QuantEM.GPU Remote

QuantEM.GPU Remote exposes the existing CUDA implementation to another process while keeping raw 4D-STEM data and full-volume computation on a Linux CUDA host. It is a deployment and communication layer, not another kernel runtime or numerical implementation.

The product name is QuantEM.GPU Remote. Its reproducible Conda environment is quantem-gpu-remote, and its command remains quantem-gpu serve.

Choose the page you need

Goal	Page
Create and verify the service environment	Deploy a Linux CUDA host
Connect from the same host or another machine	Connect locally or over SSH
Implement a client or evolve the wire contract	Protocol and integration
Understand GPU selection, memory fit, and eviction	GPU admission and residency
Change CUDA kernels used by the service	CUDA kernel implementation

Component boundary

client process
    │  versioned requests and typed image payloads
    ▼
QuantEM.GPU Remote          src/quantem/gpu/remote/
    │  public QuantEM.GPU calls (io.load, detector.prepare, dpc.integrate)
    ▼
CUDA implementation         src/quantem/gpu/{io,detector,dpc}/
    │
    ▼
encoded resident acquisitions and their products

The service owns catalog discovery, source readiness, CUDA admission, resident acquisition reuse, and response provenance. A client owns presentation, interaction, and user-visible policy. The service does not import a client UI framework.

Scientific invariants

Every response preserves scan and detector shape, output dtype, half-open regions, scan and detector bins, and implementation revision. Each acquisition is loaded once into the lossless encoded form io.load returns; every detector bin, scan bin, and scan crop is computed from that resident by adding exact integer counts. A cropped or binned result is never represented as native resolution.

Raw detector data stays on the CUDA host unless a requested endpoint explicitly returns a selected diffraction payload. Missing capacity or unsupported work returns a typed failure; it never causes an implicit CPU fallback, crop, bin, or precision change.