triotune 0.1.1__py3-none-any.whl
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.
- trio_cookbook/__init__.py +26 -0
- trio_cookbook/api/__init__.py +55 -0
- trio_cookbook/api/cli_args.py +1061 -0
- trio_cookbook/api/client_api.py +215 -0
- trio_cookbook/api/loss_api.py +743 -0
- trio_cookbook/dataset/__init__.py +44 -0
- trio_cookbook/dataset/examples.py +181 -0
- trio_cookbook/dataset/load.py +65 -0
- trio_cookbook/dataset/prepare.py +451 -0
- trio_cookbook/dataset/prompt.py +243 -0
- trio_cookbook/dataset/swe_sft/__init__.py +18 -0
- trio_cookbook/dataset/swe_sft/prepare.py +202 -0
- trio_cookbook/dataset/swe_sft/schema.py +187 -0
- trio_cookbook/dataset/swe_sft/trajectory.py +503 -0
- trio_cookbook/trainer/__init__.py +180 -0
- trio_cookbook/trainer/dpo/__init__.py +7 -0
- trio_cookbook/trainer/dpo/loss.py +90 -0
- trio_cookbook/trainer/dpo/state.py +22 -0
- trio_cookbook/trainer/dpo_trainer.py +755 -0
- trio_cookbook/trainer/opd/__init__.py +77 -0
- trio_cookbook/trainer/opd/contracts.py +287 -0
- trio_cookbook/trainer/opd/datum.py +65 -0
- trio_cookbook/trainer/opd/execution.py +219 -0
- trio_cookbook/trainer/opd/loss.py +144 -0
- trio_cookbook/trainer/opd/metrics.py +180 -0
- trio_cookbook/trainer/opd/rollout.py +128 -0
- trio_cookbook/trainer/opd/targets.py +64 -0
- trio_cookbook/trainer/opd/teacher_phase.py +98 -0
- trio_cookbook/trainer/opd_trainer.py +590 -0
- trio_cookbook/trainer/ppo/__init__.py +77 -0
- trio_cookbook/trainer/ppo/actor.py +155 -0
- trio_cookbook/trainer/ppo/advantages.py +184 -0
- trio_cookbook/trainer/ppo/contracts.py +116 -0
- trio_cookbook/trainer/ppo/critic.py +315 -0
- trio_cookbook/trainer/ppo/datum.py +99 -0
- trio_cookbook/trainer/ppo/gae.py +192 -0
- trio_cookbook/trainer/ppo/loss.py +159 -0
- trio_cookbook/trainer/ppo/rollout.py +146 -0
- trio_cookbook/trainer/rl_trainer.py +423 -0
- trio_cookbook/trainer/sft/__init__.py +7 -0
- trio_cookbook/trainer/sft/state.py +12 -0
- trio_cookbook/trainer/sft_trainer.py +343 -0
- trio_cookbook/utils/__init__.py +55 -0
- trio_cookbook/utils/checkpoint.py +32 -0
- trio_cookbook/utils/config.py +78 -0
- trio_cookbook/utils/dataloader.py +59 -0
- trio_cookbook/utils/datums.py +174 -0
- trio_cookbook/utils/metrics.py +211 -0
- trio_cookbook/utils/reporting.py +145 -0
- trio_cookbook/utils/scoring.py +236 -0
- trio_cookbook/utils/updates.py +120 -0
- triotune-0.1.1.dist-info/METADATA +213 -0
- triotune-0.1.1.dist-info/RECORD +55 -0
- triotune-0.1.1.dist-info/WHEEL +5 -0
- triotune-0.1.1.dist-info/top_level.txt +1 -0
|
@@ -0,0 +1,26 @@
|
|
|
1
|
+
"""Client-side training utilities for PyTRIO.
|
|
2
|
+
|
|
3
|
+
The package intentionally contains no model implementation. Model weights,
|
|
4
|
+
tokenization backends, sampling and optimization stay behind the PyTRIO
|
|
5
|
+
clients supplied by the caller.
|
|
6
|
+
"""
|
|
7
|
+
|
|
8
|
+
from . import utils
|
|
9
|
+
from .api.client_api import call, call_async, wait, wait_async
|
|
10
|
+
from .trainer import DPOConfig, DPOTrainer, OPDConfig, OPDTrainer
|
|
11
|
+
from .trainer import PPOConfig, PPOTrainer, RenderedPrompt, render_prompt
|
|
12
|
+
|
|
13
|
+
__all__ = [
|
|
14
|
+
"DPOConfig",
|
|
15
|
+
"DPOTrainer",
|
|
16
|
+
"OPDConfig",
|
|
17
|
+
"OPDTrainer",
|
|
18
|
+
"PPOConfig",
|
|
19
|
+
"PPOTrainer",
|
|
20
|
+
"RenderedPrompt",
|
|
21
|
+
"call",
|
|
22
|
+
"call_async",
|
|
23
|
+
"render_prompt",
|
|
24
|
+
"wait",
|
|
25
|
+
"wait_async",
|
|
26
|
+
]
|
|
@@ -0,0 +1,55 @@
|
|
|
1
|
+
"""Public configuration and client contracts for trainer algorithms."""
|
|
2
|
+
|
|
3
|
+
from .cli_args import (
|
|
4
|
+
BaseTrainerConfig,
|
|
5
|
+
DPOConfig,
|
|
6
|
+
NormConfig,
|
|
7
|
+
OPDConfig,
|
|
8
|
+
PPOConfig,
|
|
9
|
+
SFTConfig,
|
|
10
|
+
)
|
|
11
|
+
from .client_api import (
|
|
12
|
+
call,
|
|
13
|
+
call_async,
|
|
14
|
+
optimizer_params,
|
|
15
|
+
sampling_params,
|
|
16
|
+
sdk_object,
|
|
17
|
+
submit_async,
|
|
18
|
+
wait,
|
|
19
|
+
wait_async,
|
|
20
|
+
)
|
|
21
|
+
from .loss_api import (
|
|
22
|
+
BUILTIN_LOSS_FNS,
|
|
23
|
+
CE_LOSS_FNS,
|
|
24
|
+
POLICY_LOSS_FNS,
|
|
25
|
+
BuiltinObjective,
|
|
26
|
+
BuiltinPyTrioObjective,
|
|
27
|
+
LossSpec,
|
|
28
|
+
PolicyLossInput,
|
|
29
|
+
SupervisedLossInput,
|
|
30
|
+
)
|
|
31
|
+
|
|
32
|
+
__all__ = [
|
|
33
|
+
"BUILTIN_LOSS_FNS",
|
|
34
|
+
"CE_LOSS_FNS",
|
|
35
|
+
"POLICY_LOSS_FNS",
|
|
36
|
+
"BaseTrainerConfig",
|
|
37
|
+
"BuiltinObjective",
|
|
38
|
+
"BuiltinPyTrioObjective",
|
|
39
|
+
"DPOConfig",
|
|
40
|
+
"LossSpec",
|
|
41
|
+
"NormConfig",
|
|
42
|
+
"OPDConfig",
|
|
43
|
+
"PPOConfig",
|
|
44
|
+
"PolicyLossInput",
|
|
45
|
+
"SFTConfig",
|
|
46
|
+
"SupervisedLossInput",
|
|
47
|
+
"call",
|
|
48
|
+
"call_async",
|
|
49
|
+
"optimizer_params",
|
|
50
|
+
"sampling_params",
|
|
51
|
+
"sdk_object",
|
|
52
|
+
"submit_async",
|
|
53
|
+
"wait",
|
|
54
|
+
"wait_async",
|
|
55
|
+
]
|