gpuemu logo gpuemu

gpuemu for inference-as-a-service vendors

A signed Kernel Correctness Report your customers verify offline — SLA-grade evidence that your quantized, fused, and custom kernels still compute the right answer. For Fireworks, Together, Anyscale, Modal, Replicate, Baseten, and Modular.

For: Fireworks · Together · Anyscale · Modal · Replicate · Baseten · Modular

Inference vendors compete on cost, which means quantization, fused kernels, and custom sampling. Each of those is a place a kernel can be silently wrong — and the customer has no way to tell until quality drops.

The pain, with a case study

A 2026 Future AGI evaluation found that a quantized Llama-3.3-70B-Instruct kernel variant dropped JSON-mode adherence by 6 points and broke 3-step tool-calling chains — while serving the exact same OpenAI-compatible API. As the evaluation put it:

“The OpenAI-compatible API is a contract about request and response shape — not about weights, precision, kernel, or sampler defaults.”

Your customers are discovering this. The differentiator is being able to prove your kernels are correct.

The workflow gpuemu gives you

  • Signed Kernel Correctness Report. A per-op × per-dtype × per-shape pass matrix with a run id, a SHA-256 of the result database, and an ed25519 signature.
  • Offline verification. Customers check the signature with one OpenSSL command — they don’t have to trust your dashboard.
  • SLA evidence. Attach the report to your enterprise contracts as proof the served kernel computes the reference answer within calibrated tolerance.

Turn correctness from an unverifiable claim into a signed artefact.

Frequently Asked Questions

What's in a signed Kernel Correctness Report?

A per-op × per-dtype × per-shape pass matrix, a run id, a SHA-256 of the result database, and an ed25519 signature. Customers verify it offline with one OpenSSL command — no trust in your dashboard required.

Why does this matter for an OpenAI-compatible API?

The API contract guarantees request/response shape — not which kernel, precision, or sampler default served the tokens. A signed correctness report is how you make a guarantee about the thing the API doesn't cover.

Stop shipping silently-wrong kernels

Open source, dual-licensed MIT / Apache-2.0. Validate your first kernel in five minutes — or talk to us about an enterprise pilot.