| --- |
| license: other |
| tags: |
| - wheels |
| - cuda |
| - pytorch |
| - windows |
| - linux |
| --- |
| |
| # image-server-wheels |
|
|
| Prebuilt Python 3.11 wheels. |
|
|
| ## Contents |
|
|
| | File | OS | CUDA | Torch | Source | Notes | |
| |---|---|---|---|---|---| |
| | `ace_step-1.6.0-py3-none-any.whl` | any | — | — | built by us | Pure-Python, cross-platform | |
| | `qwen_tts-0.1.1+deapi3-py3-none-any.whl` | any | — | — | built by us | Pure-Python, patched fork (`+deapi3`) | |
| | `image_server_kernels-0.4.2+cu128torch2.8-cp311-cp311-win_amd64.whl` | Windows x64 | 12.8 | 2.8 | built by us | **Current.** Adds Sol-Attn sparse attention, W4A8 linear, and in-tree llama.cpp GGUF CUDA kernels (`_C_gguf`), on top of: FP8 dense + grouped GEMM, NVFP4 W4A4, INT8 W8A8 + ConvRot, fused QK-norm+RoPE. SM89 + SM120 | |
| | `image_server_kernels-0.3.0-cp311-cp311-win_amd64.whl` | Windows x64 | 12.8 | 2.8 | built by us | Previous build, kept as a fallback. No Sol-Attn, no W4A8, no bundled GGUF kernels | |
| | `image_server_kernels-0.3.0+cu129torch2.8-cp311-cp311-linux_x86_64.whl` | Linux x86_64 | 12.9 | 2.8 | built by us | Linux build is still at 0.3.0 — there is no Linux 0.4.2 yet | |
| | `block_sparse_attn-0.0.2-cp311-cp311-win_amd64.whl` | Windows x64 | 12.8 | 2.8 | built by us | Used by video pipeline | |
| | `block_sparse_attn-0.0.2-cp311-cp311-linux_x86_64.whl` | Linux x86_64 | 12.8 | 2.8 | built by us | Used by video pipeline | |
| | `q8_kernels-0.0.5-cp311-cp311-win_amd64.whl` | Windows x64 | 12.8 | 2.8 | built by us | Used by LTX video | |
| | `q8_kernels-0.0.5-cp311-cp311-linux_x86_64.whl` | Linux x86_64 | 12.8 | 2.8 | built by us | Used by LTX video | |
| | `sageattention-2.2.0+cu128torch2.8.0-cp311-cp311-win_amd64.whl` | Windows x64 | 12.8 | 2.8 | [woct0rdho/SageAttention](https://github.com/woct0rdho/SageAttention/releases/tag/v2.2.0-windows) | Mirror of upstream release | |
| | `sageattention-2.2.0+cu129torch2.8-cp311-cp311-linux_x86_64.whl` | Linux x86_64 | 12.9 | 2.8 | provenance unconfirmed — see Credits | | |
| | `flash_attn-2.8.2+cu128torch2.8-cp311-cp311-win_amd64.whl` | Windows x64 | 12.8 | 2.8 | [mjun0812/flash-attention-prebuild-wheels](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.4.10) | Mirror of upstream release | |
| | `flash_attn-2.8.3+cu130torch2.10-cp311-cp311-win_amd64.whl` | Windows x64 | 13.0 | 2.10 | [mjun0812/flash-attention-prebuild-wheels](https://github.com/mjun0812/flash-attention-prebuild-wheels) | Mirror of upstream release | |
| | `flash_attn-2.8.3+cu128torch2.8-cp311-cp311-linux_x86_64.whl` | Linux x86_64 | 12.8 | 2.8 | [mjun0812/flash-attention-prebuild-wheels](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.7.16) | Mirror of upstream release | |
| |
| ## Direct install |
| |
| ```bash |
| BASE=https://huggingface.co/deAPI-ai/image-server-wheels/resolve/main |
| |
| # Windows |
| pip install $BASE/image_server_kernels-0.4.2+cu128torch2.8-cp311-cp311-win_amd64.whl |
| pip install $BASE/q8_kernels-0.0.5-cp311-cp311-win_amd64.whl |
| pip install $BASE/block_sparse_attn-0.0.2-cp311-cp311-win_amd64.whl |
| pip install $BASE/flash_attn-2.8.2+cu128torch2.8-cp311-cp311-win_amd64.whl |
| pip install $BASE/sageattention-2.2.0+cu128torch2.8.0-cp311-cp311-win_amd64.whl |
| pip install --no-deps $BASE/ace_step-1.6.0-py3-none-any.whl |
| pip install --no-deps $BASE/qwen_tts-0.1.1+deapi3-py3-none-any.whl |
|
|
| # Linux |
| pip install $BASE/image_server_kernels-0.3.0+cu129torch2.8-cp311-cp311-linux_x86_64.whl |
| pip install $BASE/q8_kernels-0.0.5-cp311-cp311-linux_x86_64.whl |
| pip install $BASE/block_sparse_attn-0.0.2-cp311-cp311-linux_x86_64.whl |
| pip install $BASE/sageattention-2.2.0+cu129torch2.8-cp311-cp311-linux_x86_64.whl |
| pip install --no-deps $BASE/flash_attn-2.8.3+cu128torch2.8-cp311-cp311-linux_x86_64.whl |
| ``` |
| |
| `sageattention` is required by the MiniMax-H3 video model, which imports it |
| unconditionally. Every other model that can use it falls back to SDPA when it is |
| absent, so on a box that does not serve H3 the wheel is optional. |
| |
| > `image_server_kernels` is at **0.4.2 on Windows but only 0.3.0 on Linux** — the |
| > Linux build of 0.4.2 has not been made yet. Sol-Attn, W4A8 and the bundled GGUF |
| > kernels are therefore Windows-only for now. |
| |
| ## What is verified in `image_server_kernels-0.4.2` |
| |
| Built and tested on Windows 11 / RTX 5090 Laptop (SM120), CUDA 12.8, torch 2.8.0+cu128. |
| |
| - `scripts/smoke_test_rebuild.py` — ALL PASS: `fp8_dense_gemm` (per-tensor cos 0.99965, |
| per-channel+bias 0.99972), `dynamic_per_token_scaled_fp8_quant` 0.99965, |
| `fp8_dense_gemm_pre_quantized` 1.00000, `fp8_dense_gemm_sm120` 1.00000 via native |
| CUTLASS 3.x, `rmsnorm_forward` 0.999999. |
| - `scripts/validate_gguf_kernels.py` against real GGUF tensors — ALL PASS for |
| Q4_K, Q5_K, Q6_K, Q5_0, Q5_1, Q8_0 (cos 0.99984–1.00000), which covers the bf16 |
| K-quant dequant path the in-tree patch exists for. |
| - `sol_attn` on real MiniMax-H3 activations (S=15,479, 5 capture points): |
| **2.22x vs SageAttention2 at cos 0.9805** at `tau=1.0`. Both speed and accuracy |
| rise with sequence length (4k → 1.68x/0.965, 15.5k → 2.90x/0.980 in one run), |
| so longer clips do better than this figure. |
| |
| Not exercised on this box: `sol_attn` on Ada SM89 (compiles and links, but no Ada |
| GPU here), `w4a8_linear`, `nvfp4_*`, `fp8_grouped_gemm`, `w8a16_dense_gemm`. |
| |
| ## Credits |
| |
| `flash_attn` wheels are mirrored from |
| [mjun0812/flash-attention-prebuild-wheels](https://github.com/mjun0812/flash-attention-prebuild-wheels) |
| — all credit for those builds goes to the upstream author. |
| |
| `sageattention` is **a mirror, not our build**. The library is |
| [thu-ml/SageAttention](https://github.com/thu-ml/SageAttention) (Apache-2.0); the |
| Windows wheel is built and published by |
| [woct0rdho](https://github.com/woct0rdho/SageAttention/releases), and we found it |
| through [wildminder/AI-windows-whl](https://github.com/wildminder/AI-windows-whl), |
| which indexes Windows CUDA wheels. All credit for the library and the build goes to |
| them. The Windows file here is byte-identical to the upstream release |
| (`sha256:4379951403809dfcd5b1e10d35e287abf42afafb7b27615c09d6062cfbdf230a`). |
| |
| > The **Linux** `sageattention-2.2.0+cu129torch2.8` wheel's provenance is not |
| > recorded — we have not confirmed whether it is our own build or a mirror. Fill |
| > this in before relying on the attribution. |
| |
| We mirror third-party wheels so the install scripts have a single source of truth and |
| do not break if upstream release URLs change. |
| |
| `ace_step`, `qwen_tts`, `image_server_kernels`, `block_sparse_attn` and `q8_kernels` |
| were built in-house. |
| |
| Sol-Attn in `image_server_kernels` wraps CUDA kernels from comfy-kitchen |
| (Apache-2.0); see `third_party/comfy_kitchen_sol/PROVENANCE.md` in the |
| `image-server-kernels` repo. The bundled GGUF kernels are llama.cpp's `ggml-cuda` |
| (MIT), vendored per `third_party/llamacpp_gguf/PROVENANCE.md`. |
| |
