Built in-house from unmodified Dao-AILab/causal-conv1d v1.7.0 (cd81f041),
compiled from source (CAUSAL_CONV1D_FORCE_BUILD=TRUE), not a renamed
upstream wheel. CPython 3.11, torch 2.8.0+cu129, CUDA 12.9, GCC 13.3.
Debug info stripped. Requires glibc >= 2.32.
Verified against the previous in-house build of the same source: forward,
input grad, update and conv state bit-identical on 108 cases (fp32/fp16/bf16,
widths 2-4, silu/none, with/without bias). Weight/bias grads at the largest
shape are nondeterministic in both builds with the same spread (4.4e-5 rel).
sha256 a882b1c262ec7d9375feeb39d0543871adf08f3af4d8282a221f17421459e3f8
Publish the Linux cu129 image_server_kernels 0.4.3 wheel
- Add image_server_kernels 0.4.2 (Windows) as the current build; keep 0.3.0 listed as a fallback.
- The Linux image_server_kernels and Linux sageattention wheels were already in the repo but absent from the table and from the Linux install block; both added.
- Remove the note claiming there is no Linux build of image_server_kernels or sageattention here - both exist.
- Record what is verified in 0.4.2 and what was not exercised.
- Flag that the Linux sageattention wheel provenance is unrecorded.
Sol-Attn sparse attention, W4A8 linear, in-tree llama.cpp GGUF CUDA kernels (_C_gguf).
Built on Windows 11 / CUDA 12.8 / torch 2.8.0+cu128, MSVC 14.38 toolset.
Verified on RTX 5090 (SM120): smoke_test_rebuild ALL PASS, validate_gguf_kernels ALL PASS on real Q4_K/Q5_K/Q6_K tensors, sol_attn 2.22x vs SageAttention2 at cos 0.9805 on real MiniMax-H3 activations.
sha256 8089f1dacb6ea89634d5f046d6a5e33d266d85b4c4b886199ace5a0b74a16bea
SageAttention 2.2.0 Linux wheel, cu129/torch2.8 (sm_89 + sm_120a)
