airon 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.
- airon-0.1.0/.gitignore +37 -0
- airon-0.1.0/CHANGELOG.md +21 -0
- airon-0.1.0/LICENSE +21 -0
- airon-0.1.0/PKG-INFO +212 -0
- airon-0.1.0/README.md +154 -0
- airon-0.1.0/docs/protocol.md +186 -0
- airon-0.1.0/docs/releasing.md +37 -0
- airon-0.1.0/docs/usage.md +156 -0
- airon-0.1.0/examples/custom_loop.py +83 -0
- airon-0.1.0/examples/pytorch_loop.py +70 -0
- airon-0.1.0/pyproject.toml +72 -0
- airon-0.1.0/src/airon/__init__.py +30 -0
- airon-0.1.0/src/airon/__main__.py +5 -0
- airon-0.1.0/src/airon/_config.py +138 -0
- airon-0.1.0/src/airon/_console.py +125 -0
- airon-0.1.0/src/airon/_env.py +163 -0
- airon-0.1.0/src/airon/_hooks.py +116 -0
- airon-0.1.0/src/airon/_log.py +54 -0
- airon-0.1.0/src/airon/_names.py +81 -0
- airon-0.1.0/src/airon/_run.py +672 -0
- airon-0.1.0/src/airon/_sender.py +454 -0
- airon-0.1.0/src/airon/_spool.py +199 -0
- airon-0.1.0/src/airon/_sync.py +110 -0
- airon-0.1.0/src/airon/_system.py +302 -0
- airon-0.1.0/src/airon/_transport.py +242 -0
- airon-0.1.0/src/airon/_values.py +132 -0
- airon-0.1.0/src/airon/_version.py +4 -0
- airon-0.1.0/src/airon/cli.py +330 -0
- airon-0.1.0/src/airon/py.typed +0 -0
- airon-0.1.0/tests/conftest.py +84 -0
- airon-0.1.0/tests/fake_server.py +339 -0
- airon-0.1.0/tests/test_batching.py +81 -0
- airon-0.1.0/tests/test_cli.py +251 -0
- airon-0.1.0/tests/test_config.py +349 -0
- airon-0.1.0/tests/test_console.py +195 -0
- airon-0.1.0/tests/test_happy_path.py +240 -0
- airon-0.1.0/tests/test_lifecycle.py +315 -0
- airon-0.1.0/tests/test_resilience.py +202 -0
- airon-0.1.0/tests/test_retries.py +165 -0
- airon-0.1.0/tests/test_spool.py +261 -0
- airon-0.1.0/tests/test_system.py +355 -0
- airon-0.1.0/tests/test_timing.py +53 -0
- airon-0.1.0/tests/test_values.py +122 -0
airon-0.1.0/.gitignore
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
1
|
+
# Secrets and local configuration
|
|
2
|
+
.env
|
|
3
|
+
.env.*
|
|
4
|
+
!.env.example
|
|
5
|
+
*.pem
|
|
6
|
+
*.key
|
|
7
|
+
|
|
8
|
+
# Python
|
|
9
|
+
__pycache__/
|
|
10
|
+
*.py[cod]
|
|
11
|
+
.venv/
|
|
12
|
+
venv/
|
|
13
|
+
*.egg-info/
|
|
14
|
+
build/
|
|
15
|
+
dist/
|
|
16
|
+
.pytest_cache/
|
|
17
|
+
.mypy_cache/
|
|
18
|
+
.ruff_cache/
|
|
19
|
+
.coverage
|
|
20
|
+
htmlcov/
|
|
21
|
+
.python-version
|
|
22
|
+
|
|
23
|
+
# The offline spool and local credentials must never be committed
|
|
24
|
+
.airon/
|
|
25
|
+
credentials.json
|
|
26
|
+
|
|
27
|
+
# Editors and OS
|
|
28
|
+
.DS_Store
|
|
29
|
+
.idea/
|
|
30
|
+
.vscode/
|
|
31
|
+
*.swp
|
|
32
|
+
|
|
33
|
+
# Local tool state
|
|
34
|
+
.claude/
|
|
35
|
+
|
|
36
|
+
# uv resolves the dev environment itself; the SDK is a library
|
|
37
|
+
uv.lock
|
airon-0.1.0/CHANGELOG.md
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
# Changelog
|
|
2
|
+
|
|
3
|
+
All notable changes to this project are documented here. The format follows
|
|
4
|
+
[Keep a Changelog](https://keepachangelog.com/en/1.1.0/) and the project uses
|
|
5
|
+
[semantic versioning](https://semver.org/).
|
|
6
|
+
|
|
7
|
+
## [0.1.0] - unreleased
|
|
8
|
+
|
|
9
|
+
First version, implementing wire protocol 1.
|
|
10
|
+
|
|
11
|
+
### Added
|
|
12
|
+
|
|
13
|
+
- `airon.init`, `run.log`, `run.progress`, `run.config`, `run.summary`, `run.checkpoint`,
|
|
14
|
+
`run.finish`, the `with` form and the module-level `airon.log` and `airon.finish`.
|
|
15
|
+
- Non-blocking logging through a bounded queue and a background sender with batching, gzip,
|
|
16
|
+
sequence numbers, retries with backoff and `Retry-After`, and the protocol's drop rules.
|
|
17
|
+
- Offline spool, automatic redelivery from the running process, and `airon sync`.
|
|
18
|
+
- Failure capture (`failed` with traceback, `killed` on `SIGTERM`) that chains existing hooks.
|
|
19
|
+
- Console tail capture that keeps carriage-return progress bars intact.
|
|
20
|
+
- Hardware sampling (NVML or `nvidia-smi`, `/proc` or `psutil`) and environment capture.
|
|
21
|
+
- `airon login`, `logout`, `doctor`, `sync` and `version`.
|
airon-0.1.0/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 Manifestro
|
|
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.
|
airon-0.1.0/PKG-INFO
ADDED
|
@@ -0,0 +1,212 @@
|
|
|
1
|
+
Metadata-Version: 2.5
|
|
2
|
+
Name: airon
|
|
3
|
+
Version: 0.1.0
|
|
4
|
+
Summary: Track machine learning runs: metrics, progress and hardware load, sent to an Airon server without ever slowing training down.
|
|
5
|
+
Project-URL: Homepage, https://github.com/Manifestro/airon-metrics
|
|
6
|
+
Project-URL: Issues, https://github.com/Manifestro/airon-metrics/issues
|
|
7
|
+
Project-URL: Changelog, https://github.com/Manifestro/airon-metrics/blob/main/CHANGELOG.md
|
|
8
|
+
Author: Manifestro
|
|
9
|
+
License: MIT License
|
|
10
|
+
|
|
11
|
+
Copyright (c) 2026 Manifestro
|
|
12
|
+
|
|
13
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
14
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
15
|
+
in the Software without restriction, including without limitation the rights
|
|
16
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
17
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
18
|
+
furnished to do so, subject to the following conditions:
|
|
19
|
+
|
|
20
|
+
The above copyright notice and this permission notice shall be included in all
|
|
21
|
+
copies or substantial portions of the Software.
|
|
22
|
+
|
|
23
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
24
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
25
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
26
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
27
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
28
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
29
|
+
SOFTWARE.
|
|
30
|
+
License-File: LICENSE
|
|
31
|
+
Keywords: experiment-tracking,machine-learning,metrics,mlops,monitoring
|
|
32
|
+
Classifier: Development Status :: 4 - Beta
|
|
33
|
+
Classifier: Intended Audience :: Science/Research
|
|
34
|
+
Classifier: License :: OSI Approved :: MIT License
|
|
35
|
+
Classifier: Operating System :: MacOS
|
|
36
|
+
Classifier: Operating System :: POSIX :: Linux
|
|
37
|
+
Classifier: Programming Language :: Python :: 3
|
|
38
|
+
Classifier: Programming Language :: Python :: 3.9
|
|
39
|
+
Classifier: Programming Language :: Python :: 3.10
|
|
40
|
+
Classifier: Programming Language :: Python :: 3.11
|
|
41
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
42
|
+
Classifier: Programming Language :: Python :: 3.13
|
|
43
|
+
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
|
|
44
|
+
Classifier: Typing :: Typed
|
|
45
|
+
Requires-Python: >=3.9
|
|
46
|
+
Provides-Extra: dev
|
|
47
|
+
Requires-Dist: build; extra == 'dev'
|
|
48
|
+
Requires-Dist: mypy; extra == 'dev'
|
|
49
|
+
Requires-Dist: pytest-timeout; extra == 'dev'
|
|
50
|
+
Requires-Dist: pytest>=7; extra == 'dev'
|
|
51
|
+
Requires-Dist: ruff; extra == 'dev'
|
|
52
|
+
Requires-Dist: twine; extra == 'dev'
|
|
53
|
+
Provides-Extra: gpu
|
|
54
|
+
Requires-Dist: nvidia-ml-py; extra == 'gpu'
|
|
55
|
+
Provides-Extra: system
|
|
56
|
+
Requires-Dist: psutil; extra == 'system'
|
|
57
|
+
Description-Content-Type: text/markdown
|
|
58
|
+
|
|
59
|
+
# airon
|
|
60
|
+
|
|
61
|
+
Track machine learning runs from a plain Python training loop. `airon` sends metrics, progress and
|
|
62
|
+
hardware load to an [Airon](https://github.com/Manifestro/airon) server, where people watch loss,
|
|
63
|
+
ETA and GPU usage live.
|
|
64
|
+
|
|
65
|
+
- **One install, two variables, three lines.**
|
|
66
|
+
- **Never in the way of training.** Logging is a queue append; the network, the server and the disk
|
|
67
|
+
cannot slow down or break your loop (see [the guarantees](#it-will-never-break-your-training)).
|
|
68
|
+
- **Built for rented GPUs.** Flaky networks, spot preemption and pods you cannot SSH into are the
|
|
69
|
+
normal case: data that cannot be sent is kept on disk and uploaded later.
|
|
70
|
+
- **No required dependencies.** Python 3.9 to 3.13, standard library only.
|
|
71
|
+
|
|
72
|
+
## Install
|
|
73
|
+
|
|
74
|
+
```bash
|
|
75
|
+
pip install airon
|
|
76
|
+
```
|
|
77
|
+
|
|
78
|
+
Optional extras give better hardware readings: `pip install "airon[gpu]"` (NVML through
|
|
79
|
+
`nvidia-ml-py`) and `pip install "airon[system]"` (`psutil`, which also enables CPU and RAM on
|
|
80
|
+
macOS). Without them the SDK falls back to `nvidia-smi` and `/proc`.
|
|
81
|
+
|
|
82
|
+
## Set up
|
|
83
|
+
|
|
84
|
+
```bash
|
|
85
|
+
export AIRON_URL=https://airon.example.com
|
|
86
|
+
export AIRON_API_KEY=... # a key that may log runs
|
|
87
|
+
```
|
|
88
|
+
|
|
89
|
+
Or run `airon login` once on a machine, which asks for both and saves them in
|
|
90
|
+
`~/.config/airon/credentials.json` (readable only by you). `airon doctor` checks the setup and tells
|
|
91
|
+
you what to fix. There is no default server: if nothing is configured, `airon.init` says so and
|
|
92
|
+
saves the run on disk instead of failing.
|
|
93
|
+
|
|
94
|
+
## Use
|
|
95
|
+
|
|
96
|
+
```python
|
|
97
|
+
import airon
|
|
98
|
+
|
|
99
|
+
with airon.init(
|
|
100
|
+
project="memory-layer",
|
|
101
|
+
name="baseline-v2",
|
|
102
|
+
config={"lr": 3e-5, "batch_size": 32},
|
|
103
|
+
total_steps=50_000,
|
|
104
|
+
total_epochs=5,
|
|
105
|
+
) as run:
|
|
106
|
+
for step, batch in enumerate(loader):
|
|
107
|
+
loss = train_step(batch)
|
|
108
|
+
run.log(
|
|
109
|
+
{"train/loss": loss, "lr": scheduler.get_last_lr()[0]},
|
|
110
|
+
step=step,
|
|
111
|
+
epoch=step / steps_per_epoch,
|
|
112
|
+
samples=32,
|
|
113
|
+
tokens=32 * 512,
|
|
114
|
+
)
|
|
115
|
+
if step % 1000 == 0:
|
|
116
|
+
run.log({"eval/loss": evaluate()}, step=step)
|
|
117
|
+
```
|
|
118
|
+
|
|
119
|
+
Leaving the `with` block normally marks the run `finished`; an exception marks it `failed` with the
|
|
120
|
+
traceback and is raised again. Without `with`, call `run.finish()` (or do nothing: a normal exit
|
|
121
|
+
finishes the run too). Metric values can be Python numbers, numpy or torch scalars and tensors with
|
|
122
|
+
one element; `NaN` and infinity are recorded, not dropped.
|
|
123
|
+
|
|
124
|
+
More patterns, such as eval loops, resuming, checkpoint pointers and disabling, are in
|
|
125
|
+
[docs/usage.md](docs/usage.md). Runnable examples: [`examples/custom_loop.py`](examples/custom_loop.py)
|
|
126
|
+
(no dependencies) and [`examples/pytorch_loop.py`](examples/pytorch_loop.py).
|
|
127
|
+
|
|
128
|
+
## What is collected automatically
|
|
129
|
+
|
|
130
|
+
| What | Details |
|
|
131
|
+
|---|---|
|
|
132
|
+
| Environment | Python and platform, host name, GPU names and memory, git commit, branch, dirty flag and remote (credentials in a remote URL are removed), versions of well-known libraries that your program has **already imported** (airon never imports torch itself), and the cloud when it can tell (`RUNPOD_POD_ID` gives `runpod`; override with `AIRON_CLOUD`) |
|
|
133
|
+
| Hardware, every 10 s | `gpu{i}/util_pct`, `mem_used_mb`, `mem_total_mb`, `temp_c`, `power_w`; `cpu/util_pct`; `ram/used_mb`, `ram/total_mb`; `disk/used_gb`, `disk/total_gb` for the working directory. Inside a container the RAM figures follow the container's memory limit |
|
|
134
|
+
| Progress | `steps_per_sec` (smoothed, from step deltas over time, so logging every 100 steps is fine), and `samples_per_sec` and `tokens_per_sec` when you pass `samples` and `tokens`. The server computes the ETA |
|
|
135
|
+
| Console | The last lines of stdout and stderr, with `\r` progress bars (tqdm) collapsed to their final state. Output is never swallowed. Turn off with `console=False` |
|
|
136
|
+
| Heartbeat | Every 30 s when nothing else was sent, so the server can tell a quiet run from a dead one |
|
|
137
|
+
|
|
138
|
+
Environment variables are never sent wholesale, and config values whose key looks like a secret
|
|
139
|
+
(`token`, `password`, `api_key`, ...) are replaced with `***`.
|
|
140
|
+
|
|
141
|
+
## Offline behaviour
|
|
142
|
+
|
|
143
|
+
Whatever cannot be delivered goes to JSON lines under `~/.cache/airon/runs/<run id>/`
|
|
144
|
+
(`$AIRON_DIR` changes the location). While the server is unreachable the running process keeps
|
|
145
|
+
trying and uploads the backlog, in order, as soon as it answers. After a crash or a long outage run
|
|
146
|
+
`airon sync` to upload what is left. `AIRON_MODE=offline` writes only to disk, for pods without
|
|
147
|
+
internet.
|
|
148
|
+
|
|
149
|
+
## Distributed training
|
|
150
|
+
|
|
151
|
+
With `torchrun` or similar only rank zero logs by default: processes with `RANK` or `LOCAL_RANK`
|
|
152
|
+
other than 0 receive a run object whose methods do nothing, so the same script works everywhere.
|
|
153
|
+
Pass `rank_zero_only=False` to log from every process.
|
|
154
|
+
|
|
155
|
+
## It will never break your training
|
|
156
|
+
|
|
157
|
+
The first requirement of this SDK, and each point has tests:
|
|
158
|
+
|
|
159
|
+
- `log()` is an append to a bounded in-memory queue: a few microseconds (median about 4 us on a
|
|
160
|
+
laptop) under a lock that is held only for the append. No network or disk work happens in your
|
|
161
|
+
thread; a background thread sends.
|
|
162
|
+
- Nothing raises into your code because of the network, the server or the disk: down, slow, wrong
|
|
163
|
+
address, DNS failure, TLS errors, garbage replies, a read-only or full disk, and 401/403/409/5xx
|
|
164
|
+
answers are all handled. Bad arguments to `init` or `log` produce at most one warning.
|
|
165
|
+
- At most one warning per kind of problem is printed, and the API key is removed from every message.
|
|
166
|
+
- Memory is bounded: at 100,000 queued events the oldest hardware samples go first, then the oldest
|
|
167
|
+
metrics, with one warning.
|
|
168
|
+
- Retries use exponential backoff with jitter (1 s to 60 s) and honour `Retry-After`; retrying is
|
|
169
|
+
safe because the server ignores sequence numbers it already has.
|
|
170
|
+
- At exit, flushing waits at most 10 seconds; whatever is left goes to disk.
|
|
171
|
+
- Your exception hook and `SIGTERM` handler are chained, never replaced: `failed` (with traceback) is
|
|
172
|
+
reported after your hook has printed the error, and `killed` after your own handler has run.
|
|
173
|
+
- Threads are daemon threads; a forked child never reports for its parent.
|
|
174
|
+
- `AIRON_MODE=disabled` makes every call a no-op.
|
|
175
|
+
|
|
176
|
+
## Environment variables
|
|
177
|
+
|
|
178
|
+
| Variable | Meaning |
|
|
179
|
+
|---|---|
|
|
180
|
+
| `AIRON_URL` | Server address, for example `https://airon.example.com` |
|
|
181
|
+
| `AIRON_API_KEY` | The API key |
|
|
182
|
+
| `AIRON_PROJECT` | Default project (otherwise the name of the working directory) |
|
|
183
|
+
| `AIRON_RUN_ID` | Run to resume or attach to |
|
|
184
|
+
| `AIRON_MODE` | `online` (default), `offline` or `disabled` |
|
|
185
|
+
| `AIRON_DIR` | Where the spool and the credentials file live (default `~/.cache/airon` and `~/.config/airon`) |
|
|
186
|
+
| `AIRON_CLOUD` | Overrides the detected cloud name |
|
|
187
|
+
| `AIRON_QUIET` | `1` hides the start and finish lines (warnings still appear) |
|
|
188
|
+
|
|
189
|
+
Arguments to `airon.init` beat environment variables, which beat the credentials file.
|
|
190
|
+
|
|
191
|
+
## Command line
|
|
192
|
+
|
|
193
|
+
| Command | |
|
|
194
|
+
|---|---|
|
|
195
|
+
| `airon login` | Ask for the address and key, check them with the server, save them |
|
|
196
|
+
| `airon logout` | Delete the saved credentials |
|
|
197
|
+
| `airon doctor` | Check address, key, reachability and protocol version, write a tiny test run and print its link, with a fix for each failure |
|
|
198
|
+
| `airon sync` | Upload runs that were saved on disk |
|
|
199
|
+
| `airon version` | Print the version |
|
|
200
|
+
|
|
201
|
+
Exit codes: 0 success, 1 a check or an upload failed, 2 wrong usage or nothing configured.
|
|
202
|
+
|
|
203
|
+
## Links
|
|
204
|
+
|
|
205
|
+
- [Usage guide](docs/usage.md)
|
|
206
|
+
- [Wire protocol](docs/protocol.md), the contract with the server
|
|
207
|
+
- [Releasing](docs/releasing.md), [Contributing](CONTRIBUTING.md), [Security](SECURITY.md), [Changelog](CHANGELOG.md)
|
|
208
|
+
- Server: [Manifestro/airon](https://github.com/Manifestro/airon)
|
|
209
|
+
|
|
210
|
+
## License
|
|
211
|
+
|
|
212
|
+
[MIT](LICENSE)
|
airon-0.1.0/README.md
ADDED
|
@@ -0,0 +1,154 @@
|
|
|
1
|
+
# airon
|
|
2
|
+
|
|
3
|
+
Track machine learning runs from a plain Python training loop. `airon` sends metrics, progress and
|
|
4
|
+
hardware load to an [Airon](https://github.com/Manifestro/airon) server, where people watch loss,
|
|
5
|
+
ETA and GPU usage live.
|
|
6
|
+
|
|
7
|
+
- **One install, two variables, three lines.**
|
|
8
|
+
- **Never in the way of training.** Logging is a queue append; the network, the server and the disk
|
|
9
|
+
cannot slow down or break your loop (see [the guarantees](#it-will-never-break-your-training)).
|
|
10
|
+
- **Built for rented GPUs.** Flaky networks, spot preemption and pods you cannot SSH into are the
|
|
11
|
+
normal case: data that cannot be sent is kept on disk and uploaded later.
|
|
12
|
+
- **No required dependencies.** Python 3.9 to 3.13, standard library only.
|
|
13
|
+
|
|
14
|
+
## Install
|
|
15
|
+
|
|
16
|
+
```bash
|
|
17
|
+
pip install airon
|
|
18
|
+
```
|
|
19
|
+
|
|
20
|
+
Optional extras give better hardware readings: `pip install "airon[gpu]"` (NVML through
|
|
21
|
+
`nvidia-ml-py`) and `pip install "airon[system]"` (`psutil`, which also enables CPU and RAM on
|
|
22
|
+
macOS). Without them the SDK falls back to `nvidia-smi` and `/proc`.
|
|
23
|
+
|
|
24
|
+
## Set up
|
|
25
|
+
|
|
26
|
+
```bash
|
|
27
|
+
export AIRON_URL=https://airon.example.com
|
|
28
|
+
export AIRON_API_KEY=... # a key that may log runs
|
|
29
|
+
```
|
|
30
|
+
|
|
31
|
+
Or run `airon login` once on a machine, which asks for both and saves them in
|
|
32
|
+
`~/.config/airon/credentials.json` (readable only by you). `airon doctor` checks the setup and tells
|
|
33
|
+
you what to fix. There is no default server: if nothing is configured, `airon.init` says so and
|
|
34
|
+
saves the run on disk instead of failing.
|
|
35
|
+
|
|
36
|
+
## Use
|
|
37
|
+
|
|
38
|
+
```python
|
|
39
|
+
import airon
|
|
40
|
+
|
|
41
|
+
with airon.init(
|
|
42
|
+
project="memory-layer",
|
|
43
|
+
name="baseline-v2",
|
|
44
|
+
config={"lr": 3e-5, "batch_size": 32},
|
|
45
|
+
total_steps=50_000,
|
|
46
|
+
total_epochs=5,
|
|
47
|
+
) as run:
|
|
48
|
+
for step, batch in enumerate(loader):
|
|
49
|
+
loss = train_step(batch)
|
|
50
|
+
run.log(
|
|
51
|
+
{"train/loss": loss, "lr": scheduler.get_last_lr()[0]},
|
|
52
|
+
step=step,
|
|
53
|
+
epoch=step / steps_per_epoch,
|
|
54
|
+
samples=32,
|
|
55
|
+
tokens=32 * 512,
|
|
56
|
+
)
|
|
57
|
+
if step % 1000 == 0:
|
|
58
|
+
run.log({"eval/loss": evaluate()}, step=step)
|
|
59
|
+
```
|
|
60
|
+
|
|
61
|
+
Leaving the `with` block normally marks the run `finished`; an exception marks it `failed` with the
|
|
62
|
+
traceback and is raised again. Without `with`, call `run.finish()` (or do nothing: a normal exit
|
|
63
|
+
finishes the run too). Metric values can be Python numbers, numpy or torch scalars and tensors with
|
|
64
|
+
one element; `NaN` and infinity are recorded, not dropped.
|
|
65
|
+
|
|
66
|
+
More patterns, such as eval loops, resuming, checkpoint pointers and disabling, are in
|
|
67
|
+
[docs/usage.md](docs/usage.md). Runnable examples: [`examples/custom_loop.py`](examples/custom_loop.py)
|
|
68
|
+
(no dependencies) and [`examples/pytorch_loop.py`](examples/pytorch_loop.py).
|
|
69
|
+
|
|
70
|
+
## What is collected automatically
|
|
71
|
+
|
|
72
|
+
| What | Details |
|
|
73
|
+
|---|---|
|
|
74
|
+
| Environment | Python and platform, host name, GPU names and memory, git commit, branch, dirty flag and remote (credentials in a remote URL are removed), versions of well-known libraries that your program has **already imported** (airon never imports torch itself), and the cloud when it can tell (`RUNPOD_POD_ID` gives `runpod`; override with `AIRON_CLOUD`) |
|
|
75
|
+
| Hardware, every 10 s | `gpu{i}/util_pct`, `mem_used_mb`, `mem_total_mb`, `temp_c`, `power_w`; `cpu/util_pct`; `ram/used_mb`, `ram/total_mb`; `disk/used_gb`, `disk/total_gb` for the working directory. Inside a container the RAM figures follow the container's memory limit |
|
|
76
|
+
| Progress | `steps_per_sec` (smoothed, from step deltas over time, so logging every 100 steps is fine), and `samples_per_sec` and `tokens_per_sec` when you pass `samples` and `tokens`. The server computes the ETA |
|
|
77
|
+
| Console | The last lines of stdout and stderr, with `\r` progress bars (tqdm) collapsed to their final state. Output is never swallowed. Turn off with `console=False` |
|
|
78
|
+
| Heartbeat | Every 30 s when nothing else was sent, so the server can tell a quiet run from a dead one |
|
|
79
|
+
|
|
80
|
+
Environment variables are never sent wholesale, and config values whose key looks like a secret
|
|
81
|
+
(`token`, `password`, `api_key`, ...) are replaced with `***`.
|
|
82
|
+
|
|
83
|
+
## Offline behaviour
|
|
84
|
+
|
|
85
|
+
Whatever cannot be delivered goes to JSON lines under `~/.cache/airon/runs/<run id>/`
|
|
86
|
+
(`$AIRON_DIR` changes the location). While the server is unreachable the running process keeps
|
|
87
|
+
trying and uploads the backlog, in order, as soon as it answers. After a crash or a long outage run
|
|
88
|
+
`airon sync` to upload what is left. `AIRON_MODE=offline` writes only to disk, for pods without
|
|
89
|
+
internet.
|
|
90
|
+
|
|
91
|
+
## Distributed training
|
|
92
|
+
|
|
93
|
+
With `torchrun` or similar only rank zero logs by default: processes with `RANK` or `LOCAL_RANK`
|
|
94
|
+
other than 0 receive a run object whose methods do nothing, so the same script works everywhere.
|
|
95
|
+
Pass `rank_zero_only=False` to log from every process.
|
|
96
|
+
|
|
97
|
+
## It will never break your training
|
|
98
|
+
|
|
99
|
+
The first requirement of this SDK, and each point has tests:
|
|
100
|
+
|
|
101
|
+
- `log()` is an append to a bounded in-memory queue: a few microseconds (median about 4 us on a
|
|
102
|
+
laptop) under a lock that is held only for the append. No network or disk work happens in your
|
|
103
|
+
thread; a background thread sends.
|
|
104
|
+
- Nothing raises into your code because of the network, the server or the disk: down, slow, wrong
|
|
105
|
+
address, DNS failure, TLS errors, garbage replies, a read-only or full disk, and 401/403/409/5xx
|
|
106
|
+
answers are all handled. Bad arguments to `init` or `log` produce at most one warning.
|
|
107
|
+
- At most one warning per kind of problem is printed, and the API key is removed from every message.
|
|
108
|
+
- Memory is bounded: at 100,000 queued events the oldest hardware samples go first, then the oldest
|
|
109
|
+
metrics, with one warning.
|
|
110
|
+
- Retries use exponential backoff with jitter (1 s to 60 s) and honour `Retry-After`; retrying is
|
|
111
|
+
safe because the server ignores sequence numbers it already has.
|
|
112
|
+
- At exit, flushing waits at most 10 seconds; whatever is left goes to disk.
|
|
113
|
+
- Your exception hook and `SIGTERM` handler are chained, never replaced: `failed` (with traceback) is
|
|
114
|
+
reported after your hook has printed the error, and `killed` after your own handler has run.
|
|
115
|
+
- Threads are daemon threads; a forked child never reports for its parent.
|
|
116
|
+
- `AIRON_MODE=disabled` makes every call a no-op.
|
|
117
|
+
|
|
118
|
+
## Environment variables
|
|
119
|
+
|
|
120
|
+
| Variable | Meaning |
|
|
121
|
+
|---|---|
|
|
122
|
+
| `AIRON_URL` | Server address, for example `https://airon.example.com` |
|
|
123
|
+
| `AIRON_API_KEY` | The API key |
|
|
124
|
+
| `AIRON_PROJECT` | Default project (otherwise the name of the working directory) |
|
|
125
|
+
| `AIRON_RUN_ID` | Run to resume or attach to |
|
|
126
|
+
| `AIRON_MODE` | `online` (default), `offline` or `disabled` |
|
|
127
|
+
| `AIRON_DIR` | Where the spool and the credentials file live (default `~/.cache/airon` and `~/.config/airon`) |
|
|
128
|
+
| `AIRON_CLOUD` | Overrides the detected cloud name |
|
|
129
|
+
| `AIRON_QUIET` | `1` hides the start and finish lines (warnings still appear) |
|
|
130
|
+
|
|
131
|
+
Arguments to `airon.init` beat environment variables, which beat the credentials file.
|
|
132
|
+
|
|
133
|
+
## Command line
|
|
134
|
+
|
|
135
|
+
| Command | |
|
|
136
|
+
|---|---|
|
|
137
|
+
| `airon login` | Ask for the address and key, check them with the server, save them |
|
|
138
|
+
| `airon logout` | Delete the saved credentials |
|
|
139
|
+
| `airon doctor` | Check address, key, reachability and protocol version, write a tiny test run and print its link, with a fix for each failure |
|
|
140
|
+
| `airon sync` | Upload runs that were saved on disk |
|
|
141
|
+
| `airon version` | Print the version |
|
|
142
|
+
|
|
143
|
+
Exit codes: 0 success, 1 a check or an upload failed, 2 wrong usage or nothing configured.
|
|
144
|
+
|
|
145
|
+
## Links
|
|
146
|
+
|
|
147
|
+
- [Usage guide](docs/usage.md)
|
|
148
|
+
- [Wire protocol](docs/protocol.md), the contract with the server
|
|
149
|
+
- [Releasing](docs/releasing.md), [Contributing](CONTRIBUTING.md), [Security](SECURITY.md), [Changelog](CHANGELOG.md)
|
|
150
|
+
- Server: [Manifestro/airon](https://github.com/Manifestro/airon)
|
|
151
|
+
|
|
152
|
+
## License
|
|
153
|
+
|
|
154
|
+
[MIT](LICENSE)
|
|
@@ -0,0 +1,186 @@
|
|
|
1
|
+
# Wire protocol, version 1
|
|
2
|
+
|
|
3
|
+
This document is the contract between the `airon` Python SDK (this repository) and an Airon server.
|
|
4
|
+
Both sides follow it. Where the two disagree, this document wins; open an issue or a pull request to
|
|
5
|
+
change it.
|
|
6
|
+
|
|
7
|
+
Conventions
|
|
8
|
+
|
|
9
|
+
- Bodies are JSON, UTF-8. Request bodies larger than 1 KiB should be gzip-compressed with
|
|
10
|
+
`Content-Encoding: gzip`.
|
|
11
|
+
- Timestamps are Unix seconds as floats, UTC.
|
|
12
|
+
- Unknown fields must be ignored by both sides, so the protocol can grow without breaking old clients.
|
|
13
|
+
- Metric names are hierarchical with `/`, at most 200 characters, for example `train/loss`.
|
|
14
|
+
- JSON cannot carry `NaN` or infinity. A metric value that is not finite is sent as the string
|
|
15
|
+
`"NaN"`, `"Infinity"` or `"-Infinity"`. The server records it (it is a signal worth showing) and
|
|
16
|
+
leaves it out of charts.
|
|
17
|
+
- The examples use `https://airon.example.com`. The SDK has **no default server**: the address always
|
|
18
|
+
comes from `AIRON_URL` or `airon login`.
|
|
19
|
+
|
|
20
|
+
## Authentication and versioning
|
|
21
|
+
|
|
22
|
+
- `Authorization: Bearer <api key>` on every request. Keys never appear in URLs or bodies.
|
|
23
|
+
- `401` invalid or revoked key. `403` the key may not log runs (read-only or wrong scope).
|
|
24
|
+
- Clients send `X-Airon-Protocol: 1` and `User-Agent: airon/<sdk version> python/<python version>`.
|
|
25
|
+
Servers answer with `X-Airon-Protocol: 1`. A server that cannot speak the client's version answers `426`.
|
|
26
|
+
|
|
27
|
+
## GET /v1/ping
|
|
28
|
+
|
|
29
|
+
Used by `airon doctor` and `airon login` to check the address and the key.
|
|
30
|
+
|
|
31
|
+
```json
|
|
32
|
+
{"ok": true, "protocol": 1, "key": {"name": "pod-h100", "scope": "tracker"}, "server_time": 1730402880.1}
|
|
33
|
+
```
|
|
34
|
+
|
|
35
|
+
## POST /v1/runs: create or reopen a run
|
|
36
|
+
|
|
37
|
+
Idempotent by `run_id`, so a retried request is harmless. The client generates `run_id` (a UUID4 string).
|
|
38
|
+
|
|
39
|
+
```json
|
|
40
|
+
{
|
|
41
|
+
"run_id": "6f1c1c0e-3f55-4d66-9a8e-2f0f0d8f0b11",
|
|
42
|
+
"project": "memory-layer",
|
|
43
|
+
"name": "baseline-v2",
|
|
44
|
+
"experiment": "Baseline v2",
|
|
45
|
+
"config": {"lr": 3e-5, "batch_size": 32, "model": {"name": "small-encoder"}},
|
|
46
|
+
"env": {
|
|
47
|
+
"python": "3.11.6",
|
|
48
|
+
"platform": "Linux-5.15-x86_64",
|
|
49
|
+
"host": "pod-8c1f",
|
|
50
|
+
"cloud": "runpod",
|
|
51
|
+
"gpus": [{"index": 0, "name": "NVIDIA H100 80GB HBM3", "mem_total_mb": 81559}],
|
|
52
|
+
"git": {"commit": "5d1e0aa", "branch": "main", "dirty": false, "remote": "git@example.com:acme/model-code.git"},
|
|
53
|
+
"libs": {"torch": "2.5.0", "numpy": "2.1.0"}
|
|
54
|
+
},
|
|
55
|
+
"total_steps": 50000,
|
|
56
|
+
"total_epochs": 5,
|
|
57
|
+
"tags": ["baseline"],
|
|
58
|
+
"notes": "",
|
|
59
|
+
"resume": "never",
|
|
60
|
+
"started_at": 1730398080.0
|
|
61
|
+
}
|
|
62
|
+
```
|
|
63
|
+
|
|
64
|
+
Only `run_id` and `project` are required. Everything else is optional. `name` is generated by the server
|
|
65
|
+
when missing. `experiment` is the title of an Airon experiment to link the run to; if there is none, the
|
|
66
|
+
server ignores it and says so in `warnings`. `resume` is `"never"` (default), `"allow"` or `"must"`.
|
|
67
|
+
|
|
68
|
+
| Situation | Result |
|
|
69
|
+
|---|---|
|
|
70
|
+
| `run_id` is new, `resume` is not `"must"` | `201`, a new run |
|
|
71
|
+
| `run_id` exists and the run is still running | `200`, an idempotent update (covers retries) |
|
|
72
|
+
| `run_id` exists and is finished, failed, killed or crashed, `resume` is `"allow"` or `"must"` | `200`, reopened, `resumed: true` |
|
|
73
|
+
| `run_id` exists and is over, `resume` is `"never"` | `409` |
|
|
74
|
+
| `run_id` is new and `resume` is `"must"` | `404` |
|
|
75
|
+
|
|
76
|
+
```json
|
|
77
|
+
{"run_id": "6f1c…", "name": "baseline-v2", "url": "https://airon.example.com/runs/6f1c…",
|
|
78
|
+
"status": "running", "resumed": false, "last_step": null, "last_seq": 0, "warnings": [],
|
|
79
|
+
"server_time": 1730402880.2}
|
|
80
|
+
```
|
|
81
|
+
|
|
82
|
+
`last_step` and `last_seq` let a resumed run continue where the server stopped: the client starts `seq`
|
|
83
|
+
after `last_seq`.
|
|
84
|
+
|
|
85
|
+
## POST /v1/runs/{run_id}/events: send a batch
|
|
86
|
+
|
|
87
|
+
```json
|
|
88
|
+
{"events": [
|
|
89
|
+
{"seq": 41, "t": 1730402880.0, "step": 12400, "epoch": 1.98,
|
|
90
|
+
"metrics": {"train/loss": 0.412, "lr": 2.4e-5}},
|
|
91
|
+
{"seq": 42, "t": 1730402880.0, "system": {"gpu0/util_pct": 96, "gpu0/mem_used_mb": 71234, "cpu/util_pct": 38}},
|
|
92
|
+
{"seq": 43, "t": 1730402881.0,
|
|
93
|
+
"progress": {"step": 12400, "total_steps": 50000, "epoch": 1.98, "total_epochs": 5,
|
|
94
|
+
"steps_per_sec": 2.4, "samples_per_sec": 76.8, "tokens_per_sec": 18200}}
|
|
95
|
+
]}
|
|
96
|
+
```
|
|
97
|
+
|
|
98
|
+
At most 500 events and 1 MiB (uncompressed) per request.
|
|
99
|
+
|
|
100
|
+
Every event has `seq` (integer, increasing by 1 per run, assigned by the client) and `t`. The server
|
|
101
|
+
ignores an event whose `seq` it already has, which makes retries safe. An event can carry any of these
|
|
102
|
+
keys, several at once:
|
|
103
|
+
|
|
104
|
+
| Key | Content |
|
|
105
|
+
|---|---|
|
|
106
|
+
| `metrics` | `{name: number}` at an optional `step` (integer) and `epoch` (float, may be fractional) |
|
|
107
|
+
| `system` | `{name: number}` hardware samples, kept off the step axis (see well-known names) |
|
|
108
|
+
| `progress` | `step`, `total_steps`, `epoch`, `total_epochs`, `steps_per_sec`, `samples_per_sec`, `tokens_per_sec`, all optional |
|
|
109
|
+
| `log` | `[{"stream": "stdout" or "stderr", "text": "one line"}]` console lines |
|
|
110
|
+
| `checkpoint` | `{"uri": "hf://org/repo@rev", "step": 12000, "label": "best", "metrics": {}}` a pointer, never the file |
|
|
111
|
+
| `summary` | `{name: any JSON}` values set explicitly by the user |
|
|
112
|
+
| `config` | `{...}` merged into the run's config, key by key |
|
|
113
|
+
|
|
114
|
+
An event with only `seq` and `t` is a heartbeat. The server derives these without help: the ETA (from
|
|
115
|
+
the recent progress rate, so pauses for evaluation and checkpoints are included), the best and last
|
|
116
|
+
value of each metric, and the `crashed` state (no event for 5 minutes while still `running`).
|
|
117
|
+
|
|
118
|
+
Response `200`, even when some events were refused, so one bad event cannot block the queue:
|
|
119
|
+
|
|
120
|
+
```json
|
|
121
|
+
{"accepted": 3, "duplicates": 0, "last_seq": 43, "rejected": []}
|
|
122
|
+
```
|
|
123
|
+
|
|
124
|
+
A rejected event looks like `{"seq": 44, "reason": "metrics.train/loss is not a number"}`. The client
|
|
125
|
+
must not resend rejected events.
|
|
126
|
+
|
|
127
|
+
Errors: `401`, `403`, `404` unknown run, `413` body too large (the client splits the batch and retries),
|
|
128
|
+
`422` malformed request (not retried), `429` with `Retry-After` (seconds), `5xx`.
|
|
129
|
+
|
|
130
|
+
## POST /v1/runs/{run_id}/finish
|
|
131
|
+
|
|
132
|
+
```json
|
|
133
|
+
{"status": "failed", "t": 1730412345.0, "exit_code": 1,
|
|
134
|
+
"error": {"type": "torch.OutOfMemoryError", "message": "CUDA out of memory…", "traceback": "Traceback (most recent call last):\n…"},
|
|
135
|
+
"summary": {"eval/loss.best": 0.431}}
|
|
136
|
+
```
|
|
137
|
+
|
|
138
|
+
`status` is `"finished"`, `"failed"` or `"killed"` (for example a terminated pod). Idempotent. Events that
|
|
139
|
+
arrive after it are still recorded; the status does not change. `"crashed"` is never sent by a client:
|
|
140
|
+
the server sets it when a running run goes silent.
|
|
141
|
+
|
|
142
|
+
## Well-known names
|
|
143
|
+
|
|
144
|
+
These get special treatment in the interface. Everything else is shown too.
|
|
145
|
+
|
|
146
|
+
- Metrics: `train/loss`, `eval/loss`, `eval/*`, `train/*`, `lr` (or `train/lr`), `train/grad_norm`.
|
|
147
|
+
- System: `gpu{i}/util_pct`, `gpu{i}/mem_used_mb`, `gpu{i}/mem_total_mb`, `gpu{i}/temp_c`,
|
|
148
|
+
`gpu{i}/power_w`, `cpu/util_pct`, `ram/used_mb`, `ram/total_mb`, `disk/used_gb`, `disk/total_gb`.
|
|
149
|
+
|
|
150
|
+
## Client behaviour
|
|
151
|
+
|
|
152
|
+
The SDK must follow these rules. They are what lets a team adopt it without worrying about their training.
|
|
153
|
+
|
|
154
|
+
1. **Never break training.** `log()` only appends to an in-memory queue and returns. Nothing in the
|
|
155
|
+
SDK raises into user code because of the network, the server or the disk. At most one warning is
|
|
156
|
+
printed per kind of problem.
|
|
157
|
+
2. **Bounded memory.** The queue holds at most 100,000 events. When full, drop the oldest `system`
|
|
158
|
+
events first, then the oldest `metrics`, and warn once.
|
|
159
|
+
3. **Batching.** A background thread flushes every 2 seconds or 200 events, whichever comes first,
|
|
160
|
+
sends a heartbeat every 30 seconds, and samples hardware every 10 seconds. All configurable.
|
|
161
|
+
4. **Retries.** Exponential backoff from 1 s to 60 s with jitter, honouring `Retry-After`. Retry on
|
|
162
|
+
network errors, `408`, `425`, `429` and `5xx`. Do not retry other `4xx`: drop the batch and warn. On
|
|
163
|
+
`401` or `403` stop sending, keep everything in the offline spool, and warn once.
|
|
164
|
+
5. **Offline spool.** Events that cannot be delivered are written to JSON lines under `$AIRON_DIR`
|
|
165
|
+
(default `~/.cache/airon`), one directory per run. `airon sync` uploads what is there, in order.
|
|
166
|
+
`AIRON_MODE=offline` writes only to the spool; `AIRON_MODE=disabled` makes every call a no-op.
|
|
167
|
+
6. **Exit.** Flush on normal exit and on an unhandled exception, waiting at most 10 seconds. An unhandled
|
|
168
|
+
exception inside `with airon.init(...)` is reported as `failed` with its traceback and re-raised.
|
|
169
|
+
`SIGTERM` is reported as `killed` after the program's own handler has run: chain existing handlers,
|
|
170
|
+
never replace them.
|
|
171
|
+
7. **Console.** Capture the tail of stdout and stderr (default 200 lines) without breaking progress bars
|
|
172
|
+
that use carriage returns. It can be turned off with `console=False`.
|
|
173
|
+
|
|
174
|
+
## Environment variables
|
|
175
|
+
|
|
176
|
+
| Variable | Meaning |
|
|
177
|
+
|---|---|
|
|
178
|
+
| `AIRON_URL` | Server address, for example `https://airon.example.com` |
|
|
179
|
+
| `AIRON_API_KEY` | The key |
|
|
180
|
+
| `AIRON_PROJECT` | Default project |
|
|
181
|
+
| `AIRON_RUN_ID` | Run to resume or attach to |
|
|
182
|
+
| `AIRON_MODE` | `online` (default), `offline` or `disabled` |
|
|
183
|
+
| `AIRON_DIR` | Where the spool and credentials live |
|
|
184
|
+
|
|
185
|
+
`airon login` stores the address and the key in `~/.config/airon/credentials.json`, readable only by the
|
|
186
|
+
owner. Environment variables take precedence over the file.
|
|
@@ -0,0 +1,37 @@
|
|
|
1
|
+
# Releasing
|
|
2
|
+
|
|
3
|
+
Releases are built and published by GitHub Actions when a tag `v*` is pushed. Publishing uses PyPI
|
|
4
|
+
[trusted publishing](https://docs.pypi.org/trusted-publishers/), so no token is stored anywhere.
|
|
5
|
+
|
|
6
|
+
## One-time setup (the repository owner)
|
|
7
|
+
|
|
8
|
+
1. On PyPI, create the project `airon`, or add a *pending publisher* if it does not exist yet:
|
|
9
|
+
**Your account > Publishing > Add a new pending publisher**.
|
|
10
|
+
2. Fill in:
|
|
11
|
+
- PyPI project name: `airon`
|
|
12
|
+
- Owner: `Manifestro`
|
|
13
|
+
- Repository name: `airon-metrics`
|
|
14
|
+
- Workflow name: `release.yml`
|
|
15
|
+
- Environment name: `pypi`
|
|
16
|
+
3. In the GitHub repository, **Settings > Environments > New environment** named `pypi`. Add
|
|
17
|
+
*required reviewers* if every release should need an approval.
|
|
18
|
+
|
|
19
|
+
To rehearse on TestPyPI, repeat step 1 and 2 on test.pypi.org and temporarily point the workflow's
|
|
20
|
+
`repository-url` at `https://test.pypi.org/legacy/`.
|
|
21
|
+
|
|
22
|
+
## Making a release
|
|
23
|
+
|
|
24
|
+
1. Update `src/airon/_version.py` and `CHANGELOG.md` (move *unreleased* to the new version and date).
|
|
25
|
+
2. Commit and merge to `main`; wait for CI to pass.
|
|
26
|
+
3. Tag and push:
|
|
27
|
+
|
|
28
|
+
```bash
|
|
29
|
+
git tag v0.1.0
|
|
30
|
+
git push origin v0.1.0
|
|
31
|
+
```
|
|
32
|
+
|
|
33
|
+
4. The `Release` workflow builds the sdist and wheel, checks the tag matches the package version,
|
|
34
|
+
runs `twine check`, and publishes (after approval, if you set required reviewers).
|
|
35
|
+
5. Verify in a clean environment: `pip install airon==0.1.0 && airon version`.
|
|
36
|
+
|
|
37
|
+
A mistake in a published version cannot be replaced: yank it on PyPI and release a new patch version.
|