dev-python/quack-kernels
QuACK attention kernels for vLLM/FlashInfer (cute-DSL based)
ChangeLog
commit 0e21a503926f003a6b2646e7b57279b3c7ae3e65
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri May 29 12:04:16 2026 +0200
dev-python/quack-kernels: add 0.5.0
Track upstream's raised nvidia-cutlass-dsl floor (>=4.5.2).
commit f0feab8c9c0c499aeb19fb644d8b3d5d0cf5318f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed May 13 14:35:28 2026 +0200
dev-python/quack-kernels: disable py3.11
commit 2136ab74575dae3399cde0adab1e4e5302491330
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun May 10 15:24:45 2026 +0200
dev-python/quack-kernels: switch to DISTUTILS_SINGLE_IMPL
sci-ml/pytorch and the now-single-impl dev-python/torch-c-dlpack-ext
need $; the multi-impl helpers (cutlass-dsl,
apache-tvm-ffi, einops) move into python_gen_cond_dep.
commit 9b3f2a353ad27a8206075225b72edd97910c0390
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun May 10 16:51:11 2026 +0200
dev-python/quack-kernels: switch to DISTUTILS_SINGLE_IMPL
sci-ml/pytorch and the now-single-impl dev-python/torch-c-dlpack-ext
need $; the multi-impl helpers (cutlass-dsl,
apache-tvm-ffi, einops) move into python_gen_cond_dep.
(Hand-port of the same conversion landing on 0.3.3 to our newer
0.4.1 ebuild.)
commit c4956a77759b043d3ae94a8b62afcbb4ed658d0c
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun May 10 14:52:57 2026 +0200
dev-python/quack-kernels: add 0.4.1
commit 50dc7eccdb67c47628ae42212108f2e6678b22d1
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu May 7 18:28:27 2026 +0200
dev-python/quack-kernels: new package, 0.3.3
Tier 4 of the vllm CUDA target packaging cycle. QuACK attention
kernels — pure-Python wrappers that JIT-compile attention/elementwise
kernels through nvidia-cutlass-dsl + apache-tvm-ffi at first use.
All deps are already in this overlay's tracked stack. Upstream's
cu13 extra is folded into the unconditional nvidia-cutlass-dsl
dep — our nvidia-cutlass-dsl always pulls the cu13 path on this
amd64+CUDA-13 overlay, so no USE flag is needed.
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri May 29 12:04:16 2026 +0200
dev-python/quack-kernels: add 0.5.0
Track upstream's raised nvidia-cutlass-dsl floor (>=4.5.2).
commit f0feab8c9c0c499aeb19fb644d8b3d5d0cf5318f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed May 13 14:35:28 2026 +0200
dev-python/quack-kernels: disable py3.11
commit 2136ab74575dae3399cde0adab1e4e5302491330
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun May 10 15:24:45 2026 +0200
dev-python/quack-kernels: switch to DISTUTILS_SINGLE_IMPL
sci-ml/pytorch and the now-single-impl dev-python/torch-c-dlpack-ext
need $; the multi-impl helpers (cutlass-dsl,
apache-tvm-ffi, einops) move into python_gen_cond_dep.
commit 9b3f2a353ad27a8206075225b72edd97910c0390
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun May 10 16:51:11 2026 +0200
dev-python/quack-kernels: switch to DISTUTILS_SINGLE_IMPL
sci-ml/pytorch and the now-single-impl dev-python/torch-c-dlpack-ext
need $; the multi-impl helpers (cutlass-dsl,
apache-tvm-ffi, einops) move into python_gen_cond_dep.
(Hand-port of the same conversion landing on 0.3.3 to our newer
0.4.1 ebuild.)
commit c4956a77759b043d3ae94a8b62afcbb4ed658d0c
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun May 10 14:52:57 2026 +0200
dev-python/quack-kernels: add 0.4.1
commit 50dc7eccdb67c47628ae42212108f2e6678b22d1
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu May 7 18:28:27 2026 +0200
dev-python/quack-kernels: new package, 0.3.3
Tier 4 of the vllm CUDA target packaging cycle. QuACK attention
kernels — pure-Python wrappers that JIT-compile attention/elementwise
kernels through nvidia-cutlass-dsl + apache-tvm-ffi at first use.
All deps are already in this overlay's tracked stack. Upstream's
cu13 extra is folded into the unconditional nvidia-cutlass-dsl
dep — our nvidia-cutlass-dsl always pulls the cu13 path on this
amd64+CUDA-13 overlay, so no USE flag is needed.


View
Download
Browse