llama-cpp-python
PyTorch
artha-ai-studio
windows-wheels
cuda
sageattention
flash-attention
ffmpeg
gemma-4
rtx-5090
blackwell
output = llm(
"Once upon a time,",
max_tokens=512,
echo=True
)
print(output)Artha AI Studio — Models & Pre-Built Windows Wheels
This repository is the official hub for Artha AI Studio, hosting curated model weights, runtime assets, and pre-compiled Windows CUDA/C++ Python wheels (.whl) built for the portable Artha AI Studio environment.
Repository Structure
ruygar/artha-ai-studio/
├── python/ # Standalone Windows x64 Python 3.13.14 NuGet package (with include/ & libs/)
├── ffmpeg/ # Portable Windows x64 FFmpeg 8.1.3 LTS static runtime & GPLv3 provenance
├── wheels/ # Pre-built Windows x64 CUDA Python wheels (.whl)
│ ├── nvidia/ # NVIDIA CUDA 13.2 runtime headers & libraries
│ ├── torch/ # PyTorch CUDA 13.2 wheels for Python 3.13
│ ├── sage/ # SageAttention CUDA wheels (Ampere, Ada, Hopper, Blackwell)
│ ├── flash/ # FlashAttention-2 CUDA wheels (Ampere, Ada, Blackwell)
│ └── llama/ # llama-cpp-python CUDA wheels (Ampere, Ada, Blackwell)
└── models/ # Studio-specific model weights and checkpoints
└── llm/
└── gemma/ # Gemma-4 INT8 ConvRot quantized weights (.safetensors)
Pre-Built Runtime & Windows Wheels
All packages in this repository are prepared for Windows x64 using the portable Artha AI Studio build toolchain (Python 3.13.14 + MSVC 14.44 + CUDA 13.2).
1. Python Runtime & Development Headers (python/)
| Package File | Version | Platform | Contents | Direct Link |
|---|---|---|---|---|
python-3.13.14-win_amd64.nupkg |
3.13.14 |
win_amd64 |
Full standalone Python 3.13.14 (tools/python.exe, Lib/, DLLs/, include/, libs/) |
Download |
Direct Download (curl):
curl -L -o python_nuget.zip "https://huggingface.co/ruygar/artha-ai-studio/resolve/main/python/python-3.13.14-win_amd64.nupkg"
2. NVIDIA CUDA Runtime (wheels/nvidia/)
| Wheel File | Version | CUDA | Python ABI | Direct Link |
|---|---|---|---|---|
nvidia_cuda_runtime-13.2.86-py3-none-win_amd64.whl |
13.2.86 |
13.2 |
py3-none (Python 3.x) |
Download |
Direct Installation (pip):
pip install "https://huggingface.co/ruygar/artha-ai-studio/resolve/main/wheels/nvidia/nvidia_cuda_runtime-13.2.86-py3-none-win_amd64.whl"
3. PyTorch (wheels/torch/)
| Wheel File | Version | CUDA | Python ABI | Direct Link |
|---|---|---|---|---|
torch-2.13.0+cu132-cp313-cp313-win_amd64.whl |
2.13.0 |
13.2 |
cp313 (Python 3.13) |
Download |
torchvision-0.28.0+cu132-cp313-cp313-win_amd64.whl |
0.28.0 |
13.2 |
cp313 (Python 3.13) |
Download |
torchaudio-2.11.1+cu132-py3-none-win_amd64.whl |
2.11.1 |
13.2 |
py3-none (Python 3.x) |
Download |
Direct Installation (pip):
pip install ^
"https://huggingface.co/ruygar/artha-ai-studio/resolve/main/wheels/torch/torch-2.13.0%2Bcu132-cp313-cp313-win_amd64.whl" ^
"https://huggingface.co/ruygar/artha-ai-studio/resolve/main/wheels/torch/torchvision-0.28.0%2Bcu132-cp313-cp313-win_amd64.whl" ^
"https://huggingface.co/ruygar/artha-ai-studio/resolve/main/wheels/torch/torchaudio-2.11.1%2Bcu132-py3-none-win_amd64.whl"
4. SageAttention (wheels/sage/)
| Wheel File | Version | CUDA | PyTorch | Python ABI | Supported GPU Architectures | Direct Link |
|---|---|---|---|---|---|---|
sageattention-2.2.0+cu132torch2.13.0-cp310-abi3-win_amd64.whl |
2.2.0 |
13.2 |
2.13.0+ |
cp310-abi3 (Python 3.10–3.13+) |
sm_80, sm_86, sm_89, sm_90a, sm_120a (RTX 30xx / 40xx / 50xx) |
Download |
Key Features of This Build:
- Python Stable ABI (
cp310-abi3) & LibTorch Stable ABI (TORCH_STABLE_ONLY): A single wheel works across Python 3.10, 3.11, 3.12, and 3.13+. - Full RTX 5090 (
sm_120) Support Across Both FP8 & FP16 Kernels: Includes nativesm_89andsm_120a+ PTX binaries in both_qattn_sm89.pyd(sageattn_qk_int8_pv_fp8_cuda) and_qattn_sm80.pyd(sageattn_qk_int8_pv_fp16_cuda).
Direct Installation (pip):
pip install "https://huggingface.co/ruygar/artha-ai-studio/resolve/main/wheels/sage/sageattention-2.2.0%2Bcu132torch2.13.0-cp310-abi3-win_amd64.whl"
(Note: Requires triton-windows on Windows: pip install triton-windows)
5. FlashAttention-2 (wheels/flash/)
| Wheel File | Version | CUDA | PyTorch | Python ABI | Supported GPU Architectures | Direct Link |
|---|---|---|---|---|---|---|
flash_attn-2.8.4+cu132torch2.13.0-cp313-cp313-win_amd64.whl |
2.8.4 |
13.2 |
2.13.0+ |
cp313 (Python 3.13) |
sm_80 (Ampere/Ada) + sm_120f / compute_120 (Blackwell RTX 50xx) |
Download |
Direct Installation (pip):
pip install "https://huggingface.co/ruygar/artha-ai-studio/resolve/main/wheels/flash/flash_attn-2.8.4%2Bcu132torch2.13.0-cp313-cp313-win_amd64.whl"
(Note: Requires einops: pip install einops)
6. llama-cpp-python (wheels/llama/)
| Wheel File | Version | CUDA | Python ABI | Supported GPU Architectures | Direct Link |
|---|---|---|---|---|---|
llama_cpp_python-0.3.36+cu132-py3-none-win_amd64.whl |
0.3.36 |
13.2 |
py3-none (Python 3.8–3.14+) |
sm_80, sm_86, sm_89, sm_120a (Blackwell native FP4 + FlashAttention q4_0/q8_0/f16/bf16) |
Download |
Direct Installation (pip):
pip install "https://huggingface.co/ruygar/artha-ai-studio/resolve/main/wheels/llama/llama_cpp_python-0.3.36%2Bcu132-py3-none-win_amd64.whl"
7. FFmpeg 8.1.3 LTS Static Runtime (ffmpeg/)
| Archive File | Version | Platform | License | Contents | Direct Link |
|---|---|---|---|---|---|
ffmpeg-n8.1.3-34-ga35157abb9-win64-gpl-8.1.zip |
8.1.3 (a35157abb9) |
win64 |
GPLv3+ / MIT (build scripts) |
Static ffmpeg.exe, ffprobe.exe, ffplay.exe, doc/, presets/, LICENSE.txt (GPLv3), LICENSE-BtbN-MIT.txt, SOURCE_PROVENANCE.txt |
Download |
License & Corresponding Source Provenance (GPLv3 Section 6):
- Binary License: GNU General Public License v3.0 or later (
COPYING.GPLv3, included asLICENSE.txtinside the archive). - Build Scripts License: MIT License (
Copyright 2020-2021 BtbN <btbn@btbn.de>, included asLICENSE-BtbN-MIT.txtinside the archive). - Upstream FFmpeg Source Revision:
FFmpeg/FFmpeg@a35157abb9(n8.1.3-34-ga35157abb9, release branch8.1) - Upstream Build Recipe & Dependency Pins:
BtbN/FFmpeg-Builds@c63bb308a1(release tagautobuild-2026-11-07-12-47)
Direct Download & Extraction (curl + tar):
curl -L -o ffmpeg.zip "https://huggingface.co/ruygar/artha-ai-studio/resolve/main/ffmpeg/ffmpeg-n8.1.3-34-ga35157abb9-win64-gpl-8.1.zip"
tar -xf ffmpeg.zip --strip-components=1 -C bin\ffmpeg
8. Gemma-4 INT8 ConvRot Models (models/llm/gemma/)
| Model | File | Quantized 2D Layers (I8) | Preserved 16-bit Layers (BF16) | Size (BF16 → INT8 ConvRot) | Direct Link |
|---|---|---|---|---|---|
| Gemma-4 E2B | gemma-4-E2B-it-int8-convrot.safetensors |
316 |
251 (vision_tower, audio_tower, embeddings) |
9.54 GB → 7.78 GB (-18.5%) |
Download |
| Gemma-4 E4B | gemma-4-E4B-it-int8-convrot.safetensors |
379 |
251 (vision_tower, audio_tower, embeddings) |
14.89 GB → 11.15 GB (-25.2%) |
Download |
| Gemma-4 12B | gemma-4-12B-it-int8-convrot.safetensors |
328 |
4 (vision_embedder, embed_vision, embed_audio, embed_tokens) |
22.28 GB → 12.13 GB (-45.5%) |
Download |
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="ruygar/artha-ai-studio", filename="{{GGUF_FILE}}", )