---
title: "Install and Use NVIDIA PAIR with llama.cpp — Troubleshooting"
canonical: "https://build.nvidia.com/rtx-spark/pair/troubleshooting.md"
---

## Troubleshooting

Start with the symptom and then inspect **Settings → Service**, the desktop
error surface, or the terminal interface's **Errors** tab for more detail.
In the terminal interface, also check **Service** for worker state and **Logs**
for engine installation or startup output.

| Symptom | Usual meaning | What to do |
| --- | --- | --- |
| PAIR remains on **Loading...** | A background service did not start correctly. | Wait one or two minutes, then inspect **Settings → Service** and restart the affected service. |
| llama.cpp is missing, or the terminal shows the old ten-tab layout | The installed PAIR version predates this playbook's 1.0.0 workflow. | Check the PAIR version and install the matching 1.0.0 package for your OS and architecture. |
| An installer does not match the system | The architecture choice is incorrect. | Use Windows ARM64 for RTX Spark, Linux ARM64 for DGX Spark, and macOS ARM64 for Apple silicon. Intel/AMD systems use x64 packages, named `amd64` for Debian/Ubuntu. |
| LM Studio remains after uninstalling PAIR 1.0.0 | LM Studio installations created by older PAIR versions are not removed by the newer PAIR uninstaller. | If you also want to remove LM Studio, uninstall that older installation manually from `~/.lmstudio` on Linux or macOS, or `%USERPROFILE%\.lmstudio` on Windows. |
| A node is not discovered | mDNS is blocked or unavailable. | Confirm both nodes are on the same trusted local network, allow `5353/udp`, or add the node by IP address. |
| Pairing stalls or fails | The invitation expired, the PIN is incorrect, or port `14321` is blocked. | Start a new invitation, enter the new PIN, and confirm cluster ports are reachable. |
| Connection refused | Nothing is listening at that address. | Copy the current URL from **Endpoints** and confirm the PAIR service is running. |
| `403` from another machine | PAIR endpoints accept loopback traffic only. | Run PAIR on the machine hosting the client and use its local endpoint. |
| `502` with `no active node` | No node is eligible for the request. | Start a compatible engine and make the requested model available on at least one online node. |
| A model is listed but the inference request fails | The request's model ID, API route, or engine does not match the prepared model. | Copy the exact ID from `<PAIR_BASE_URL>/v1/models`. For llama.cpp, preserve `owner/repository:quantization` and use `/v1/chat/completions`, not Ollama's `/api/chat`. |
| `400` or `422` | The request is malformed. | Correct its JSON, route, model name, or required fields; malformed requests are not retried. |
| A response arrives but **Jobs** is empty | Another process owns the expected proxy port. | Check **Endpoints** and **Settings → Service**, then change the PAIR port or stop the conflicting process. |
| Requests do not use every GPU | PAIR routes each request to one eligible node. | Send independent requests and verify **Ran on** for each job. PAIR does not split one request across GPUs. |
| Desktop and terminal behavior conflicts | Both PAIR interfaces are running on one system. | Stop one interface and use only the desktop application or terminal interface. |
| `nvpair` is not found in native Windows PowerShell | The packaged launcher has not been created or the shell has not picked up its PATH entry. | Open the installed PAIR app once, quit it, and open a new PowerShell window before running `nvpair`. |
| A llama.cpp model download or load fails | The GGUF selection or available resources need checking. | Confirm the selected engine is llama.cpp, choose a compatible catalog entry, and inspect **Errors** and **Logs**. Check available disk space and memory, including context overhead. |
| llama.cpp fails after custom launch arguments | A fixed-model argument can conflict with PAIR's managed router mode. | Remove custom `-m` or `-hf` arguments and use PAIR's model download and load controls. |
| A loaded llama.cpp model becomes idle and the next request takes longer | Managed llama.cpp puts idle models to sleep after five minutes and wakes them for a new request. | Allow the model to wake; inspect **Jobs** to confirm that the request completes. |
| A browser client reports a CORS error | PAIR 1.0.0 follows the selected engine's origin policy. | Configure the allowed browser origin on the system running that engine. For Ollama, set `OLLAMA_ORIGINS` in the engine's **Settings** launch arguments. |