dev-python/tokenspeed-mla-bin
TokenSpeed multi-head latent attention CUDA kernels
Screenshots
-
tokenspeed-mla-bin-0.2.8~amd64python_single_target_python3_12 python_single_target_python3_13 python_single_target_python3_14 debug
View 
Download 
Browse License: MIT Overlay: stuff -
tokenspeed-mla-bin-0.2.7~amd64python_single_target_python3_12 python_single_target_python3_13 python_single_target_python3_14
View Download Browse License: MIT Overlay: stuff
ChangeLog
commit 6121863cea3fce11dbb0609bcc0c085f44842ec1
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Sep 12 13:53:59 2026 +0200
dev-python/tokenspeed-mla-bin: bump to 0.2.8
Upstream 0.2.8 keeps the exact apache-tvm-ffi pin and minimum TokenSpeed
Triton version, so its dependency contract is unchanged.
The wheel has extension modules, which are marked and byte-compiled.
`~amd64` remains the only keyword because dependencies are amd64-only.
A full staged install passed for the enabled Python target.
commit 0367bbe9d96ad1e1e1edb67402b9a39f20c6b9f4
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Sep 9 10:34:07 2026 +0200
dev-python/tokenspeed-mla-bin: drop 0.2.5
Last two of the 0.2 line are 0.2.6 and 0.2.7. 0.1.8 stays as the 0.1
anchor -- dev-python/vllm pins it exactly, and it is what holds
tokenspeed-triton-bin 3.7.10_p20260531 in the tree.
commit 06d803c0ba6e3459d301a8f8df93b84d89164203
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Sep 9 09:48:12 2026 +0200
dev-python/tokenspeed-mla-bin: add 0.2.7
Upstream raised the tokenspeed-triton floor to >=3.8.10.post20260906; the
atom follows the version added in the preceding commit. apache-tvm-ffi is
still pinned at exactly 0.1.13.post3, so that pair of bounds is unchanged.
commit 4b49d69d779b8d05dca3bb1651813730b23ac2e0
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 22:02:12 2026 +0200
dev-python/tokenspeed-mla-bin: add 0.2.6
Wheel repackage, so this is not a bare $ bump: SRC_URI hardcodes a
content-addressed pythonhosted path, and the new wheel's hash directory had
to be pulled from the PyPI JSON. Fetched size matches PyPI's declared 755954
bytes exactly.
RDEPEND change, read from the wheel's own METADATA rather than the PyPI
summary: upstream moved its exact tvm-ffi pin from apache-tvm-ffi==0.1.13 to
==0.1.13.post3, so the >=/<= exact-pin pair moves to 0.1.13_p3. That version
is already in the overlay, so the pin resolves. The other three requirements
(nvidia-cutlass-dsl, tokenspeed-triton>=3.8.10.post20260721, torch) are
unchanged.
Adds-only, nothing dropped: every vllm ebuild in the tree exact-pins
~dev-python/tokenspeed-mla-bin-0.1.8, so 0.1.8 is the only version any
consumer can reach and dropping it would break all three. 0.2.5 and 0.2.6
sit unconsumed until a vllm bump re-pins them -- the same chain-driven
retention shape this package has had since it was added.
Build-verified as root; installs the two objs/*.so kernels and the
dist-info. The unresolved-soname QA notice on libcute_dsl_runtime.so is
host-derived, not a missing dep: nvidia-cutlass-dsl is declared in RDEPEND
but is not installed here, and these are sm_100a/sm_103a Blackwell CUDA
blobs on an AMD-only machine. 0.2.5 carries the identical QA_PREBUILT shape.
commit c563f555bfd7a07238f9ba0b7be911fbb9b6be2f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Aug 15 17:53:23 2026 +0200
dev-python/tokenspeed-mla-bin: add 0.2.5
Add the current upstream kernels and follow their exact apache-tvm-ffi pin and
TokenSpeed Triton floor.
commit c1a398c9fd10a70d6708081bfd1631d99624ff0d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Aug 14 21:56:10 2026 +0200
dev-python/tokenspeed-mla-bin: add 0.1.8
vLLM pins TokenSpeed MLA for speculative-decoding attention kernels. Declare the
exact upstream runtime closure and preserve its Blackwell-native wheel payload.
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Sep 12 13:53:59 2026 +0200
dev-python/tokenspeed-mla-bin: bump to 0.2.8
Upstream 0.2.8 keeps the exact apache-tvm-ffi pin and minimum TokenSpeed
Triton version, so its dependency contract is unchanged.
The wheel has extension modules, which are marked and byte-compiled.
`~amd64` remains the only keyword because dependencies are amd64-only.
A full staged install passed for the enabled Python target.
commit 0367bbe9d96ad1e1e1edb67402b9a39f20c6b9f4
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Sep 9 10:34:07 2026 +0200
dev-python/tokenspeed-mla-bin: drop 0.2.5
Last two of the 0.2 line are 0.2.6 and 0.2.7. 0.1.8 stays as the 0.1
anchor -- dev-python/vllm pins it exactly, and it is what holds
tokenspeed-triton-bin 3.7.10_p20260531 in the tree.
commit 06d803c0ba6e3459d301a8f8df93b84d89164203
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Sep 9 09:48:12 2026 +0200
dev-python/tokenspeed-mla-bin: add 0.2.7
Upstream raised the tokenspeed-triton floor to >=3.8.10.post20260906; the
atom follows the version added in the preceding commit. apache-tvm-ffi is
still pinned at exactly 0.1.13.post3, so that pair of bounds is unchanged.
commit 4b49d69d779b8d05dca3bb1651813730b23ac2e0
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 22:02:12 2026 +0200
dev-python/tokenspeed-mla-bin: add 0.2.6
Wheel repackage, so this is not a bare $ bump: SRC_URI hardcodes a
content-addressed pythonhosted path, and the new wheel's hash directory had
to be pulled from the PyPI JSON. Fetched size matches PyPI's declared 755954
bytes exactly.
RDEPEND change, read from the wheel's own METADATA rather than the PyPI
summary: upstream moved its exact tvm-ffi pin from apache-tvm-ffi==0.1.13 to
==0.1.13.post3, so the >=/<= exact-pin pair moves to 0.1.13_p3. That version
is already in the overlay, so the pin resolves. The other three requirements
(nvidia-cutlass-dsl, tokenspeed-triton>=3.8.10.post20260721, torch) are
unchanged.
Adds-only, nothing dropped: every vllm ebuild in the tree exact-pins
~dev-python/tokenspeed-mla-bin-0.1.8, so 0.1.8 is the only version any
consumer can reach and dropping it would break all three. 0.2.5 and 0.2.6
sit unconsumed until a vllm bump re-pins them -- the same chain-driven
retention shape this package has had since it was added.
Build-verified as root; installs the two objs/*.so kernels and the
dist-info. The unresolved-soname QA notice on libcute_dsl_runtime.so is
host-derived, not a missing dep: nvidia-cutlass-dsl is declared in RDEPEND
but is not installed here, and these are sm_100a/sm_103a Blackwell CUDA
blobs on an AMD-only machine. 0.2.5 carries the identical QA_PREBUILT shape.
commit c563f555bfd7a07238f9ba0b7be911fbb9b6be2f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Aug 15 17:53:23 2026 +0200
dev-python/tokenspeed-mla-bin: add 0.2.5
Add the current upstream kernels and follow their exact apache-tvm-ffi pin and
TokenSpeed Triton floor.
commit c1a398c9fd10a70d6708081bfd1631d99624ff0d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Aug 14 21:56:10 2026 +0200
dev-python/tokenspeed-mla-bin: add 0.1.8
vLLM pins TokenSpeed MLA for speculative-decoding attention kernels. Declare the
exact upstream runtime closure and preserve its Blackwell-native wheel payload.