torchnative 0.0.1a0__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 +1,21 @@
1
+ MIT License
2
+
3
+ Copyright (c) 2025 thisisthepy
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
@@ -0,0 +1,293 @@
1
+ Metadata-Version: 2.4
2
+ Name: torchnative
3
+ Version: 0.0.1a0
4
+ Summary: Run the real PyTorch ecosystem on device -- not a reimplementation of it
5
+ Author: thisisthepy
6
+ License-Expression: MIT
7
+ Project-URL: Homepage, https://github.com/thisisthepy/torchnative
8
+ Project-URL: Documentation, https://github.com/thisisthepy/torchnative/tree/develop/docs
9
+ Project-URL: Source, https://github.com/thisisthepy/torchnative
10
+ Keywords: pytorch,on-device,edge,federated-learning,test-time-adaptation
11
+ Classifier: Development Status :: 2 - Pre-Alpha
12
+ Classifier: Intended Audience :: Developers
13
+ Classifier: Programming Language :: Python :: 3.13
14
+ Classifier: Programming Language :: Rust
15
+ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
16
+ Classifier: Operating System :: Android
17
+ Classifier: Operating System :: iOS
18
+ Classifier: Operating System :: MacOS
19
+ Classifier: Operating System :: POSIX :: Linux
20
+ Requires-Python: >=3.13
21
+ Description-Content-Type: text/markdown
22
+ License-File: LICENSE
23
+ Provides-Extra: federated
24
+ Provides-Extra: test
25
+ Requires-Dist: torch<2.14,>=2.13; extra == "test"
26
+ Dynamic: license-file
27
+
28
+ <div align="center">
29
+
30
+ # torchnative
31
+
32
+ **Run the real PyTorch ecosystem on device — not a reimplementation of it.**
33
+
34
+ [![PyPI](https://img.shields.io/pypi/v/torchnative?color=blue)](https://pypi.org/project/torchnative/)
35
+ [![Python](https://img.shields.io/badge/python-3.13%2B-blue)](https://www.python.org/)
36
+ [![License](https://img.shields.io/badge/license-MIT-green)](LICENSE)
37
+ [![Platforms](https://img.shields.io/badge/platforms-macOS%20%7C%20Linux%20%7C%20Android%20%7C%20iOS-lightgrey)](#install)
38
+ [![Status](https://img.shields.io/badge/status-pre--alpha-orange)](#status)
39
+
40
+ </div>
41
+
42
+ ---
43
+
44
+ `torchnative` replaces PyTorch's compiled core — `torch._C` — with a native extension, so the
45
+ genuine `torch` and `transformers` packages run on a phone the way they run on a workstation.
46
+
47
+ Models are not ported, converted, or re-expressed. They are imported.
48
+
49
+ ```python
50
+ from transformers import AutoModelForCausalLM # the real one
51
+ model = AutoModelForCausalLM.from_pretrained("...")
52
+ model.generate(...) # on the device
53
+ ```
54
+
55
+ > [!WARNING]
56
+ > **Pre-alpha.** The operator layer matches upstream PyTorch numerically and 15 of 20 tested
57
+ > architectures reach zero missing operators — but `import transformers` does not work yet, no
58
+ > checkpoint has ever been loaded, and no device has run the built artefact.
59
+ > See [Status](#status) before depending on this.
60
+
61
+ ---
62
+
63
+ ## Why not a reimplementation
64
+
65
+ Every other route to on-device inference re-expresses the model somewhere else.
66
+
67
+ | | approach | cost |
68
+ |---|---|---|
69
+ | llama.cpp | architectures rewritten in C++ | each new architecture is a porting task |
70
+ | ExecuTorch · CoreML | ahead-of-time compiled graph | export step, and what runs is not what you wrote |
71
+ | MLC | lowered to its own runtime | same |
72
+ | **torchnative** | **the real Python package** | **the substrate is hard; architectures are free** |
73
+
74
+ The reason nobody runs the real thing is that `torch._C` cannot be built for mobile. PyTorch's own
75
+ build sets `INTERN_BUILD_MOBILE` for any Android or iOS toolchain, and that path forces
76
+ `BUILD_PYTHON` off — so the mobile build is structurally incapable of producing the Python
77
+ extension module the Python package needs.
78
+
79
+ `torchnative` supplies that module instead. Everything above it is upstream source, unmodified.
80
+
81
+ ---
82
+
83
+ ## What it does
84
+
85
+ ### 1 · LLM inference
86
+
87
+ Run `transformers` models directly. No conversion step, no per-architecture port — if
88
+ `transformers` supports it and the operators are covered, it runs.
89
+
90
+ ```python
91
+ import torch
92
+ from transformers import AutoModelForCausalLM, AutoTokenizer
93
+
94
+ model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B")
95
+ tok = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")
96
+
97
+ out = model.generate(**tok("On-device inference is", return_tensors="pt"),
98
+ max_new_tokens=32, do_sample=True)
99
+ ```
100
+
101
+ **Today:** the operator layer under this is complete for 15 of 20 architectures, and Llama and
102
+ GPT-2 match upstream **token for token and logit for logit** in both greedy and sampling mode.
103
+ The `from_pretrained` path itself is still blocked — see [Status](#status).
104
+
105
+ ### 2 · Federated learning
106
+
107
+ Devices train locally and share updates, not data. Federated averaging *is* collective
108
+ communication, so this is built on `torch.distributed` rather than beside it — broadcast the
109
+ model, gather the updates, weighted all-reduce.
110
+
111
+ ```python
112
+ from torchnative.nn import federated
113
+
114
+ engine = federated.Engine(model, rounds=..., aggregator=federated.FedAvg())
115
+ engine.participate() # local epochs, then contribute a delta
116
+ ```
117
+
118
+ **Today:** planned. `torch.distributed` is being implemented from `world_size = 1` upward, which
119
+ is a truthful description of a single device rather than a stub.
120
+
121
+ ### 3 · Test-time adaptation & training
122
+
123
+ A model that ships to a device meets data the training set never had. TTA, TTT and the wider
124
+ test-time learning family let it adapt in place — and every method reduces to the same thing: a
125
+ weight delta over base weights, differing only in lifetime and destination.
126
+
127
+ ```python
128
+ from torchnative import adapt
129
+
130
+ model = adapt.wrap(model, method=adapt.Tent()) # or TTT, memory-based, entropy-based
131
+ model.online() # adapt as it serves
132
+ ```
133
+
134
+ **Today:** planned; the delta abstraction is specified in [`docs/DESIGN.md`](docs/DESIGN.md) §3.
135
+ Lifetime is driven by system events — backgrounding, user switch, sync window — rather than by
136
+ the domain boundaries a benchmark hands you.
137
+
138
+ ---
139
+
140
+ ## How it works
141
+
142
+ ```
143
+ your code · transformers · torch/*.py upstream Python, unmodified
144
+ ──────────────────────────────────────────
145
+ torch._C ← replaced
146
+ ├── _aten_dispatch the single door every operator passes
147
+ ├── Python spellings torch.mm, x.softmax(), F.linear, ...
148
+ └── kernels Rust, backed by candle
149
+ ──────────────────────────────────────────
150
+ CPU today · Metal, Vulkan, NPU planned
151
+ ```
152
+
153
+ **One door.** Every operator reaches its kernel through `_aten_dispatch`, and nothing bypasses
154
+ it. That makes the surface measurable — an unimplemented operator names itself rather than
155
+ failing downstream — and it gives graph capture, which NPU backends will need, exactly one place
156
+ to attach.
157
+
158
+ **Demand-driven.** Nothing is implemented because it might be needed. The shim refuses by name,
159
+ the refusal names the next thing to build, and that list comes from running real models.
160
+
161
+ **Stable ABI.** Built against CPython's limited API (`abi3-py313`), so one binary per platform
162
+ loads on 3.13, 3.14 and later without a rebuild.
163
+
164
+ ---
165
+
166
+ ## Status
167
+
168
+ <table>
169
+ <tr><th align="left">Working</th><th align="left"></th></tr>
170
+ <tr><td>ATen operators</td><td><b>91</b>, each compared against upstream</td></tr>
171
+ <tr><td>Golden comparison cases</td><td><b>2095 / 2095</b> — values, shapes, dtypes</td></tr>
172
+ <tr><td>Python spellings</td><td><b>204</b> verified against upstream signatures</td></tr>
173
+ <tr><td>Architectures complete</td><td><b>15 of 20</b> measured</td></tr>
174
+ <tr><td>Build targets</td><td>macOS · Linux · Android arm64 · iOS arm64</td></tr>
175
+ </table>
176
+
177
+ Complete: Llama · GPT-2 · Qwen2 · Mistral · Gemma · GPT-NeoX · OPT · MPT · StarCoder2 ·
178
+ Persimmon · Cohere · StableLM · OLMo · Phi · BERT
179
+
180
+ `uniform_` and `normal_` are **bit-identical** to upstream, and `multinomial` consumes the same
181
+ generator stream — a seeded run reproduces exactly.
182
+
183
+ **Not working yet**
184
+
185
+ - `import transformers` fails — `torch.distributed` is unimplemented and an unguarded import
186
+ inside `torch._dynamo` reaches `dist.Store`. Every result above was measured against models
187
+ transcribed by hand.
188
+ - No checkpoint has ever been loaded. All weights so far are randomly initialised.
189
+ - Mobile is link-verified only — the artefacts build and link, but no device has loaded one.
190
+ - CPU only. No GPU or NPU backend.
191
+
192
+ Tracked with the measurements behind them in [`docs/DESIGN.md`](docs/DESIGN.md) §11.1.
193
+
194
+ ---
195
+
196
+ ## Verification
197
+
198
+ Correctness here means *agreeing with upstream PyTorch*, so the strategy is comparison rather
199
+ than assertion.
200
+
201
+ | | |
202
+ |---|---|
203
+ | **Golden comparison** | Every operator runs on both upstream torch and this shim, compared on value, shape and dtype. It has caught a `float16` GEMM accumulating in `float16` where torch accumulates in `float32`, `cumsum` routed through the wrong kernel, and integer overflow where torch refuses. |
204
+ | **The harness tests itself** | `--self-test` injects a fault shaped like a plausible misimplementation at each comparator and fails if the comparator accepts it — 11 comparators × 11 fault modes, with any comparator never exercised reported as failure. It found that the previous fault injection reached exactly one case out of 1781. |
205
+ | **Tokens are not enough** | A wrong `gelu` approximation produced *identical tokens* while logits differed by 5.9e-04. End-to-end tests compare logits too, with a tolerance measured to sit between normal float32 noise and that failure. |
206
+
207
+ ```sh
208
+ sh rust/torch_c/pytests/run.sh # smoke tests + harness self-test
209
+ python tools/golden/compare.py # golden comparison against upstream
210
+ python rust/torch_c/pytests/verify_schemas.py # signature tables vs upstream
211
+ ```
212
+
213
+ ---
214
+
215
+ ## Roadmap
216
+
217
+ The next milestone is the device abstraction, because everything waits on it — a distributed rank
218
+ needs a device to point at, and every accelerator attaches there.
219
+
220
+ ```
221
+ torchnative.nn.federated rounds · client selection · aggregation · dropout
222
+ └ torch.distributed ProcessGroup · collectives (transport)
223
+ └ backends ours, via register_backend
224
+ └ devices CPU · Metal · Vulkan · NPU
225
+ ```
226
+
227
+ | | |
228
+ |---|---|
229
+ | **Device abstraction** | `torch.device`, per-device dispatch. Everything else waits on it. |
230
+ | **Metal** | candle already has the backend; disabled here for build isolation, not absent. |
231
+ | **`torch.distributed`** | From `world_size = 1` upward. Unblocks `transformers` as a side effect. |
232
+ | **Android GPU** | No candle backend and no `vulkan` slot in the `kernels` contract — genuinely new work. |
233
+ | **NPU** | ANE, NNAPI and QNN take a whole graph ahead of time, so this needs a capture layer rather than another device. The single door is where it attaches. |
234
+
235
+ ---
236
+
237
+ ## Install
238
+
239
+ > Not published yet. The alpha will carry platform wheels for macOS, Linux, Android and iOS —
240
+ > the extension is native, so `py3-none-any` is not the shape this ships in.
241
+
242
+ ```sh
243
+ pip install torchnative
244
+ ```
245
+
246
+ ### Building from source
247
+
248
+ Requires a Rust toolchain and CPython 3.13+.
249
+
250
+ ```sh
251
+ bash vendor/vendor_torch.sh # assemble the vendored torch tree
252
+ bash vendor/install_shim.sh # build the extension and install it
253
+ ```
254
+
255
+ Cross-compilation is documented in [`docs/RUST_CROSSBUILD.md`](docs/RUST_CROSSBUILD.md),
256
+ including the PyO3 configuration iOS needs in order not to link `libpython`.
257
+
258
+ ---
259
+
260
+ ## Repository layout
261
+
262
+ ```
263
+ torchnative/ the Python library
264
+ rust/torch_c/ the torch._C replacement (Rust · PyO3 · candle)
265
+ tools/golden/ the upstream comparison harness
266
+ vendor/ scripts that assemble the vendored torch tree (not checked in)
267
+ docs/ design, measurements, and the reasoning behind open decisions
268
+ ```
269
+
270
+ `docs/` is written to be read. It records what was measured, what was assumed, and where an
271
+ earlier conclusion turned out to be wrong — corrections are left visible rather than edited away.
272
+ Start with [`DESIGN.md`](docs/DESIGN.md); [`SURFACE_HONESTY.md`](docs/SURFACE_HONESTY.md) and
273
+ [`HARNESS.md`](docs/HARNESS.md) show the standard the rest aims for.
274
+
275
+ ---
276
+
277
+ ## Related
278
+
279
+ - [PythonMultiplatform](https://github.com/thisisthepy/PythonMultiplatform) — embeds CPython 3.13
280
+ into Kotlin Multiplatform; the deployment target for this library
281
+ - [pypackpack](https://github.com/thisisthepy/pypackpack) — the build and bundling tool
282
+ - [Hugging Face `kernels`](https://github.com/huggingface/kernels) — the fused-kernel contract
283
+ this adopts, with resolution moved from runtime download to build time, since downloading
284
+ executable code is not permitted on every target platform
285
+
286
+ ---
287
+
288
+ ## License
289
+
290
+ MIT — see [LICENSE](LICENSE).
291
+
292
+ PyTorch is vendored under its own BSD-3-Clause license. The vendored tree is assembled at build
293
+ time and is not redistributed in this repository.
@@ -0,0 +1,266 @@
1
+ <div align="center">
2
+
3
+ # torchnative
4
+
5
+ **Run the real PyTorch ecosystem on device — not a reimplementation of it.**
6
+
7
+ [![PyPI](https://img.shields.io/pypi/v/torchnative?color=blue)](https://pypi.org/project/torchnative/)
8
+ [![Python](https://img.shields.io/badge/python-3.13%2B-blue)](https://www.python.org/)
9
+ [![License](https://img.shields.io/badge/license-MIT-green)](LICENSE)
10
+ [![Platforms](https://img.shields.io/badge/platforms-macOS%20%7C%20Linux%20%7C%20Android%20%7C%20iOS-lightgrey)](#install)
11
+ [![Status](https://img.shields.io/badge/status-pre--alpha-orange)](#status)
12
+
13
+ </div>
14
+
15
+ ---
16
+
17
+ `torchnative` replaces PyTorch's compiled core — `torch._C` — with a native extension, so the
18
+ genuine `torch` and `transformers` packages run on a phone the way they run on a workstation.
19
+
20
+ Models are not ported, converted, or re-expressed. They are imported.
21
+
22
+ ```python
23
+ from transformers import AutoModelForCausalLM # the real one
24
+ model = AutoModelForCausalLM.from_pretrained("...")
25
+ model.generate(...) # on the device
26
+ ```
27
+
28
+ > [!WARNING]
29
+ > **Pre-alpha.** The operator layer matches upstream PyTorch numerically and 15 of 20 tested
30
+ > architectures reach zero missing operators — but `import transformers` does not work yet, no
31
+ > checkpoint has ever been loaded, and no device has run the built artefact.
32
+ > See [Status](#status) before depending on this.
33
+
34
+ ---
35
+
36
+ ## Why not a reimplementation
37
+
38
+ Every other route to on-device inference re-expresses the model somewhere else.
39
+
40
+ | | approach | cost |
41
+ |---|---|---|
42
+ | llama.cpp | architectures rewritten in C++ | each new architecture is a porting task |
43
+ | ExecuTorch · CoreML | ahead-of-time compiled graph | export step, and what runs is not what you wrote |
44
+ | MLC | lowered to its own runtime | same |
45
+ | **torchnative** | **the real Python package** | **the substrate is hard; architectures are free** |
46
+
47
+ The reason nobody runs the real thing is that `torch._C` cannot be built for mobile. PyTorch's own
48
+ build sets `INTERN_BUILD_MOBILE` for any Android or iOS toolchain, and that path forces
49
+ `BUILD_PYTHON` off — so the mobile build is structurally incapable of producing the Python
50
+ extension module the Python package needs.
51
+
52
+ `torchnative` supplies that module instead. Everything above it is upstream source, unmodified.
53
+
54
+ ---
55
+
56
+ ## What it does
57
+
58
+ ### 1 · LLM inference
59
+
60
+ Run `transformers` models directly. No conversion step, no per-architecture port — if
61
+ `transformers` supports it and the operators are covered, it runs.
62
+
63
+ ```python
64
+ import torch
65
+ from transformers import AutoModelForCausalLM, AutoTokenizer
66
+
67
+ model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B")
68
+ tok = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")
69
+
70
+ out = model.generate(**tok("On-device inference is", return_tensors="pt"),
71
+ max_new_tokens=32, do_sample=True)
72
+ ```
73
+
74
+ **Today:** the operator layer under this is complete for 15 of 20 architectures, and Llama and
75
+ GPT-2 match upstream **token for token and logit for logit** in both greedy and sampling mode.
76
+ The `from_pretrained` path itself is still blocked — see [Status](#status).
77
+
78
+ ### 2 · Federated learning
79
+
80
+ Devices train locally and share updates, not data. Federated averaging *is* collective
81
+ communication, so this is built on `torch.distributed` rather than beside it — broadcast the
82
+ model, gather the updates, weighted all-reduce.
83
+
84
+ ```python
85
+ from torchnative.nn import federated
86
+
87
+ engine = federated.Engine(model, rounds=..., aggregator=federated.FedAvg())
88
+ engine.participate() # local epochs, then contribute a delta
89
+ ```
90
+
91
+ **Today:** planned. `torch.distributed` is being implemented from `world_size = 1` upward, which
92
+ is a truthful description of a single device rather than a stub.
93
+
94
+ ### 3 · Test-time adaptation & training
95
+
96
+ A model that ships to a device meets data the training set never had. TTA, TTT and the wider
97
+ test-time learning family let it adapt in place — and every method reduces to the same thing: a
98
+ weight delta over base weights, differing only in lifetime and destination.
99
+
100
+ ```python
101
+ from torchnative import adapt
102
+
103
+ model = adapt.wrap(model, method=adapt.Tent()) # or TTT, memory-based, entropy-based
104
+ model.online() # adapt as it serves
105
+ ```
106
+
107
+ **Today:** planned; the delta abstraction is specified in [`docs/DESIGN.md`](docs/DESIGN.md) §3.
108
+ Lifetime is driven by system events — backgrounding, user switch, sync window — rather than by
109
+ the domain boundaries a benchmark hands you.
110
+
111
+ ---
112
+
113
+ ## How it works
114
+
115
+ ```
116
+ your code · transformers · torch/*.py upstream Python, unmodified
117
+ ──────────────────────────────────────────
118
+ torch._C ← replaced
119
+ ├── _aten_dispatch the single door every operator passes
120
+ ├── Python spellings torch.mm, x.softmax(), F.linear, ...
121
+ └── kernels Rust, backed by candle
122
+ ──────────────────────────────────────────
123
+ CPU today · Metal, Vulkan, NPU planned
124
+ ```
125
+
126
+ **One door.** Every operator reaches its kernel through `_aten_dispatch`, and nothing bypasses
127
+ it. That makes the surface measurable — an unimplemented operator names itself rather than
128
+ failing downstream — and it gives graph capture, which NPU backends will need, exactly one place
129
+ to attach.
130
+
131
+ **Demand-driven.** Nothing is implemented because it might be needed. The shim refuses by name,
132
+ the refusal names the next thing to build, and that list comes from running real models.
133
+
134
+ **Stable ABI.** Built against CPython's limited API (`abi3-py313`), so one binary per platform
135
+ loads on 3.13, 3.14 and later without a rebuild.
136
+
137
+ ---
138
+
139
+ ## Status
140
+
141
+ <table>
142
+ <tr><th align="left">Working</th><th align="left"></th></tr>
143
+ <tr><td>ATen operators</td><td><b>91</b>, each compared against upstream</td></tr>
144
+ <tr><td>Golden comparison cases</td><td><b>2095 / 2095</b> — values, shapes, dtypes</td></tr>
145
+ <tr><td>Python spellings</td><td><b>204</b> verified against upstream signatures</td></tr>
146
+ <tr><td>Architectures complete</td><td><b>15 of 20</b> measured</td></tr>
147
+ <tr><td>Build targets</td><td>macOS · Linux · Android arm64 · iOS arm64</td></tr>
148
+ </table>
149
+
150
+ Complete: Llama · GPT-2 · Qwen2 · Mistral · Gemma · GPT-NeoX · OPT · MPT · StarCoder2 ·
151
+ Persimmon · Cohere · StableLM · OLMo · Phi · BERT
152
+
153
+ `uniform_` and `normal_` are **bit-identical** to upstream, and `multinomial` consumes the same
154
+ generator stream — a seeded run reproduces exactly.
155
+
156
+ **Not working yet**
157
+
158
+ - `import transformers` fails — `torch.distributed` is unimplemented and an unguarded import
159
+ inside `torch._dynamo` reaches `dist.Store`. Every result above was measured against models
160
+ transcribed by hand.
161
+ - No checkpoint has ever been loaded. All weights so far are randomly initialised.
162
+ - Mobile is link-verified only — the artefacts build and link, but no device has loaded one.
163
+ - CPU only. No GPU or NPU backend.
164
+
165
+ Tracked with the measurements behind them in [`docs/DESIGN.md`](docs/DESIGN.md) §11.1.
166
+
167
+ ---
168
+
169
+ ## Verification
170
+
171
+ Correctness here means *agreeing with upstream PyTorch*, so the strategy is comparison rather
172
+ than assertion.
173
+
174
+ | | |
175
+ |---|---|
176
+ | **Golden comparison** | Every operator runs on both upstream torch and this shim, compared on value, shape and dtype. It has caught a `float16` GEMM accumulating in `float16` where torch accumulates in `float32`, `cumsum` routed through the wrong kernel, and integer overflow where torch refuses. |
177
+ | **The harness tests itself** | `--self-test` injects a fault shaped like a plausible misimplementation at each comparator and fails if the comparator accepts it — 11 comparators × 11 fault modes, with any comparator never exercised reported as failure. It found that the previous fault injection reached exactly one case out of 1781. |
178
+ | **Tokens are not enough** | A wrong `gelu` approximation produced *identical tokens* while logits differed by 5.9e-04. End-to-end tests compare logits too, with a tolerance measured to sit between normal float32 noise and that failure. |
179
+
180
+ ```sh
181
+ sh rust/torch_c/pytests/run.sh # smoke tests + harness self-test
182
+ python tools/golden/compare.py # golden comparison against upstream
183
+ python rust/torch_c/pytests/verify_schemas.py # signature tables vs upstream
184
+ ```
185
+
186
+ ---
187
+
188
+ ## Roadmap
189
+
190
+ The next milestone is the device abstraction, because everything waits on it — a distributed rank
191
+ needs a device to point at, and every accelerator attaches there.
192
+
193
+ ```
194
+ torchnative.nn.federated rounds · client selection · aggregation · dropout
195
+ └ torch.distributed ProcessGroup · collectives (transport)
196
+ └ backends ours, via register_backend
197
+ └ devices CPU · Metal · Vulkan · NPU
198
+ ```
199
+
200
+ | | |
201
+ |---|---|
202
+ | **Device abstraction** | `torch.device`, per-device dispatch. Everything else waits on it. |
203
+ | **Metal** | candle already has the backend; disabled here for build isolation, not absent. |
204
+ | **`torch.distributed`** | From `world_size = 1` upward. Unblocks `transformers` as a side effect. |
205
+ | **Android GPU** | No candle backend and no `vulkan` slot in the `kernels` contract — genuinely new work. |
206
+ | **NPU** | ANE, NNAPI and QNN take a whole graph ahead of time, so this needs a capture layer rather than another device. The single door is where it attaches. |
207
+
208
+ ---
209
+
210
+ ## Install
211
+
212
+ > Not published yet. The alpha will carry platform wheels for macOS, Linux, Android and iOS —
213
+ > the extension is native, so `py3-none-any` is not the shape this ships in.
214
+
215
+ ```sh
216
+ pip install torchnative
217
+ ```
218
+
219
+ ### Building from source
220
+
221
+ Requires a Rust toolchain and CPython 3.13+.
222
+
223
+ ```sh
224
+ bash vendor/vendor_torch.sh # assemble the vendored torch tree
225
+ bash vendor/install_shim.sh # build the extension and install it
226
+ ```
227
+
228
+ Cross-compilation is documented in [`docs/RUST_CROSSBUILD.md`](docs/RUST_CROSSBUILD.md),
229
+ including the PyO3 configuration iOS needs in order not to link `libpython`.
230
+
231
+ ---
232
+
233
+ ## Repository layout
234
+
235
+ ```
236
+ torchnative/ the Python library
237
+ rust/torch_c/ the torch._C replacement (Rust · PyO3 · candle)
238
+ tools/golden/ the upstream comparison harness
239
+ vendor/ scripts that assemble the vendored torch tree (not checked in)
240
+ docs/ design, measurements, and the reasoning behind open decisions
241
+ ```
242
+
243
+ `docs/` is written to be read. It records what was measured, what was assumed, and where an
244
+ earlier conclusion turned out to be wrong — corrections are left visible rather than edited away.
245
+ Start with [`DESIGN.md`](docs/DESIGN.md); [`SURFACE_HONESTY.md`](docs/SURFACE_HONESTY.md) and
246
+ [`HARNESS.md`](docs/HARNESS.md) show the standard the rest aims for.
247
+
248
+ ---
249
+
250
+ ## Related
251
+
252
+ - [PythonMultiplatform](https://github.com/thisisthepy/PythonMultiplatform) — embeds CPython 3.13
253
+ into Kotlin Multiplatform; the deployment target for this library
254
+ - [pypackpack](https://github.com/thisisthepy/pypackpack) — the build and bundling tool
255
+ - [Hugging Face `kernels`](https://github.com/huggingface/kernels) — the fused-kernel contract
256
+ this adopts, with resolution moved from runtime download to build time, since downloading
257
+ executable code is not permitted on every target platform
258
+
259
+ ---
260
+
261
+ ## License
262
+
263
+ MIT — see [LICENSE](LICENSE).
264
+
265
+ PyTorch is vendored under its own BSD-3-Clause license. The vendored tree is assembled at build
266
+ time and is not redistributed in this repository.
@@ -0,0 +1,87 @@
1
+ [build-system]
2
+ requires = ["setuptools>=77"]
3
+ build-backend = "setuptools.build_meta"
4
+
5
+ [project]
6
+ name = "torchnative"
7
+ version = "0.0.1a0"
8
+ description = "Run the real PyTorch ecosystem on device -- not a reimplementation of it"
9
+ readme = { file = "README.md", content-type = "text/markdown" }
10
+ license = "MIT"
11
+ license-files = ["LICENSE"]
12
+ requires-python = ">=3.13"
13
+
14
+ # Empty on purpose. This distribution **provides** `torch` -- the upstream
15
+ # Python tree with `_C` replaced (DESIGN.md §2) -- rather than consuming
16
+ # somebody else's, so requiring `torch` here would be a package declaring a
17
+ # dependency on itself.
18
+ #
19
+ # Two earlier attempts at this line were both wrong, in opposite directions.
20
+ # Empty with a module-scope `from torch import nn` installed cleanly and then
21
+ # failed on import. Requiring `torch` unconditionally was unresolvable on
22
+ # Android and iOS, where upstream publishes no wheel at all -- which is to say
23
+ # on exactly the platforms this project exists for. The mistake underneath both
24
+ # was reading upstream torch as a runtime dependency when it is a *comparison
25
+ # baseline*: the golden harness needs one installed to diff against, and nothing
26
+ # at runtime does.
27
+ #
28
+ # It follows that this cannot coexist with an installed PyTorch once the wheels
29
+ # carry the tree. That is not a defect to route around; there is one `torch` on
30
+ # a given interpreter, and this is a build of it.
31
+ dependencies = []
32
+ authors = [{ name = "thisisthepy" }]
33
+ keywords = ["pytorch", "on-device", "edge", "federated-learning", "test-time-adaptation"]
34
+
35
+ # Pre-alpha on purpose. The Python surface here is still a skeleton; the working
36
+ # part of this project is the `torch._C` replacement under `rust/torch_c`, which
37
+ # is not yet built into this wheel. See the README's Status section.
38
+ classifiers = [
39
+ "Development Status :: 2 - Pre-Alpha",
40
+ "Intended Audience :: Developers",
41
+ "Programming Language :: Python :: 3.13",
42
+ "Programming Language :: Rust",
43
+ "Topic :: Scientific/Engineering :: Artificial Intelligence",
44
+ "Operating System :: Android",
45
+ "Operating System :: iOS",
46
+ "Operating System :: MacOS",
47
+ "Operating System :: POSIX :: Linux",
48
+ ]
49
+
50
+ [project.urls]
51
+ Homepage = "https://github.com/thisisthepy/torchnative"
52
+ Documentation = "https://github.com/thisisthepy/torchnative/tree/develop/docs"
53
+ Source = "https://github.com/thisisthepy/torchnative"
54
+
55
+ [project.optional-dependencies]
56
+ # Aggregation, transport and privacy live behind this extra so that using
57
+ # adaptation alone does not pull in the federated stack. See docs/DESIGN.md §10.
58
+ federated = []
59
+
60
+ # Upstream PyTorch, as the thing the golden harness diffs against -- not as a
61
+ # runtime dependency. Pinned to what is actually compared: this shim implements
62
+ # one release's `_C` surface, so a tree from another release expects different
63
+ # symbols from it. Installing this alongside the wheels that carry our own tree
64
+ # will conflict, which is why it is an extra and not a dependency.
65
+ test = ["torch>=2.13,<2.14"]
66
+
67
+ # Without this, setuptools auto-discovery treats `src` as the root and ships
68
+ # `main/` and `test/` as importable packages -- which is what it did, so a wheel
69
+ # built before this section answered `import main.torchnative` and not
70
+ # `import torchnative`. The source-set layout has to be spelled out.
71
+ [tool.setuptools]
72
+ package-dir = { "" = "torchnative/src/main" }
73
+
74
+ [tool.setuptools.packages.find]
75
+ where = ["torchnative/src/main"]
76
+ # `torch` is deliberately not shipped from this distribution. src/main/torch is
77
+ # the add-hook that grafts `torchnative.nn.federated` onto the torch namespace
78
+ # (DESIGN.md §2), and writing files into another distribution's package would
79
+ # collide with an installed PyTorch, break its uninstall, and conflict outright
80
+ # if upstream ever adds that path. It stays a build-time graft against the
81
+ # vendored tree we assemble ourselves.
82
+ include = ["torchnative*"]
83
+
84
+ [tool.ppp]
85
+ # Platform source sets follow the pypackpack layout: src/main is scanned for
86
+ # top-level Python packages, which is what lets one package provide both
87
+ # `torch` and `torchnative`. See docs/DESIGN.md §10.
@@ -0,0 +1,4 @@
1
+ [egg_info]
2
+ tag_build =
3
+ tag_date = 0
4
+
@@ -0,0 +1,4 @@
1
+ """torchnative — on-device test-time learning and federated learning.
2
+
3
+ See docs/DESIGN.md for the design and its reasoning.
4
+ """
@@ -0,0 +1,8 @@
1
+ """Test-time learning methods.
2
+
3
+ TTL contains TTA contains TTT; the nesting is not flattened into sibling
4
+ modules. Each method declares its own differentiation requirement rather than
5
+ living in a directory named for one, because normalization calibration sits on
6
+ both sides of that line -- recomputing statistics needs no backward pass,
7
+ updating affine parameters by a loss does. See DESIGN.md §3.
8
+ """
@@ -0,0 +1,24 @@
1
+ """Deployment, lifetime policy and device orchestration.
2
+
3
+ The `torch` import is deferred rather than done at module scope. This
4
+ distribution *provides* `torch` -- the vendored tree plus our `_C` -- rather
5
+ than depending on someone else's, so at import time there may not be one yet:
6
+ the wheels that carry the tree are not wired up, and until they are, importing
7
+ `torchnative` should not require a PyTorch that this package is itself meant to
8
+ supply. See DESIGN.md §2 and the README's Status section.
9
+ """
10
+
11
+ from __future__ import annotations
12
+
13
+ from typing import TYPE_CHECKING
14
+
15
+ if TYPE_CHECKING:
16
+ from torch import nn
17
+
18
+
19
+ class TorchNativeAPI(object):
20
+ def __init__(self, *args, **kwargs):
21
+ pass
22
+
23
+ def deploy(self, model: "nn.Module"):
24
+ pass
@@ -0,0 +1,11 @@
1
+ """Weight deltas over base weights.
2
+
3
+ Every adaptation method produces one of these; they differ only in lifetime
4
+ and destination. The boundary is ``model(x)`` -- state a model manages inside
5
+ its own forward (fast weights, caches) is the model's, not ours.
6
+
7
+ Lifetime names are deliberately unnamed for now. They are driven by system
8
+ events -- backgrounding, user switch, sync window -- not by the domain
9
+ boundaries a benchmark hands you, and they get chosen once the first
10
+ integration shows the real usage. See DESIGN.md §3.
11
+ """
@@ -0,0 +1,7 @@
1
+ """Bundle-backed resolver for the Hugging Face ``kernels`` contract.
2
+
3
+ The contract is adopted as-is; only resolution is inverted. On mobile, kernel
4
+ variants are selected and compiled at build time and read from the app bundle,
5
+ because iOS will not execute downloaded native code. Desktop keeps resolving
6
+ against the Hub. See DESIGN.md §8.
7
+ """
@@ -0,0 +1,7 @@
1
+ """Federated learning.
2
+
3
+ A layer above the single-device adaptation methods, not a sibling of them: the
4
+ local step is the same mechanism, with aggregation, transport and privacy on
5
+ top. Its dependencies stay behind the ``federated`` extra so that using
6
+ adaptation alone does not pull in the aggregation stack. See DESIGN.md §3.
7
+ """
@@ -0,0 +1,293 @@
1
+ Metadata-Version: 2.4
2
+ Name: torchnative
3
+ Version: 0.0.1a0
4
+ Summary: Run the real PyTorch ecosystem on device -- not a reimplementation of it
5
+ Author: thisisthepy
6
+ License-Expression: MIT
7
+ Project-URL: Homepage, https://github.com/thisisthepy/torchnative
8
+ Project-URL: Documentation, https://github.com/thisisthepy/torchnative/tree/develop/docs
9
+ Project-URL: Source, https://github.com/thisisthepy/torchnative
10
+ Keywords: pytorch,on-device,edge,federated-learning,test-time-adaptation
11
+ Classifier: Development Status :: 2 - Pre-Alpha
12
+ Classifier: Intended Audience :: Developers
13
+ Classifier: Programming Language :: Python :: 3.13
14
+ Classifier: Programming Language :: Rust
15
+ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
16
+ Classifier: Operating System :: Android
17
+ Classifier: Operating System :: iOS
18
+ Classifier: Operating System :: MacOS
19
+ Classifier: Operating System :: POSIX :: Linux
20
+ Requires-Python: >=3.13
21
+ Description-Content-Type: text/markdown
22
+ License-File: LICENSE
23
+ Provides-Extra: federated
24
+ Provides-Extra: test
25
+ Requires-Dist: torch<2.14,>=2.13; extra == "test"
26
+ Dynamic: license-file
27
+
28
+ <div align="center">
29
+
30
+ # torchnative
31
+
32
+ **Run the real PyTorch ecosystem on device — not a reimplementation of it.**
33
+
34
+ [![PyPI](https://img.shields.io/pypi/v/torchnative?color=blue)](https://pypi.org/project/torchnative/)
35
+ [![Python](https://img.shields.io/badge/python-3.13%2B-blue)](https://www.python.org/)
36
+ [![License](https://img.shields.io/badge/license-MIT-green)](LICENSE)
37
+ [![Platforms](https://img.shields.io/badge/platforms-macOS%20%7C%20Linux%20%7C%20Android%20%7C%20iOS-lightgrey)](#install)
38
+ [![Status](https://img.shields.io/badge/status-pre--alpha-orange)](#status)
39
+
40
+ </div>
41
+
42
+ ---
43
+
44
+ `torchnative` replaces PyTorch's compiled core — `torch._C` — with a native extension, so the
45
+ genuine `torch` and `transformers` packages run on a phone the way they run on a workstation.
46
+
47
+ Models are not ported, converted, or re-expressed. They are imported.
48
+
49
+ ```python
50
+ from transformers import AutoModelForCausalLM # the real one
51
+ model = AutoModelForCausalLM.from_pretrained("...")
52
+ model.generate(...) # on the device
53
+ ```
54
+
55
+ > [!WARNING]
56
+ > **Pre-alpha.** The operator layer matches upstream PyTorch numerically and 15 of 20 tested
57
+ > architectures reach zero missing operators — but `import transformers` does not work yet, no
58
+ > checkpoint has ever been loaded, and no device has run the built artefact.
59
+ > See [Status](#status) before depending on this.
60
+
61
+ ---
62
+
63
+ ## Why not a reimplementation
64
+
65
+ Every other route to on-device inference re-expresses the model somewhere else.
66
+
67
+ | | approach | cost |
68
+ |---|---|---|
69
+ | llama.cpp | architectures rewritten in C++ | each new architecture is a porting task |
70
+ | ExecuTorch · CoreML | ahead-of-time compiled graph | export step, and what runs is not what you wrote |
71
+ | MLC | lowered to its own runtime | same |
72
+ | **torchnative** | **the real Python package** | **the substrate is hard; architectures are free** |
73
+
74
+ The reason nobody runs the real thing is that `torch._C` cannot be built for mobile. PyTorch's own
75
+ build sets `INTERN_BUILD_MOBILE` for any Android or iOS toolchain, and that path forces
76
+ `BUILD_PYTHON` off — so the mobile build is structurally incapable of producing the Python
77
+ extension module the Python package needs.
78
+
79
+ `torchnative` supplies that module instead. Everything above it is upstream source, unmodified.
80
+
81
+ ---
82
+
83
+ ## What it does
84
+
85
+ ### 1 · LLM inference
86
+
87
+ Run `transformers` models directly. No conversion step, no per-architecture port — if
88
+ `transformers` supports it and the operators are covered, it runs.
89
+
90
+ ```python
91
+ import torch
92
+ from transformers import AutoModelForCausalLM, AutoTokenizer
93
+
94
+ model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B")
95
+ tok = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")
96
+
97
+ out = model.generate(**tok("On-device inference is", return_tensors="pt"),
98
+ max_new_tokens=32, do_sample=True)
99
+ ```
100
+
101
+ **Today:** the operator layer under this is complete for 15 of 20 architectures, and Llama and
102
+ GPT-2 match upstream **token for token and logit for logit** in both greedy and sampling mode.
103
+ The `from_pretrained` path itself is still blocked — see [Status](#status).
104
+
105
+ ### 2 · Federated learning
106
+
107
+ Devices train locally and share updates, not data. Federated averaging *is* collective
108
+ communication, so this is built on `torch.distributed` rather than beside it — broadcast the
109
+ model, gather the updates, weighted all-reduce.
110
+
111
+ ```python
112
+ from torchnative.nn import federated
113
+
114
+ engine = federated.Engine(model, rounds=..., aggregator=federated.FedAvg())
115
+ engine.participate() # local epochs, then contribute a delta
116
+ ```
117
+
118
+ **Today:** planned. `torch.distributed` is being implemented from `world_size = 1` upward, which
119
+ is a truthful description of a single device rather than a stub.
120
+
121
+ ### 3 · Test-time adaptation & training
122
+
123
+ A model that ships to a device meets data the training set never had. TTA, TTT and the wider
124
+ test-time learning family let it adapt in place — and every method reduces to the same thing: a
125
+ weight delta over base weights, differing only in lifetime and destination.
126
+
127
+ ```python
128
+ from torchnative import adapt
129
+
130
+ model = adapt.wrap(model, method=adapt.Tent()) # or TTT, memory-based, entropy-based
131
+ model.online() # adapt as it serves
132
+ ```
133
+
134
+ **Today:** planned; the delta abstraction is specified in [`docs/DESIGN.md`](docs/DESIGN.md) §3.
135
+ Lifetime is driven by system events — backgrounding, user switch, sync window — rather than by
136
+ the domain boundaries a benchmark hands you.
137
+
138
+ ---
139
+
140
+ ## How it works
141
+
142
+ ```
143
+ your code · transformers · torch/*.py upstream Python, unmodified
144
+ ──────────────────────────────────────────
145
+ torch._C ← replaced
146
+ ├── _aten_dispatch the single door every operator passes
147
+ ├── Python spellings torch.mm, x.softmax(), F.linear, ...
148
+ └── kernels Rust, backed by candle
149
+ ──────────────────────────────────────────
150
+ CPU today · Metal, Vulkan, NPU planned
151
+ ```
152
+
153
+ **One door.** Every operator reaches its kernel through `_aten_dispatch`, and nothing bypasses
154
+ it. That makes the surface measurable — an unimplemented operator names itself rather than
155
+ failing downstream — and it gives graph capture, which NPU backends will need, exactly one place
156
+ to attach.
157
+
158
+ **Demand-driven.** Nothing is implemented because it might be needed. The shim refuses by name,
159
+ the refusal names the next thing to build, and that list comes from running real models.
160
+
161
+ **Stable ABI.** Built against CPython's limited API (`abi3-py313`), so one binary per platform
162
+ loads on 3.13, 3.14 and later without a rebuild.
163
+
164
+ ---
165
+
166
+ ## Status
167
+
168
+ <table>
169
+ <tr><th align="left">Working</th><th align="left"></th></tr>
170
+ <tr><td>ATen operators</td><td><b>91</b>, each compared against upstream</td></tr>
171
+ <tr><td>Golden comparison cases</td><td><b>2095 / 2095</b> — values, shapes, dtypes</td></tr>
172
+ <tr><td>Python spellings</td><td><b>204</b> verified against upstream signatures</td></tr>
173
+ <tr><td>Architectures complete</td><td><b>15 of 20</b> measured</td></tr>
174
+ <tr><td>Build targets</td><td>macOS · Linux · Android arm64 · iOS arm64</td></tr>
175
+ </table>
176
+
177
+ Complete: Llama · GPT-2 · Qwen2 · Mistral · Gemma · GPT-NeoX · OPT · MPT · StarCoder2 ·
178
+ Persimmon · Cohere · StableLM · OLMo · Phi · BERT
179
+
180
+ `uniform_` and `normal_` are **bit-identical** to upstream, and `multinomial` consumes the same
181
+ generator stream — a seeded run reproduces exactly.
182
+
183
+ **Not working yet**
184
+
185
+ - `import transformers` fails — `torch.distributed` is unimplemented and an unguarded import
186
+ inside `torch._dynamo` reaches `dist.Store`. Every result above was measured against models
187
+ transcribed by hand.
188
+ - No checkpoint has ever been loaded. All weights so far are randomly initialised.
189
+ - Mobile is link-verified only — the artefacts build and link, but no device has loaded one.
190
+ - CPU only. No GPU or NPU backend.
191
+
192
+ Tracked with the measurements behind them in [`docs/DESIGN.md`](docs/DESIGN.md) §11.1.
193
+
194
+ ---
195
+
196
+ ## Verification
197
+
198
+ Correctness here means *agreeing with upstream PyTorch*, so the strategy is comparison rather
199
+ than assertion.
200
+
201
+ | | |
202
+ |---|---|
203
+ | **Golden comparison** | Every operator runs on both upstream torch and this shim, compared on value, shape and dtype. It has caught a `float16` GEMM accumulating in `float16` where torch accumulates in `float32`, `cumsum` routed through the wrong kernel, and integer overflow where torch refuses. |
204
+ | **The harness tests itself** | `--self-test` injects a fault shaped like a plausible misimplementation at each comparator and fails if the comparator accepts it — 11 comparators × 11 fault modes, with any comparator never exercised reported as failure. It found that the previous fault injection reached exactly one case out of 1781. |
205
+ | **Tokens are not enough** | A wrong `gelu` approximation produced *identical tokens* while logits differed by 5.9e-04. End-to-end tests compare logits too, with a tolerance measured to sit between normal float32 noise and that failure. |
206
+
207
+ ```sh
208
+ sh rust/torch_c/pytests/run.sh # smoke tests + harness self-test
209
+ python tools/golden/compare.py # golden comparison against upstream
210
+ python rust/torch_c/pytests/verify_schemas.py # signature tables vs upstream
211
+ ```
212
+
213
+ ---
214
+
215
+ ## Roadmap
216
+
217
+ The next milestone is the device abstraction, because everything waits on it — a distributed rank
218
+ needs a device to point at, and every accelerator attaches there.
219
+
220
+ ```
221
+ torchnative.nn.federated rounds · client selection · aggregation · dropout
222
+ └ torch.distributed ProcessGroup · collectives (transport)
223
+ └ backends ours, via register_backend
224
+ └ devices CPU · Metal · Vulkan · NPU
225
+ ```
226
+
227
+ | | |
228
+ |---|---|
229
+ | **Device abstraction** | `torch.device`, per-device dispatch. Everything else waits on it. |
230
+ | **Metal** | candle already has the backend; disabled here for build isolation, not absent. |
231
+ | **`torch.distributed`** | From `world_size = 1` upward. Unblocks `transformers` as a side effect. |
232
+ | **Android GPU** | No candle backend and no `vulkan` slot in the `kernels` contract — genuinely new work. |
233
+ | **NPU** | ANE, NNAPI and QNN take a whole graph ahead of time, so this needs a capture layer rather than another device. The single door is where it attaches. |
234
+
235
+ ---
236
+
237
+ ## Install
238
+
239
+ > Not published yet. The alpha will carry platform wheels for macOS, Linux, Android and iOS —
240
+ > the extension is native, so `py3-none-any` is not the shape this ships in.
241
+
242
+ ```sh
243
+ pip install torchnative
244
+ ```
245
+
246
+ ### Building from source
247
+
248
+ Requires a Rust toolchain and CPython 3.13+.
249
+
250
+ ```sh
251
+ bash vendor/vendor_torch.sh # assemble the vendored torch tree
252
+ bash vendor/install_shim.sh # build the extension and install it
253
+ ```
254
+
255
+ Cross-compilation is documented in [`docs/RUST_CROSSBUILD.md`](docs/RUST_CROSSBUILD.md),
256
+ including the PyO3 configuration iOS needs in order not to link `libpython`.
257
+
258
+ ---
259
+
260
+ ## Repository layout
261
+
262
+ ```
263
+ torchnative/ the Python library
264
+ rust/torch_c/ the torch._C replacement (Rust · PyO3 · candle)
265
+ tools/golden/ the upstream comparison harness
266
+ vendor/ scripts that assemble the vendored torch tree (not checked in)
267
+ docs/ design, measurements, and the reasoning behind open decisions
268
+ ```
269
+
270
+ `docs/` is written to be read. It records what was measured, what was assumed, and where an
271
+ earlier conclusion turned out to be wrong — corrections are left visible rather than edited away.
272
+ Start with [`DESIGN.md`](docs/DESIGN.md); [`SURFACE_HONESTY.md`](docs/SURFACE_HONESTY.md) and
273
+ [`HARNESS.md`](docs/HARNESS.md) show the standard the rest aims for.
274
+
275
+ ---
276
+
277
+ ## Related
278
+
279
+ - [PythonMultiplatform](https://github.com/thisisthepy/PythonMultiplatform) — embeds CPython 3.13
280
+ into Kotlin Multiplatform; the deployment target for this library
281
+ - [pypackpack](https://github.com/thisisthepy/pypackpack) — the build and bundling tool
282
+ - [Hugging Face `kernels`](https://github.com/huggingface/kernels) — the fused-kernel contract
283
+ this adopts, with resolution moved from runtime download to build time, since downloading
284
+ executable code is not permitted on every target platform
285
+
286
+ ---
287
+
288
+ ## License
289
+
290
+ MIT — see [LICENSE](LICENSE).
291
+
292
+ PyTorch is vendored under its own BSD-3-Clause license. The vendored tree is assembled at build
293
+ time and is not redistributed in this repository.
@@ -0,0 +1,15 @@
1
+ LICENSE
2
+ README.md
3
+ pyproject.toml
4
+ torchnative/src/main/torchnative/__init__.py
5
+ torchnative/src/main/torchnative.egg-info/PKG-INFO
6
+ torchnative/src/main/torchnative.egg-info/SOURCES.txt
7
+ torchnative/src/main/torchnative.egg-info/dependency_links.txt
8
+ torchnative/src/main/torchnative.egg-info/requires.txt
9
+ torchnative/src/main/torchnative.egg-info/top_level.txt
10
+ torchnative/src/main/torchnative/adapt/__init__.py
11
+ torchnative/src/main/torchnative/api/__init__.py
12
+ torchnative/src/main/torchnative/delta/__init__.py
13
+ torchnative/src/main/torchnative/kernels/__init__.py
14
+ torchnative/src/main/torchnative/nn/__init__.py
15
+ torchnative/src/main/torchnative/nn/federated/__init__.py
@@ -0,0 +1,5 @@
1
+
2
+ [federated]
3
+
4
+ [test]
5
+ torch<2.14,>=2.13