Русский · Hugging Face · TalasCloud
I optimize local LLM inference and generative media for low-spec PCs. My projects include portable Windows builds and GGUF quantization. I work with NVIDIA Pascal and Tesla P40 over CUDA, AMD Polaris and RX 580 over Vulkan, Intel Ivy Bridge with AVX1, and Haswell with AVX2.
| Project | Use it for |
|---|---|
| Strata Desktop | LLMs on Tesla P40 · local Swift 256K, CUDA and AVX1 |
| pascallama.cpp | llama.cpp DP4A / MMQ patches for Pascal |
| KoboldCpp Core | Haswell AVX2 + Polaris Vulkan runtime |
| Polaris AI Toolkit | RX 580 toolkit for ZLUDA, HIP/ROCm and ComfyUI |
| GGUF Converter | Portable Windows conversion + Vulkan calibration |
| KoboldCpp Media | Local API for images, video, music and sound |
| Project | Use it for |
|---|---|
| Open WebUI AVX1 | Self-hosted interface for older Windows PCs |
| Open WebUI Desktop AVX1 | Desktop client with isolated connections |
| Open Terminal AVX1 | Standalone terminal and agent service |
| Conduit ARM64 | Open WebUI + native Cline on Android |
| Cline Desktop AVX1 | Archived Ivy Bridge Windows client |
- Tesla P40 · Strata. One Swift IQ3_S request with 261,998 input tokens and 128 output tokens completed in 913.43 s without a cache read. Evidence
- RX 580 · Polaris. Synthetic 4096×4096 attention took 87.68 → 53.86 ms and 1,024 → 69 MiB in additional peak tensor allocations. Evidence
- RX 580 / Haswell · Media. LightVAE-Pro saved 17 Wan frames in 30.53 s vs 33.04 s. The SA3 CPU worker processed three chunks of a 15-second effect in 21.57 s vs 26.53 s. Outputs matched byte for byte. Evidence
Calibrated GGUF, ONNX exports and original components. Model cards include upstream revisions, licenses, hashes and compatible runtimes.
Images / Изображения · Video / Видео · Music / Музыка · Sound effects / Звуковые эффекты

