Troubleshooting

Troubleshooting

Start with the Logs page on the TV (or ts5.log in the app's folder): every request and every problem has a line there in plain words.

PS5 title

• The title doesn't appear on the home screen. ShadowMountPlus must be running and scanning the folder (/data/homebrew). If the console says it can't start the app (CE-107750-0), set the folder and its files to permissions 777 in your FTP client; apps the Prospero Store installs are set for you.
• Models disappeared after an update. Start the title once: the first start of a new version brings the models back through the jailbreak's FTP server (port 2121). Without ftpsrv, copy them in again from a computer or download them again. The store keeps the previous folder in /data/prosperostore/previous/PPSA99555 until the next update.
• A download is slow. The PS5 lets a title write at about 1 to 4 MB/s after the first few hundred MB. Copy models in over FTP instead: Models.
• A download stopped. Cross on it continues from where it left off. A dropped connection is retried on its own; any other reason is on the page and in Logs.
• The model won't start. Square on it (the GPU check) says why, and Logs has the details. A model that doesn't fit the memory with at least some context is refused; pick a smaller quant.
• R1 talk does nothing. A speech model must be downloaded (Settings > Speech model). The mic is open only while R1 is held.
• No spoken replies. Download Kokoro and a voice in Settings, and check that spoken replies are on and aimed at the TV or the controller.

Linux

• "no Vulkan device" / ts5 finds no GPU on a PS5. Mesa is too old for the PS5's GPU (it needs 25.3 or newer). Debian 12's Mesa doesn't know the chip; Debian 13 with trixie-backports, Fedora, Ubuntu 26.04 and Bazzite work. vulkaninfo --summary should show "RADV GFX1013".
• A 14B model doesn't fit. tokenstation-memory status, then sudo tokenstation-memory set max and reboot (re-run the loader). See PS5 hardware on Linux.
• Port 8081 is in use. The TV service and the API-only service can't both run; the desktop app also serves the API itself. Disable the one you don't want (sudo systemctl disable --now tokenstation).
• The hardware settings do nothing. Your user must be in the tokenstation group (sudo usermod -aG tokenstation $USER, log in again), and the boost needs ps5_control from ps5-linux-tools on the PATH or in PS5_CONTROL.
• The boost keeps coming back on. Disable ps5-linux-tools' ps5boost.service.
• The console won't boot after a compute-unit change. Open cmdline.txt on the boot drive's FAT32 partition from another computer and set amdgpu.ps5_cu_unlock=0. Keep the file on one line.
• The GPU check fails after a compute-unit change. Go back to 36. That chip has a faulty unit among the extra ones.
• It's slower than the numbers here. Check the boost is off, the GPU memory isn't swapping (tokenstation-memory status), and nothing else is using the GPU (close Steam or the desktop for the TV service). ts5 bench -m model.gguf gives the raw engine speed.

Apps and the API

• The app can't connect. It must use http://, not https://, with /v1 at the end, and the PS5's current address. A request that never shows in Logs never reached the PS5.
• 401 invalid_api_key. Require a key is on and the app sent none, or sent one that's off. Make a new key in Settings > Apps.
• "maximum context length". The conversation is longer than the model's context. Tell the app the context_length from /v1/models, or start a new task.
• Away from home nothing connects. The PS5 needs to be on your tailnet and an API key is always required over Tailscale. Check Settings shows the device, and your tailnet's access rules allow port 8081.
• Web search never happens. The key must start with tvly- and the console's clock must be right (HTTPS fails otherwise). Only chats that look like they need the web get the tool; "web_search": true in the request forces it.