Challenge setup

Evaluates KV-cache planner accuracy under simulated bursty tenant workloads.

Challenge input

JSON array of incoming requests with start times, context lengths, and priority tiers.

LangChain

Provides structured tool abstractions for VRAM profiling.

Microsoft AutoGen

Simulates multi-tenant bargaining for memory allocation.

Evaluated output

JSON execution trace showing block allocations, swap events, and total throughput achieved.

Evaluation checks
  • Ensures no OOM events occurred during peak multi-tenant contention.
  • Measures continuous token generation throughput. • target: 850 • range: 0-2000
Configured evaluation evidence
  • Benchmark: vLLM Multi-Tenant Benchmark
  • Throughput Tok Per Sec target: 850
  • 1 public reference case
Configured runtime
  • Python execution harness
  • Python sandbox (unavailable on Versalist)

Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.

View technical recipe

Configured tools

Action Space
  • Langchain · Required
  • LangChain · Optional
  • AutoGen · Optional
Orchestration
  • Langchain · Required
  • LangChain · Optional

Evaluation contract

  • zero_oom_policy · Weight 1
  • throughput_tok_per_sec · Weight 1

Metrics

  • throughput_tok_per_sec · Target 850

Benchmarks

  • vLLM Multi-Tenant Benchmark
  • general

Reference cases

  • 1 public reference case

Workflow tasks

  1. Task 1

    Input: JSON array of incoming requests with start times, context lengths, and priority tiers.

    Output: JSON execution trace showing block allocations, swap events, and total throughput achieved.

    Execution harness: Python

Evaluation tools

  • LangChain · Agent Framework
  • Microsoft AutoGen · Agent System

Recipe freeze marker

No recipe freeze time is recorded.

Frequently Asked Questions about LangChain & AutoGen KV-Cache Multi-Tenant Capacity Planner