Understand TechField recordAn AI application stack on two office Sparks
UT’s move from rented cloud GPUs to machines in its own office, and the engineering record behind it.
INFRASTRUCTURE / ACROSS PLATFORMS
Build on someone else’s
starting point.
Model servers, runtime configurations and deployment recipes. The machine, software, evidence and open questions in one place.
10 recordsHow to read the evidence
Understand TechField record
NVIDIA · indexed by the LabRecipeA browser chat interface, served by your Spark
NVIDIA’s Open WebUI and Ollama guide, connected to the hardware and operating questions it raises.
Understand TechField recordLocal AI inside an accounting and advisory firm
A professional-services firm’s on-premises deployment: applications, confidentiality and the expertise to operate the system. The firm is not named at its request.
NVIDIA · indexed by the LabRecipeAn inference endpoint for your local applications
Connect applications to a model server on a GB10 or GB300 reference platform using vLLM.
NVIDIA · indexed by the LabRecipeStructured responses from a local Station
Explore SGLang serving, structured output and the conditions to check on a GB300 workstation.
NVIDIA · indexed by the LabRecipeA GGUF model server on GB10
A source-based starting point for running llama.cpp on Spark and documenting your own configuration.
AMD · indexed by the LabRecipeA local model server on Ryzen AI Max
AMD’s gpt-oss-20b example connects a GGUF model to llama.cpp with ROCm acceleration.
A local language model in the Apple ecosystem
Use MLX-LM and converted model weights as a starting point on Apple Silicon.
Qualcomm · indexed by the LabRecipeCompile a model for a Cloud AI workstation
Follow Qualcomm’s Efficient Transformers path from model weights to accelerator execution.
NVIDIA · indexed by the LabRecipeOpen-weight models on an RTX PRO workstation
NVIDIA’s local gpt-oss support, connected to the details needed for an OEM workstation evaluation.
The useful part is what you changed.
Follow a recipe, adapt it to your machine, and share the outcome. Include unsuccessful attempts.