Executable programs and extensions
Last updated
Strata troubleshooting: downloads, first launch, and slow responses
The error message and the stage where setup stopped point to the next check.
Strata prepares the engine and model at startup, loads the model into RAM1 and GPU2 memory, and then opens the local server. Resume an interrupted download with the same update file, and keep the progress window open if the PC slows during the first load. Low-RAM resident and mmap modes supply experts differently; PLE3 is separate data read from storage. Version 0.1.43 also fixes parallel NTFS reads for Windows low-RAM GGUF-mmap setups.
At startup, Strata's launcher prepares the engine and model, loads the model into memory, and opens the local server. When an error occurs, note the last completed stage and the first error in the window or log. The next check depends on whether setup stopped during download, GPU initialization, model loading, or browser connection.
Record the OS, GPU model, driver, and Strata version. AMD on Windows requires an Adrenalin driver, and validation of model execution on individual discrete Radeons with the current HIP4 ZIP remains limited. Consult the installation guide and AMD HIP documentation for your OS to find support status and diagnostic commands.
Distinguish an interrupted download from the first model load
If the model download stops during first setup, rerun START-HERE.bat on Windows or ./setup.sh in the same Strata folder on Linux. Setup resumes with files already downloaded. For an update to an existing branch-based Git installation, rerun UPDATE.bat or ./update.sh, then use the launcher to start the model. A fixed-tag installation needs a new release ZIP. If downloads keep failing, check free disk space and the network connection.
During the first model load, Strata places model experts in RAM and prepares space for the GPU. The official installation guide says the PC may slow down or appear unresponsive for 1–3 minutes on the first run. Keep the window open while progress information changes. Distinguish a brief unchanged screen from repeated error messages in the log.
If there is no progress for more than 10 minutes and errors continue, restart the PC and close memory-heavy apps such as browsers, games, or editors. Select a smaller model or shorter context in setup and try again. If the error appears during GPU initialization, record the driver version and the end of the Strata log.
For slow responses, distinguish resident, mmap, and PLE
In low-RAM setups, resident mode copies experts outside GPU memory into RAM at startup and pins them there. If the remaining experts fit, their weights are not read from the SSD5 during responses. When RAM cannot hold the resident set, setup may use mmap instead. Mmap maps the expert file and supplies data through the OS file cache, so accessing uncached pages involves storage reads.
In Monitor, watch free RAM and disk reads together. If many experts outside the GPU are being read while RAM is short, test a smaller quantized model, a shorter context, and fewer concurrently running apps, one at a time. Low-RAM mode changes how experts are placed, and an mmap configuration that frequently reads experts from the SSD can be slow with limited VRAM6.
PLE (per-layer embeddings) is an embedding table used by individual model layers. It is read from the SSD, so file access can continue in resident mode. Record available RAM and disk-read timing in Monitor and logs, then check whether the reads supply experts or PLE data.
Check Windows GGUF-mmap reads on 0.1.43
In low-memory Windows setups, keeping GGUF7 shards mmaped while using unbuffered NTFS reads could serialize expert-file reads. Strata 0.1.43 closes each shard's mapping view when the file tier starts unbuffered reads, restoring parallel reads. If GGUF-mmap responses were unusually slow, confirm the log shows the 0.1.43 engine and compare the same files, model, and prompt8.
Scroll sideways to compare all columns.
| Experts in RAM | Story decode (tok/s) | Code decode (tok/s) | Short-prompt decode (tok/s) |
|---|---|---|---|
| 22 GiB | 18.9 → 25.5 tok/s | 13.5 → 21.1 tok/s | 13.3 → 18.2 tok/s |
| 12 GiB | 9.6 → 16.0 tok/s | 6.5 → 11.6 tok/s | 4.6 → 8.5 tok/s |
Test conditions: RTX 5070, Ryzen 5 7600, 64 GB RAM, Windows, IQ3_XXS GGUF. Results are medians from five interleaved old/new pairs with either 22 GiB or 12 GiB of experts in RAM.
Check GPU detection and port 8080
If Strata cannot find the GPU, first check whether the operating system sees the card. On Windows AMD, update the driver in AMD Software: Adrenalin Edition and restart. On Linux AMD, check the amdgpu driver and HIP device-detection output in the setup log. For NVIDIA, run nvidia-smi in a terminal and check that the GPU and driver appear.
If you see port 8080 is already in use, check whether Strata is already running in another window. If it is, open the existing server at http://127.0.0.1:8080. If another app owns the port, use the documented option to select another port, such as --port 8081, and update the connected app to use the same port.
If the issue continues, record the OS, GPU and driver, Strata version, failed stage, and final log lines, then compare them with the official troubleshooting guide. Turning off Windows protections or ignoring security warnings can expose the PC without identifying the cause. Start with documented support, memory, driver, and port checks.
Terminology notes
System RAM — System memory that temporarily holds data while programs run.
Back to the textGPU — A processor designed to handle many calculations in parallel. It performs model computations during AI inference.
Back to the textPLE embedding table — Short for Per-Layer Embedding, a structure that adds embedding information at particular model layers. The Qwen Flash-Next family uses n-gram tables; supported runtimes can offload these tables to RAM or SSD.
Back to the textHIP — An API and runtime for portable GPU C++ code. It helps with porting CUDA code.
Back to the textSSD — A data storage device that uses flash memory and retains data without power.
Back to the textVRAM — Memory used by a graphics card’s GPU to store model weights and intermediate values.
Back to the textGGUF — A file format for model data, widely used by llama.cpp-based tools.
Back to the textPrompt — The input text sent to a model. For generative models it describes the desired result and strongly shapes the output.
Back to the text
Frequently asked questions
- What should you check first in this article?
- Strata prepares the engine and model at startup, loads the model into RAM and GPU memory, and then opens the local server. Resume an interrupted download with the same update file, and keep the progress window open if the PC slows during the first load. Low-RAM resident and mmap modes supply experts differently; PLE is separate data read from storage. Version 0.1.43 also fixes parallel NTFS reads for Windows low-RAM GGUF-mmap setups.
- How much memory does it need?
- In low-memory Windows setups, keeping GGUF shards mmaped while using unbuffered NTFS reads could serialize expert-file reads. Strata 0.1.43 closes each shard's mapping view when the file tier starts unbuffered reads, restoring parallel reads. If GGUF-mmap responses were unusually slow, confirm the log shows the 0.1.43 engine and compare the same files, model, and prompt.