Spaces:

---

dystrio

/

gpu-optimizer

like 2

Running

I built Dystrio as an interactive demo to make GPU placement decisions visible
and explainable for distributed inference on Kubernetes.

This Space:

• Ingests real PyTorch / NCCL profiler traces
• Reconstructs rank-to-rank communication graphs
• Detects whether patterns are stable across runs
• Emits Kubernetes-native podAffinity YAML only when it’s provably safe
• Explains why a recommendation is strong, weak, or downgraded

The UI here is static by design — all real analysis runs in backend services —
so it’s safe to explore without exposing models or cluster internals.

Primary focus:

• Hugging Face TGI
• Tensor-parallel / multi-GPU inference
• Latency-sensitive distributed workloads

I’m actively looking for feedback and design partners running real inference
systems on Kubernetes.

Happy to answer questions or walk through specific workloads.