helix-lm 0.1.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,188 @@
1
+ # Initially taken from Github's Python gitignore file
2
+
3
+ # Byte-compiled / optimized / DLL files
4
+ __pycache__/
5
+ *.py[cod]
6
+ *$py.class
7
+
8
+ # C extensions
9
+ *.so
10
+
11
+ # tests and logs
12
+ tests/fixtures/cached_*_text.txt
13
+ logs/
14
+ lightning_logs/
15
+ lang_code_data/
16
+ reports/
17
+
18
+ # Distribution / packaging
19
+ .Python
20
+ build/
21
+ develop-eggs/
22
+ dist/
23
+ downloads/
24
+ eggs/
25
+ .eggs/
26
+ lib/
27
+ lib64/
28
+ parts/
29
+ sdist/
30
+ var/
31
+ wheels/
32
+ *.egg-info/
33
+ .installed.cfg
34
+ *.egg
35
+ MANIFEST
36
+
37
+ # PyInstaller
38
+ # Usually these files are written by a python script from a template
39
+ # before PyInstaller builds the exe, so as to inject date/other infos into it.
40
+ *.manifest
41
+ *.spec
42
+
43
+ # Installer logs
44
+ pip-log.txt
45
+ pip-delete-this-directory.txt
46
+
47
+ # Unit test / coverage reports
48
+ htmlcov/
49
+ .tox/
50
+ .nox/
51
+ .coverage
52
+ .coverage.*
53
+ .cache
54
+ nosetests.xml
55
+ coverage.xml
56
+ *.cover
57
+ .hypothesis/
58
+ .pytest_cache/
59
+
60
+ # Translations
61
+ *.mo
62
+ *.pot
63
+
64
+ # Django stuff:
65
+ *.log
66
+ local_settings.py
67
+ db.sqlite3
68
+
69
+ # Flask stuff:
70
+ instance/
71
+ .webassets-cache
72
+
73
+ # Scrapy stuff:
74
+ .scrapy
75
+
76
+ # Sphinx documentation
77
+ docs/_build/
78
+
79
+ # PyBuilder
80
+ target/
81
+
82
+ # Jupyter Notebook
83
+ .ipynb_checkpoints
84
+
85
+ # IPython
86
+ profile_default/
87
+ ipython_config.py
88
+
89
+ # pyenv
90
+ .python-version
91
+
92
+ # celery beat schedule file
93
+ celerybeat-schedule
94
+
95
+ # SageMath parsed files
96
+ *.sage.py
97
+
98
+ # Environments
99
+ .env
100
+ .venv
101
+ .venv*
102
+ env/
103
+ venv/
104
+ ENV/
105
+ env.bak/
106
+ venv.bak/
107
+
108
+ # Spyder project settings
109
+ .spyderproject
110
+ .spyproject
111
+
112
+ # Rope project settings
113
+ .ropeproject
114
+
115
+ # mkdocs documentation
116
+ /site
117
+
118
+ # mypy
119
+ .mypy_cache/
120
+ .dmypy.json
121
+ dmypy.json
122
+
123
+ # Pyre type checker
124
+ .pyre/
125
+
126
+ # vscode
127
+ .vs
128
+ .vscode
129
+
130
+ # Pycharm
131
+ .idea
132
+
133
+ # TF code
134
+ tensorflow_code
135
+
136
+ # Models
137
+ proc_data
138
+
139
+ # examples
140
+ runs
141
+ /runs_old
142
+ /wandb
143
+ /examples/runs
144
+ /examples/**/*.args
145
+ /examples/rag/sweep
146
+
147
+ # data
148
+ /data
149
+ serialization_dir
150
+
151
+ # emacs
152
+ *.*~
153
+ debug.env
154
+
155
+ # vim
156
+ .*.swp
157
+
158
+ #ctags
159
+ tags
160
+
161
+ # pre-commit
162
+ .pre-commit*
163
+
164
+ # .lock
165
+ *.lock
166
+
167
+ # DS_Store (MacOS)
168
+ .DS_Store
169
+
170
+ # ruff
171
+ .ruff_cache
172
+
173
+ # checkers cache
174
+ utils/.checkers_cache.json
175
+ utils/.checkers_requirements.json
176
+ utils/.noisy_comments_cache.json
177
+ mlinter-findings.json
178
+
179
+ # modular conversion
180
+ *.modular_backup
181
+
182
+ # Cursor IDE files
183
+ .cursor/
184
+ test-results/
185
+
186
+ # AI agent local setup artifacts
187
+ /.agents/skills
188
+ /.claude/skills
helix_lm-0.1.0/LICENSE ADDED
@@ -0,0 +1,203 @@
1
+ Copyright 2018- The Hugging Face team. All rights reserved.
2
+
3
+ Apache License
4
+ Version 2.0, January 2004
5
+ http://www.apache.org/licenses/
6
+
7
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
8
+
9
+ 1. Definitions.
10
+
11
+ "License" shall mean the terms and conditions for use, reproduction,
12
+ and distribution as defined by Sections 1 through 9 of this document.
13
+
14
+ "Licensor" shall mean the copyright owner or entity authorized by
15
+ the copyright owner that is granting the License.
16
+
17
+ "Legal Entity" shall mean the union of the acting entity and all
18
+ other entities that control, are controlled by, or are under common
19
+ control with that entity. For the purposes of this definition,
20
+ "control" means (i) the power, direct or indirect, to cause the
21
+ direction or management of such entity, whether by contract or
22
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
23
+ outstanding shares, or (iii) beneficial ownership of such entity.
24
+
25
+ "You" (or "Your") shall mean an individual or Legal Entity
26
+ exercising permissions granted by this License.
27
+
28
+ "Source" form shall mean the preferred form for making modifications,
29
+ including but not limited to software source code, documentation
30
+ source, and configuration files.
31
+
32
+ "Object" form shall mean any form resulting from mechanical
33
+ transformation or translation of a Source form, including but
34
+ not limited to compiled object code, generated documentation,
35
+ and conversions to other media types.
36
+
37
+ "Work" shall mean the work of authorship, whether in Source or
38
+ Object form, made available under the License, as indicated by a
39
+ copyright notice that is included in or attached to the work
40
+ (an example is provided in the Appendix below).
41
+
42
+ "Derivative Works" shall mean any work, whether in Source or Object
43
+ form, that is based on (or derived from) the Work and for which the
44
+ editorial revisions, annotations, elaborations, or other modifications
45
+ represent, as a whole, an original work of authorship. For the purposes
46
+ of this License, Derivative Works shall not include works that remain
47
+ separable from, or merely link (or bind by name) to the interfaces of,
48
+ the Work and Derivative Works thereof.
49
+
50
+ "Contribution" shall mean any work of authorship, including
51
+ the original version of the Work and any modifications or additions
52
+ to that Work or Derivative Works thereof, that is intentionally
53
+ submitted to Licensor for inclusion in the Work by the copyright owner
54
+ or by an individual or Legal Entity authorized to submit on behalf of
55
+ the copyright owner. For the purposes of this definition, "submitted"
56
+ means any form of electronic, verbal, or written communication sent
57
+ to the Licensor or its representatives, including but not limited to
58
+ communication on electronic mailing lists, source code control systems,
59
+ and issue tracking systems that are managed by, or on behalf of, the
60
+ Licensor for the purpose of discussing and improving the Work, but
61
+ excluding communication that is conspicuously marked or otherwise
62
+ designated in writing by the copyright owner as "Not a Contribution."
63
+
64
+ "Contributor" shall mean Licensor and any individual or Legal Entity
65
+ on behalf of whom a Contribution has been received by Licensor and
66
+ subsequently incorporated within the Work.
67
+
68
+ 2. Grant of Copyright License. Subject to the terms and conditions of
69
+ this License, each Contributor hereby grants to You a perpetual,
70
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
71
+ copyright license to reproduce, prepare Derivative Works of,
72
+ publicly display, publicly perform, sublicense, and distribute the
73
+ Work and such Derivative Works in Source or Object form.
74
+
75
+ 3. Grant of Patent License. Subject to the terms and conditions of
76
+ this License, each Contributor hereby grants to You a perpetual,
77
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
78
+ (except as stated in this section) patent license to make, have made,
79
+ use, offer to sell, sell, import, and otherwise transfer the Work,
80
+ where such license applies only to those patent claims licensable
81
+ by such Contributor that are necessarily infringed by their
82
+ Contribution(s) alone or by combination of their Contribution(s)
83
+ with the Work to which such Contribution(s) was submitted. If You
84
+ institute patent litigation against any entity (including a
85
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
86
+ or a Contribution incorporated within the Work constitutes direct
87
+ or contributory patent infringement, then any patent licenses
88
+ granted to You under this License for that Work shall terminate
89
+ as of the date such litigation is filed.
90
+
91
+ 4. Redistribution. You may reproduce and distribute copies of the
92
+ Work or Derivative Works thereof in any medium, with or without
93
+ modifications, and in Source or Object form, provided that You
94
+ meet the following conditions:
95
+
96
+ (a) You must give any other recipients of the Work or
97
+ Derivative Works a copy of this License; and
98
+
99
+ (b) You must cause any modified files to carry prominent notices
100
+ stating that You changed the files; and
101
+
102
+ (c) You must retain, in the Source form of any Derivative Works
103
+ that You distribute, all copyright, patent, trademark, and
104
+ attribution notices from the Source form of the Work,
105
+ excluding those notices that do not pertain to any part of
106
+ the Derivative Works; and
107
+
108
+ (d) If the Work includes a "NOTICE" text file as part of its
109
+ distribution, then any Derivative Works that You distribute must
110
+ include a readable copy of the attribution notices contained
111
+ within such NOTICE file, excluding those notices that do not
112
+ pertain to any part of the Derivative Works, in at least one
113
+ of the following places: within a NOTICE text file distributed
114
+ as part of the Derivative Works; within the Source form or
115
+ documentation, if provided along with the Derivative Works; or,
116
+ within a display generated by the Derivative Works, if and
117
+ wherever such third-party notices normally appear. The contents
118
+ of the NOTICE file are for informational purposes only and
119
+ do not modify the License. You may add Your own attribution
120
+ notices within Derivative Works that You distribute, alongside
121
+ or as an addendum to the NOTICE text from the Work, provided
122
+ that such additional attribution notices cannot be construed
123
+ as modifying the License.
124
+
125
+ You may add Your own copyright statement to Your modifications and
126
+ may provide additional or different license terms and conditions
127
+ for use, reproduction, or distribution of Your modifications, or
128
+ for any such Derivative Works as a whole, provided Your use,
129
+ reproduction, and distribution of the Work otherwise complies with
130
+ the conditions stated in this License.
131
+
132
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
133
+ any Contribution intentionally submitted for inclusion in the Work
134
+ by You to the Licensor shall be under the terms and conditions of
135
+ this License, without any additional terms or conditions.
136
+ Notwithstanding the above, nothing herein shall supersede or modify
137
+ the terms of any separate license agreement you may have executed
138
+ with Licensor regarding such Contributions.
139
+
140
+ 6. Trademarks. This License does not grant permission to use the trade
141
+ names, trademarks, service marks, or product names of the Licensor,
142
+ except as required for reasonable and customary use in describing the
143
+ origin of the Work and reproducing the content of the NOTICE file.
144
+
145
+ 7. Disclaimer of Warranty. Unless required by applicable law or
146
+ agreed to in writing, Licensor provides the Work (and each
147
+ Contributor provides its Contributions) on an "AS IS" BASIS,
148
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
149
+ implied, including, without limitation, any warranties or conditions
150
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
151
+ PARTICULAR PURPOSE. You are solely responsible for determining the
152
+ appropriateness of using or redistributing the Work and assume any
153
+ risks associated with Your exercise of permissions under this License.
154
+
155
+ 8. Limitation of Liability. In no event and under no legal theory,
156
+ whether in tort (including negligence), contract, or otherwise,
157
+ unless required by applicable law (such as deliberate and grossly
158
+ negligent acts) or agreed to in writing, shall any Contributor be
159
+ liable to You for damages, including any direct, indirect, special,
160
+ incidental, or consequential damages of any character arising as a
161
+ result of this License or out of the use or inability to use the
162
+ Work (including but not limited to damages for loss of goodwill,
163
+ work stoppage, computer failure or malfunction, or any and all
164
+ other commercial damages or losses), even if such Contributor
165
+ has been advised of the possibility of such damages.
166
+
167
+ 9. Accepting Warranty or Additional Liability. While redistributing
168
+ the Work or Derivative Works thereof, You may choose to offer,
169
+ and charge a fee for, acceptance of support, warranty, indemnity,
170
+ or other liability obligations and/or rights consistent with this
171
+ License. However, in accepting such obligations, You may act only
172
+ on Your own behalf and on Your sole responsibility, not on behalf
173
+ of any other Contributor, and only if You agree to indemnify,
174
+ defend, and hold each Contributor harmless for any liability
175
+ incurred by, or claims asserted against, such Contributor by reason
176
+ of your accepting any such warranty or additional liability.
177
+
178
+ END OF TERMS AND CONDITIONS
179
+
180
+ APPENDIX: How to apply the Apache License to your work.
181
+
182
+ To apply the Apache License to your work, attach the following
183
+ boilerplate notice, with the fields enclosed by brackets "[]"
184
+ replaced with your own identifying information. (Don't include
185
+ the brackets!) The text should be enclosed in the appropriate
186
+ comment syntax for the file format. We also recommend that a
187
+ file or class name and description of purpose be included on the
188
+ same "printed page" as the copyright notice for easier
189
+ identification within third-party archives.
190
+
191
+ Copyright [yyyy] [name of copyright owner]
192
+
193
+ Licensed under the Apache License, Version 2.0 (the "License");
194
+ you may not use this file except in compliance with the License.
195
+ You may obtain a copy of the License at
196
+
197
+ http://www.apache.org/licenses/LICENSE-2.0
198
+
199
+ Unless required by applicable law or agreed to in writing, software
200
+ distributed under the License is distributed on an "AS IS" BASIS,
201
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
202
+ See the License for the specific language governing permissions and
203
+ limitations under the License.
helix_lm-0.1.0/NOTICE ADDED
@@ -0,0 +1,23 @@
1
+ helix-lm
2
+ Copyright 2026 Nathan
3
+
4
+ Licensed under the Apache License, Version 2.0 (the "License"); you may not use
5
+ these files except in compliance with the License. A copy of the License is in
6
+ the LICENSE file.
7
+
8
+ --------------------------------------------------------------------------------
9
+
10
+ This product includes software developed at Hugging Face, Inc.
11
+
12
+ Portions of `helix_lm/model.py` are derived from HuggingFace Transformers
13
+ (https://github.com/huggingface/transformers), which is licensed under the
14
+ Apache License, Version 2.0. In particular, the chunkwise-parallel and
15
+ single-step gated delta-rule routines (`torch_chunk_gated_delta_rule`,
16
+ `torch_recurrent_gated_delta_rule`), the depthwise causal convolution helper
17
+ (`causal_conv1d_fn`), the rotary-embedding helpers (`rotate_half`,
18
+ `apply_rotary_pos_emb`), `repeat_kv`, and the RMS normalization layers follow
19
+ that project's reference implementations.
20
+
21
+ The HELIX architecture itself -- the three-strand braid, the hierarchical
22
+ landmark index and its causal beam descent, the multi-scale local staircase,
23
+ surprise-gated writes, and the strand fusion -- is original to this project.
@@ -0,0 +1,189 @@
1
+ Metadata-Version: 2.5
2
+ Name: helix-lm
3
+ Version: 0.1.0
4
+ Summary: HELIX: a linear-scaling sequence architecture that keeps exact long-range recall
5
+ Project-URL: Homepage, https://github.com/Wisk-inc/transformers/tree/main/helix-lm
6
+ Project-URL: Source, https://github.com/Wisk-inc/transformers/tree/main/helix-lm
7
+ Project-URL: Issues, https://github.com/Wisk-inc/transformers/issues
8
+ Author: Nathan
9
+ Maintainer: Nathan
10
+ License-Expression: Apache-2.0
11
+ License-File: LICENSE
12
+ License-File: NOTICE
13
+ Keywords: associative-recall,attention,delta-rule,linear-attention,long-context,pytorch,sparse-attention,state-space-model,transformer
14
+ Classifier: Development Status :: 3 - Alpha
15
+ Classifier: Intended Audience :: Developers
16
+ Classifier: Intended Audience :: Science/Research
17
+ Classifier: License :: OSI Approved :: Apache Software License
18
+ Classifier: Operating System :: OS Independent
19
+ Classifier: Programming Language :: Python :: 3
20
+ Classifier: Programming Language :: Python :: 3.10
21
+ Classifier: Programming Language :: Python :: 3.11
22
+ Classifier: Programming Language :: Python :: 3.12
23
+ Classifier: Programming Language :: Python :: 3.13
24
+ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
25
+ Classifier: Typing :: Typed
26
+ Requires-Python: >=3.10
27
+ Requires-Dist: torch>=2.5
28
+ Provides-Extra: dev
29
+ Requires-Dist: pytest>=7; extra == 'dev'
30
+ Requires-Dist: ruff>=0.6; extra == 'dev'
31
+ Description-Content-Type: text/markdown
32
+
33
+ # HELIX
34
+
35
+ **H**ierarchical **E**pisodic **L**inear **I**nde**X** — a sequence architecture that keeps what attention
36
+ is good at while dropping the quadratic bill.
37
+
38
+ ```bash
39
+ pip install helix-lm
40
+ ```
41
+
42
+ ```python
43
+ import torch
44
+ from helix_lm import HelixConfig, HelixForCausalLM
45
+
46
+ model = HelixForCausalLM(HelixConfig(vocab_size=32000, hidden_size=1024, num_hidden_layers=12))
47
+ logits = model(torch.randint(0, 32000, (1, 8192))).logits
48
+ text = model.generate(torch.randint(0, 32000, (1, 64)), max_new_tokens=32, temperature=0.8)
49
+ ```
50
+
51
+ ```bash
52
+ helix info # what it is and what it costs
53
+ helix demo # build a model, generate, verify the decode state
54
+ helix bench # cost against context length
55
+ helix recall # train with and without the index strand, and compare
56
+ ```
57
+
58
+ ---
59
+
60
+ ## The idea
61
+
62
+ A transformer compares every token to every other token. Ten thousand tokens is fifty million
63
+ comparisons; double the input and you quadruple the work. Linear-time alternatives fix that by squeezing
64
+ the past into a fixed-size state — and then they cannot quote you a name they saw once, because a state of
65
+ `S` bits cannot tell apart more than `2^S` histories. That is a counting argument, not an engineering gap.
66
+
67
+ HELIX does not try to compress its way around it. It **separates memory capacity from memory bandwidth**.
68
+ Every block braids three mixers over the same residual stream:
69
+
70
+ | strand | reads | training cost | held at decode |
71
+ | --- | --- | --- | --- |
72
+ | **L** local | the last `local_blocks × block_size` tokens, exactly, with a different window per head | `O(N · span)` | a window-sized ring buffer |
73
+ | **R** recurrent | everything, compressed into a matrix-valued delta-rule state | `O(N · d_k · d_v)` | one matrix + two conv states per head |
74
+ | **I** index | `index_topk` blocks chosen anywhere in the past by a beam descent over a landmark tree | `O(N · topk · block_size)` + `O(N log N)` routing | the key/value history and a landmark tree, of which a step reads `O(topk · block_size + branching · log N)` |
75
+
76
+ Storage is `O(N)`, append-only, never rewritten, and can live in cold memory. What a single token *touches*
77
+ is bounded. So the claim is **`O(1)` hot state and `O(log N)` probes** — not `O(1)` total memory, which is
78
+ impossible for anything that can quote an exact name from a million tokens back.
79
+
80
+ ## How the index works
81
+
82
+ The past is cut into blocks of `block_size` tokens. Each closed block gets a **landmark** — a learned pooled
83
+ summary that mixes a plain mean with an attention pool, because a mean alone erases the rare token a later
84
+ query will be hunting for. Landmarks are pooled again, `index_branching` at a time, into a tree, using one
85
+ weight-shared pooler at every internal level.
86
+
87
+ Two rules keep it causal *and* parallel:
88
+
89
+ 1. **A node is eligible for query block `J` only if the whole span it summarizes ends at or before `J`.** A
90
+ summary is never consulted by a query it partly describes, so a landmark is computed once and reused by
91
+ every query.
92
+ 2. **Routing for block `J` comes from the hidden state at the end of block `J − 1`.** Strictly past for
93
+ every token of block `J`, so one gather serves `block_size` queries. Anything routing could not
94
+ anticipate inside that block is covered exactly by strand L.
95
+
96
+ The descent keeps `index_beam_width` nodes per level. Candidates at each level are the beam's children
97
+ **plus** that level's *frontier* — eligible nodes whose parent is not eligible. At most `branching − 1` per
98
+ level, and the frontiers across all levels tile `[0, J)` exactly, so nothing reachable is cut off by a beam
99
+ that descended elsewhere. Selection scores are fed back as additive attention logits, accumulated along the
100
+ descent path, which is what lets gradient reach every level of the tree rather than just the leaves.
101
+
102
+ ## What has been measured
103
+
104
+ On small models, 4 CPU cores. Reproduce with `helix bench` and `helix recall`.
105
+
106
+ **Cost against context length.** Attention pairs are exact — the query/key products the forward pass
107
+ actually forms, counted by instrumenting the attention call.
108
+
109
+ | tokens | HELIX pairs/token | full attention pairs/token | HELIX | full attention |
110
+ | ---: | ---: | ---: | ---: | ---: |
111
+ | 4 096 | 8 192 | 32 776 | 0.4 s | 0.2 s |
112
+ | 8 192 | 8 192 | 65 544 | 0.9 s | 0.6 s |
113
+ | 16 384 | 8 192 | 131 080 | 1.7 s | 1.9 s |
114
+ | 32 768 | 8 192 | 262 152 | 4.3 s | 6.7 s |
115
+
116
+ HELIX fits `N^1.00` in attention pairs; full attention fits `N^2.00`. By 32k tokens HELIX forms **32×
117
+ fewer** and runs **1.6× faster** — as unfused reference PyTorch against a fused SDPA kernel.
118
+
119
+ **Recall from far back.** Multi-query associative recall: bindings written near the start, queried ~200
120
+ tokens later, well outside the local window.
121
+
122
+ | model | recall | 95% CI |
123
+ | --- | ---: | :---: |
124
+ | HELIX (L+R+I) | **30.8%** | [28.5%, 33.1%] |
125
+ | HELIX no index (L+R) | 6.0% | [4.8%, 7.2%] |
126
+ | full attention | 32.0% | [29.7%, 34.3%] |
127
+ | *chance* | *6.2%* | |
128
+
129
+ Read the two gaps separately. **HELIX against its own ablation is the result**: strip the index strand and
130
+ the same model, same width, same depth, sits at chance — its window cannot see the bindings and its
131
+ recurrent state cannot hold them. That gap is ~21 standard errors. **HELIX against full attention is a
132
+ tie**, not a win or a loss: 1.2 points with a 1.7-point standard error, z = 0.72.
133
+
134
+ ## Correctness
135
+
136
+ `pytest` covers the properties you can check without training anything:
137
+
138
+ - **strictly causal** — editing token `t` leaves every earlier logit bitwise unchanged;
139
+ - **decode == prefill** — token-by-token generation from the fixed-size state reproduces the one-shot
140
+ forward, as does unaligned chunked prefill;
141
+ - **right padding is exactly invariant**;
142
+ - **constant work per token** — each query reads the same number of keys at any context length;
143
+ - **the cache is what the architecture claims** — fixed-size recurrent and convolution states everywhere,
144
+ a window-capped key/value cache on non-indexing layers;
145
+ - **tiling is bitwise inert** — `attention_tile_blocks` caps peak activation memory and changes nothing;
146
+ - **every parameter gets gradient**, landmark poolers included.
147
+
148
+ ## What this is not
149
+
150
+ **There are no trained checkpoints.** Nobody has trained HELIX on real text. Everything above is about
151
+ complexity, causality and recall mechanics — properties provable on an untrained model or a synthetic
152
+ probe — not about perplexity or language quality. The inductive-bias arguments (multi-scale windows, short
153
+ convolutions, surprise-gated writes) are untested hypotheses.
154
+
155
+ Treat this as an architecture proposal with a working, tested reference implementation, not a better model.
156
+ Finding out whether it is actually better needs real training runs on real hardware.
157
+
158
+ Known design limits, stated plainly:
159
+
160
+ - One retrieval serves a whole query block, so `index_topk` must cover the diversity a `block_size` span
161
+ asks for.
162
+ - Routing is one block stale — the index cannot react to the token currently being generated.
163
+ - Landmarks are pooled from keys that already carry their rotary phase, so routing is not purely
164
+ content-addressed.
165
+ - Left padding shifts the block grid, so it is not bit-identical to an unpadded run. Right padding is.
166
+ - No fused kernel for the gathered attention yet; the scaling is right, the constant factor is not.
167
+
168
+ ## Configuration
169
+
170
+ Every knob is on `HelixConfig`, documented in its docstring. The ones that matter most:
171
+
172
+ ```python
173
+ HelixConfig(
174
+ block_size=64, # memory-block granularity
175
+ local_blocks=4, # strand L sees 4 previous blocks plus its own
176
+ num_window_scales=4, # head groups get geometrically different windows
177
+ index_topk=8, # blocks strand I retrieves per query block
178
+ index_layer_stride=3, # ...on every third layer; the rest keep an O(1) KV cache
179
+ index_branching=8, # landmark tree fan-out
180
+ attention_tile_blocks=64, # caps peak activation memory; no effect on the result
181
+ )
182
+ ```
183
+
184
+ ## License
185
+
186
+ Apache-2.0. Portions of the delta-rule and rotary helpers derive from
187
+ [HuggingFace Transformers](https://github.com/huggingface/transformers) (Apache-2.0); see `NOTICE`.
188
+
189
+ **Made by Nathan.**