`rosa-torch-native`

Optional native companion for rosa-torch, accelerating the exact CPU suffix-automaton and Link-Cut Tree inference step. The distribution installs the importable rosa_native_step extension module; it does not replace the main Python package.

The C++ core implements the validated exact production state machine. It binds the NumPy arrays of a _StatefulInferenceState once, updates them in place, and releases the GIL during computation. It neither includes nor calls libtorch. The runtime dependencies rosa-torch>=0.3,<0.4 and NumPy provide the compatible tensor and array contracts. Install the rosa-torch[numba] extra to enable the native suffix-automaton and rich-candidate integration; the native RLBWT backends do not require Numba.

The constructor validates every shape, dtype, counter, and ABI version before retaining any pointer. The current native state ABI is 1.

The module also provides fused online RLBWT inference for exact dense top-1 retrieval. NativeRLBWTState is the explicit baseline, NativeRLBWTCompactState stores vocabularies of at most 256 IDs in adaptive packed leaves, and NativeRLBWTStateMC exposes the separately named opt-in Monte-Carlo variants. The Python package keeps these backends explicit; backend="auto" continues to select the production suffix-automaton path.

The module also exposes NativeCandidateState for the exact rich state in rosa._stateful_candidates_numba. Its batched step maintains the same K suffixes, R newest occurrences, unbounded frequencies, and newest-prefix + delta LCT tags as the Python/Numba implementation. The R capacity remains owned by the NumPy arrays of the Python CandidateState, whose lifetime is retained by the native object. Capability can be detected through the presence of NativeCandidateState and candidate_abi_version == 1.

The rich ABI retains step, global reset, and position, while detecting the optional prefill, step_masked, reset_masked, positions, step_into, and prefill_into capabilities. prefill emits all five native arrays at every position in one C++ call and leaves the state ready for continuation. The caller-owned *_into methods avoid repeated output allocation after validating dtype, shape, contiguity, writability, and memory overlap.

Ragged mode stores one position per row. Uniform and ragged paths are intentionally incompatible so that exactly one position authority exists at any time. The Python wrapper falls back exactly to Numba when an older ABI-1 wheel does not provide a newer optional method.

Installation and usage

Install the matching wheel for the current platform and Python ABI from PyPI:

uv add 'rosa-torch[numba]' rosa-torch-native

rosa-torch detects the extension automatically from its Numba inference backend. The low-level API remains available for diagnostics:

from rosa_native_step import NativeState

NativeState(state).step(tokens_numpy) expects a contiguous NumPy vector of shape [batch_size] that is convertible to int64. The object retains a reference to the Python state and exposes its read-only position.

NativeCandidateState(candidate_state).step(tokens_numpy) requires a C-contiguous int64 NumPy vector and returns the low-level tuple (source, match_length, state_id, frequency, count). reset() recycles the whole batch in time proportional to the state nodes and hash slots that are actually occupied.

Each native state lazily creates a small dependency-free persistent C++17 thread pool. It parallelizes independent batch rows for top-1 and rich prefill. Prefill keeps a serial path below batch 4, where waking a worker costs more than the measured computation. Uniform steps use the pool only from batch 64 onward to preserve the inexpensive serial path for small batches.

The pool is bounded by the batch size, reported CPU count, and 16 total threads. ROSA_NATIVE_THREADS=1 forces the serial path; another positive integer sets a lower cap. An unset variable selects the available limit automatically, while an invalid value falls back conservatively to one thread. Only non-empty digit strings representing a strictly positive integer are accepted.

Using one pool per state avoids fragile extension-singleton shutdown ordering, and workers never call Python. No worker is created for a ragged state or a batch below 4 that never uses prefill. A sufficiently large first call may create at most 15 workers, which are reused until state destruction. Concurrent mutating calls on the same instance are serialized by a mutex acquired while the GIL is released. Public output allocation and array ownership remain unchanged. Worker exceptions are captured and rethrown on the calling thread.

Isolated local build

The PEP 517 backend is setuptools, with setuptools, wheel, and pybind11 as build dependencies. Pybind11Extension selects C++17, while setuptools provides the platform extension flags for macOS and Linux. Both platforms additionally use -O3 and NDEBUG.

From the repository root:

uv sync --extra numba
uv build --python .venv/bin/python --wheel native \
  --out-dir /tmp/rosa-native-dist
uv run --isolated \
  --with '.[numba]' \
  --with /tmp/rosa-native-dist/rosa_torch_native-0.3.0-*.whl \
  native/tests/smoke.py

The smoke test forces Numba as the oracle, lets the regular ROSA path load the companion, and compares predictions step by step.

Run the rich smoke test and direct Numba comparison with native/tests/candidate_smoke.py and native/benchmark_candidates.py, respectively, in the same isolated environment.

Multi-platform publication

The release workflow builds CPython 3.10–3.14 wheels for Linux x86-64, macOS arm64, and Windows x64, plus CPython 3.10–3.12 wheels for macOS x86-64. Each wheel is smoke-tested by importing the extension and checking its production and compact RLBWT ABI versions before it is attached to the GitHub Release and published through PyPI Trusted Publishing.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

rosa_torch_native-0.4.0.tar.gz (49.7 kB view details)

Uploaded Aug 19, 2026 Source

Built Distributions

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

rosa_torch_native-0.4.0-cp314-cp314-win_amd64.whl (410.4 kB view details)

Uploaded Aug 19, 2026 CPython 3.14Windows x86-64

rosa_torch_native-0.4.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl (266.2 kB view details)

Uploaded Aug 19, 2026 CPython 3.14manylinux: glibc 2.27+ x86-64manylinux: glibc 2.28+ x86-64

rosa_torch_native-0.4.0-cp314-cp314-macosx_11_0_arm64.whl (231.7 kB view details)

Uploaded Aug 19, 2026 CPython 3.14macOS 11.0+ ARM64

rosa_torch_native-0.4.0-cp313-cp313-win_amd64.whl (397.8 kB view details)

Uploaded Aug 19, 2026 CPython 3.13Windows x86-64

rosa_torch_native-0.4.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl (266.1 kB view details)

Uploaded Aug 19, 2026 CPython 3.13manylinux: glibc 2.27+ x86-64manylinux: glibc 2.28+ x86-64

rosa_torch_native-0.4.0-cp313-cp313-macosx_11_0_arm64.whl (230.6 kB view details)

Uploaded Aug 19, 2026 CPython 3.13macOS 11.0+ ARM64

rosa_torch_native-0.4.0-cp312-cp312-win_amd64.whl (397.8 kB view details)

Uploaded Aug 19, 2026 CPython 3.12Windows x86-64

rosa_torch_native-0.4.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl (266.1 kB view details)

Uploaded Aug 19, 2026 CPython 3.12manylinux: glibc 2.27+ x86-64manylinux: glibc 2.28+ x86-64

rosa_torch_native-0.4.0-cp312-cp312-macosx_11_0_arm64.whl (230.7 kB view details)

Uploaded Aug 19, 2026 CPython 3.12macOS 11.0+ ARM64

rosa_torch_native-0.4.0-cp312-cp312-macosx_10_13_x86_64.whl (244.9 kB view details)

Uploaded Aug 19, 2026 CPython 3.12macOS 10.13+ x86-64

rosa_torch_native-0.4.0-cp311-cp311-win_amd64.whl (393.2 kB view details)

Uploaded Aug 19, 2026 CPython 3.11Windows x86-64

rosa_torch_native-0.4.0-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl (263.5 kB view details)

Uploaded Aug 19, 2026 CPython 3.11manylinux: glibc 2.27+ x86-64manylinux: glibc 2.28+ x86-64

rosa_torch_native-0.4.0-cp311-cp311-macosx_11_0_arm64.whl (229.7 kB view details)

Uploaded Aug 19, 2026 CPython 3.11macOS 11.0+ ARM64

rosa_torch_native-0.4.0-cp311-cp311-macosx_10_9_x86_64.whl (244.2 kB view details)

Uploaded Aug 19, 2026 CPython 3.11macOS 10.9+ x86-64

rosa_torch_native-0.4.0-cp310-cp310-win_amd64.whl (392.8 kB view details)

Uploaded Aug 19, 2026 CPython 3.10Windows x86-64

rosa_torch_native-0.4.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl (262.7 kB view details)

Uploaded Aug 19, 2026 CPython 3.10manylinux: glibc 2.27+ x86-64manylinux: glibc 2.28+ x86-64

rosa_torch_native-0.4.0-cp310-cp310-macosx_11_0_arm64.whl (228.5 kB view details)

Uploaded Aug 19, 2026 CPython 3.10macOS 11.0+ ARM64

rosa_torch_native-0.4.0-cp310-cp310-macosx_10_9_x86_64.whl (243.1 kB view details)

Uploaded Aug 19, 2026 CPython 3.10macOS 10.9+ x86-64

File details

Details for the file rosa_torch_native-0.4.0.tar.gz.

File metadata

Download URL: rosa_torch_native-0.4.0.tar.gz
Upload date: Aug 19, 2026
Size: 49.7 kB
Tags: Source
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for rosa_torch_native-0.4.0.tar.gz
Algorithm	Hash digest
SHA256	`3000b8dadbe8a58c50c9f807ebd01472e1ab0c59fae5be35f851927893f4e527`
MD5	`ca8df7c6d74b35ee9c2a91ef72c8a679`
BLAKE2b-256	`b15ff4c9da713581045d47714415f66d709fb5691b4471d942909db260ec005c`

Algorithm	Hash digest
SHA256	`caf02a3a4f8c7559d32a5b8100754bf7119e63ddeeabbd9195128e4bd7923c31`
MD5	`63641cad2c98b380b2fcb9b09e010a5e`
BLAKE2b-256	`a7a724701a29016b8cca40163b622d772df061ff9d328d6c1d54e0bf68e30190`

Algorithm	Hash digest
SHA256	`bb06cd488ecae2526e4c29ac6e61508cb094f1107b573ede090bb4a39c4b64ce`
MD5	`1612efe64e1fabe02f13799074963b1f`
BLAKE2b-256	`bfd3aedd7f1a2e48acd6f3776784449308e330eef7c4d5c4ca31e9b1ec26f732`

Algorithm	Hash digest
SHA256	`9e300e5797501f8c458d6705d5a711547ffeba52725d77bc025757cfc2a0caee`
MD5	`40888f1e99677c8aaae31f643f9e98cb`
BLAKE2b-256	`4e27227f0aa769caa8b3c584ea694ea88ced854db590e5aa514a711023ebf2c2`

Algorithm	Hash digest
SHA256	`5c0cfc59c10f72c0a284cc4d2d086f9a1a224503c8e5b9d40df70087039f6e6a`
MD5	`d5c150a6faaef3d768b9ba29987b91fe`
BLAKE2b-256	`27c1d8258f7dc3cb7775dbe82292da17cccaef85aed2fc2681eca0c6e678d675`

Algorithm	Hash digest
SHA256	`cb71e600fca599b946c4218dfb23e7f071c8c195f29f55fca08015e881a0639e`
MD5	`90cc719251ac4bc19c7443d043b961dc`
BLAKE2b-256	`d682ffa8a587ee7df892b2ca2cde0d3cea975a87271aa83d626150f0a72f5e6b`

Algorithm	Hash digest
SHA256	`21ee956f2aa20e8bda15bc0172f6b1e908d95025c89274535edc9708450fb40c`
MD5	`b2e20783d06e280b10649212cfa87922`
BLAKE2b-256	`1f0b40953ad96c4e2cef569715ff76210af38b585bebc8643a7aca137397879b`

Algorithm	Hash digest
SHA256	`b97f6032a3e3bf7daee520932d58bf7451f8029377a906c00cb808e552dae1b9`
MD5	`dfabb1999262f1e12263d86507f4dafc`
BLAKE2b-256	`0fc1bc4f6230c63420fd185ff24fea91243709d77852533800b53b972f58b2c0`

Algorithm	Hash digest
SHA256	`88c0f12ffd6d6f95cd121d49518134eba585cece06d723b8517bbfa3b934e407`
MD5	`7614801426db40e6730614552d01c370`
BLAKE2b-256	`fcf395faf3d869aab1ce34f3a8553074408b03de1a8f5dcae06ccf7d3d89e664`

Algorithm	Hash digest
SHA256	`2209a5893eb45c6620bbb14ce2340d6042fc9440533d32e242cc239af8c41fa3`
MD5	`e91f8e7d3a90a2ec61c167e826835393`
BLAKE2b-256	`e335829182de14511a057d8950a32e863dc2d4dbf006d7eb4d4981f435921b23`

Algorithm	Hash digest
SHA256	`c8deb60c6800d0e3b8a7f6d928628cdcc55db1c9641e84689e6d27187014048f`
MD5	`69ae4aa1f1ba93cc11f7f441700f5286`
BLAKE2b-256	`b549a3e63f09eaeb0b23ff0bb9b85d7949b87d85868d134969efdcec92d7d872`

Algorithm	Hash digest
SHA256	`599347c029cfef30ec0237f17af455bd85968b3246a7bed78b8d119a5bc55350`
MD5	`96e6c74f5ebd75ed22ab3cdda6624f2c`
BLAKE2b-256	`9c1aba8d290eb511521329c4de876b67f07e894970b52f0e441c0c99cfb13d36`

Algorithm	Hash digest
SHA256	`84a65b7ca986c9333c15165c79b12fa3dd27eec46f115a0fb956dd307686824d`
MD5	`95ee0c8a12c63bf2714819ccea7af4de`
BLAKE2b-256	`3b489c7d30cb52ee635e8564ba53fa57073147338492871376626fcc0ab625ba`

Algorithm	Hash digest
SHA256	`4a226f2815c99979239c422521751a676fc7f321a5ffc6fc1e3a3c02f8da0a67`
MD5	`2842fb7118740aea3cf368ea9b0e2995`
BLAKE2b-256	`5a0f14aedf733da3439ae1d2698e389476b96daac1fb29d00540e9346eddfcc4`

Algorithm	Hash digest
SHA256	`c902da35f7c1a7ef9087dd3ed56f23f2eb7b101e1161f1114f92a104eca5fb6c`
MD5	`8b335a56a6ccdf0e2f2ebd41eb7de709`
BLAKE2b-256	`3023fdd0fed8d09781de84f476d6899402e332aa0b5a0a6d84994d4adec344d5`

Algorithm	Hash digest
SHA256	`7e6f5353957f41613092bfdfe679bb902c25c07721108768d6c243ea2f7d5b07`
MD5	`2e391c016d4c7cd53c8fd0240da1b415`
BLAKE2b-256	`e03b7d38f08467a60e1183984365f08d21a257100b1945f9cdc4f26922047863`

Algorithm	Hash digest
SHA256	`eefee42bfb7d3c817003ea8eb1969b664ae6d7d00f08bda74de612e7d99a16b8`
MD5	`5efe86be8a14e407468c005e062fd41e`
BLAKE2b-256	`59afa2a4bbaafeefeb9857965ca7fdeab900f7d5249a6238dbb01c03aac89481`

Algorithm	Hash digest
SHA256	`1bee9c912359a19675bd311240a469f604591f2172cd769c1c751a836fba6f2c`
MD5	`82caac03ec6959d5d85b10452bf134c1`
BLAKE2b-256	`855c96f3f020a6f3ac9b21e216a519ea195c04cc4e1bf15e196f99db31c44201`

Algorithm	Hash digest
SHA256	`ec239c03b6b2fa44d00e0bff56586b7100b437758a58ddb0361d9614c26e3e3f`
MD5	`783b9d711ddc8983f25cc31edcf99228`
BLAKE2b-256	`19e69fb26434b6fafc6003d1f7a9aa16d57918a4a2b4b0c707381d27b9e025ac`

rosa-torch-native 0.4.0

rosa-torch-native

Installation and usage

Isolated local build

Multi-platform publication

Project links

Key dates

1 maintainer

GitHub Statistics

License expression

Requires

Classifiers

Download files

Source Distribution

Built Distributions

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

`rosa-torch-native`