Skip to content

[feat] Hardware-validated CI gate for NPU regression — proof on this repo? #11

Description

@ashish-frozo

Hi — saw npurun. NPU-first LLM runtime on Snapdragon X Elite (Windows on ARM) is a tight target and the kind of project where small QNN graph changes can silently route ops to CPU between releases. Your topic tags (genie, npu, qualcomm, snapdragon-x-elite, windows-arm64) put you squarely in the validation gap we work on.

We built EdgeGate (https://edgegate.frozo.ai) around this: a CI gate that runs your runtime on real Snapdragon via Qualcomm AI Hub, asserts NPU routing op-by-op, and blocks the merge if latency or memory regresses. Free tier covers small projects.

Would be a natural validation layer for a runtime project like this. Happy to set up a one-shot gate on npurun as a proof — you'd see it as a passing/failing check on a sample PR, no commitment, just data.

Open to it? Either way, nice work on the runtime.

— Ashish (EdgeGate)

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions