CrewAI Distributed Interconnect & Tensor Parallelism Simulator
Design an interconnect-aware distributed inference simulator using CrewAI agents to model multi-node tensor parallel and pipeline parallel workloads across NVLink and AWS EFA topologies. Implement automated bottleneck detection that optimizes cross-node message passing efficiency, achieving a 20% latency reduction in simulated multi-GPU clusters.
What you are building
Coordinate role-playing CrewAI agents to simulate NVLink and AWS EFA cross-node bandwidth constraints for distributed tensor parallelism.
How work is evaluated
Evaluates distributed parallel configuration suggestions on simulated multi-node clusters.
Datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
gpu_topology_conservation
Validates TP * PP equals total cluster GPU count.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
latency_reduction_pct
Simulated latency improvement percentage over baseline naive TP=16. • target: 20 • range: 0-50
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Learning goals
Build a multi-agent distributed simulation using CrewAI framework
Model NVLink 900 GB/s intra-node and AWS EFA 800 Gbps inter-node network topologies
Simulate AllReduce and AllGather collective communication overheads in Tensor Parallelism (TP=8, PP=2)
Optimize parallel splitting parameters to reduce inter-node communication latency by 20%
Resources
Network trace profiles of NCCL AllReduce and AllGather operations captured across multi-node H100 and AWS Trainium2 cluster topologies.
Challenge setup
Evaluates distributed parallel configuration suggestions on simulated multi-node clusters.
Challenge input
JSON cluster spec containing GPU count, node count, interconnect bandwidth, and model size (70B parameters).
CrewAI
Multi-agent framework for collaborative role-playing simulation.
Evaluated output
JSON optimization layout specifying TP size, PP size, expected AllReduce latency, and throughput gain %.
- Validates TP * PP equals total cluster GPU count.
- Simulated latency improvement percentage over baseline naive TP=16. • target: 20 • range: 0-50
- Benchmark: Megatron-LM Distributed Benchmark
- Latency Reduction Pct target: 20
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Shows the first task, up to three tools, and up to three evidence items.
View technical recipe
Configured tools
- CrewAI · Required
- crewAI · Optional
- Anthropic · Optional
Evaluation contract
- gpu_topology_conservation · Weight 1
- latency_reduction_pct · Weight 1
Metrics
- latency_reduction_pct · Target 20
Benchmarks
- Megatron-LM Distributed Benchmark
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON cluster spec containing GPU count, node count, interconnect bandwidth, and model size (70B parameters).
Output: JSON optimization layout specifying TP size, PP size, expected AllReduce latency, and throughput gain %.
Execution harness: Python
Evaluation tools
- CrewAI · Agent Framework
Recipe freeze marker
No recipe freeze time is recorded.