dromeus 0.4.0.dev1__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.
- dromeus/__init__.py +109 -0
- dromeus/__main__.py +6 -0
- dromeus/algorithms/__init__.py +31 -0
- dromeus/algorithms/base.py +188 -0
- dromeus/algorithms/codec.py +623 -0
- dromeus/algorithms/noloco.py +486 -0
- dromeus/application.py +523 -0
- dromeus/cli.py +341 -0
- dromeus/diagnostics.py +303 -0
- dromeus/errors.py +57 -0
- dromeus/gossip/__init__.py +29 -0
- dromeus/gossip/_async.py +15 -0
- dromeus/gossip/axl.py +497 -0
- dromeus/gossip/engine.py +378 -0
- dromeus/gossip/interfaces.py +158 -0
- dromeus/gossip/peer_scheduler.py +117 -0
- dromeus/manifests/__init__.py +1 -0
- dromeus/manifests/canonical.py +157 -0
- dromeus/manifests/models.py +694 -0
- dromeus/membership/__init__.py +1 -0
- dromeus/membership/formation.py +576 -0
- dromeus/node.py +338 -0
- dromeus/persistence/__init__.py +20 -0
- dromeus/persistence/archive.py +576 -0
- dromeus/persistence/files.py +39 -0
- dromeus/persistence/privacy_ledger.py +104 -0
- dromeus/persistence/run_store.py +352 -0
- dromeus/protocol/__init__.py +25 -0
- dromeus/protocol/codec.py +127 -0
- dromeus/protocol/models.py +199 -0
- dromeus/protocol/version.py +5 -0
- dromeus/py.typed +1 -0
- dromeus/runtime.py +1081 -0
- dromeus/telemetry/__init__.py +61 -0
- dromeus/telemetry/consensus.py +909 -0
- dromeus/telemetry/divergence.py +321 -0
- dromeus/telemetry/divergence_monitor.py +97 -0
- dromeus/telemetry/events.py +214 -0
- dromeus/telemetry/evidence.py +429 -0
- dromeus/telemetry/metrics.py +200 -0
- dromeus/telemetry/progress.py +255 -0
- dromeus/training/__init__.py +1 -0
- dromeus/training/base.py +60 -0
- dromeus/training/model_state.py +111 -0
- dromeus/training/native_gradients.py +289 -0
- dromeus/training/optimizer_factory.py +426 -0
- dromeus/training/ordinary_task.py +408 -0
- dromeus/training/privacy.py +96 -0
- dromeus/training/private_data.py +96 -0
- dromeus/training/private_model.py +414 -0
- dromeus/training/private_state.py +97 -0
- dromeus/training/private_trainer.py +534 -0
- dromeus/training/private_workload.py +217 -0
- dromeus/training/state.py +57 -0
- dromeus/training/task.py +464 -0
- dromeus/training/trainer.py +294 -0
- dromeus/training/versions.py +15 -0
- dromeus/transport/__init__.py +1 -0
- dromeus/transport/axl.py +168 -0
- dromeus/transport/envelope_sender.py +73 -0
- dromeus/transport/interface.py +31 -0
- dromeus/transport/outbound_scheduler.py +326 -0
- dromeus/transport/receiver.py +345 -0
- dromeus/transport/transfer.py +854 -0
- dromeus-0.4.0.dev1.dist-info/METADATA +150 -0
- dromeus-0.4.0.dev1.dist-info/RECORD +69 -0
- dromeus-0.4.0.dev1.dist-info/WHEEL +4 -0
- dromeus-0.4.0.dev1.dist-info/entry_points.txt +2 -0
- dromeus-0.4.0.dev1.dist-info/licenses/LICENSE +21 -0
dromeus/__init__.py
ADDED
|
@@ -0,0 +1,109 @@
|
|
|
1
|
+
"""Decentralized federated learning over Gensyn AXL.
|
|
2
|
+
|
|
3
|
+
Every node trains the same `TrainingTask` on its own records and exchanges NoLoCo
|
|
4
|
+
updates with one peer per round. Typical use:
|
|
5
|
+
|
|
6
|
+
1. Define a `TrainingTask` that every node builds identically.
|
|
7
|
+
2. One person writes the shared draft with `make_draft(task, "draft.yaml", ...)`.
|
|
8
|
+
3. Each node runs `await Node.from_file("node.yaml").run(task, local=LocalData(...))`.
|
|
9
|
+
|
|
10
|
+
Only the names exported here are the public, stable API.
|
|
11
|
+
"""
|
|
12
|
+
|
|
13
|
+
from __future__ import annotations
|
|
14
|
+
|
|
15
|
+
from importlib import import_module
|
|
16
|
+
from importlib.metadata import version as _version
|
|
17
|
+
from typing import TYPE_CHECKING, Any
|
|
18
|
+
|
|
19
|
+
if TYPE_CHECKING:
|
|
20
|
+
from dromeus.application import (
|
|
21
|
+
LocalData,
|
|
22
|
+
NoLoCo,
|
|
23
|
+
Privacy,
|
|
24
|
+
check_draft,
|
|
25
|
+
code_hash,
|
|
26
|
+
make_draft,
|
|
27
|
+
)
|
|
28
|
+
from dromeus.errors import (
|
|
29
|
+
ConfigError,
|
|
30
|
+
DraftError,
|
|
31
|
+
DromeusError,
|
|
32
|
+
FormationError,
|
|
33
|
+
PrivacyError,
|
|
34
|
+
RunError,
|
|
35
|
+
TaskError,
|
|
36
|
+
)
|
|
37
|
+
from dromeus.manifests.models import DivergencePolicy, TransportLimits
|
|
38
|
+
from dromeus.node import Node, RunResult
|
|
39
|
+
from dromeus.persistence.privacy_ledger import PrivacyLedger
|
|
40
|
+
from dromeus.training.optimizer_factory import OptimizerRegistration
|
|
41
|
+
from dromeus.training.task import Optimizer, TensorBatchAdapter, TrainingTask
|
|
42
|
+
|
|
43
|
+
__version__ = _version("dromeus")
|
|
44
|
+
|
|
45
|
+
# Imported on first use, so `import dromeus` stays light (no PyTorch until needed).
|
|
46
|
+
_EXPORTS = {
|
|
47
|
+
"LocalData": "dromeus.application",
|
|
48
|
+
"NoLoCo": "dromeus.application",
|
|
49
|
+
"Privacy": "dromeus.application",
|
|
50
|
+
"check_draft": "dromeus.application",
|
|
51
|
+
"code_hash": "dromeus.application",
|
|
52
|
+
"make_draft": "dromeus.application",
|
|
53
|
+
"ConfigError": "dromeus.errors",
|
|
54
|
+
"DraftError": "dromeus.errors",
|
|
55
|
+
"DromeusError": "dromeus.errors",
|
|
56
|
+
"FormationError": "dromeus.errors",
|
|
57
|
+
"PrivacyError": "dromeus.errors",
|
|
58
|
+
"RunError": "dromeus.errors",
|
|
59
|
+
"TaskError": "dromeus.errors",
|
|
60
|
+
"DivergencePolicy": "dromeus.manifests.models",
|
|
61
|
+
"TransportLimits": "dromeus.manifests.models",
|
|
62
|
+
"Node": "dromeus.node",
|
|
63
|
+
"RunResult": "dromeus.node",
|
|
64
|
+
"PrivacyLedger": "dromeus.persistence.privacy_ledger",
|
|
65
|
+
"OptimizerRegistration": "dromeus.training.optimizer_factory",
|
|
66
|
+
"Optimizer": "dromeus.training.task",
|
|
67
|
+
"TensorBatchAdapter": "dromeus.training.task",
|
|
68
|
+
"TrainingTask": "dromeus.training.task",
|
|
69
|
+
}
|
|
70
|
+
|
|
71
|
+
|
|
72
|
+
def __getattr__(name: str) -> Any:
|
|
73
|
+
module = _EXPORTS.get(name)
|
|
74
|
+
if module is None:
|
|
75
|
+
raise AttributeError(f"module 'dromeus' has no attribute {name!r}")
|
|
76
|
+
value = getattr(import_module(module), name)
|
|
77
|
+
globals()[name] = value
|
|
78
|
+
return value
|
|
79
|
+
|
|
80
|
+
|
|
81
|
+
def __dir__() -> list[str]:
|
|
82
|
+
return sorted(__all__)
|
|
83
|
+
|
|
84
|
+
|
|
85
|
+
__all__ = [
|
|
86
|
+
"ConfigError",
|
|
87
|
+
"DivergencePolicy",
|
|
88
|
+
"DraftError",
|
|
89
|
+
"DromeusError",
|
|
90
|
+
"FormationError",
|
|
91
|
+
"LocalData",
|
|
92
|
+
"NoLoCo",
|
|
93
|
+
"Node",
|
|
94
|
+
"Optimizer",
|
|
95
|
+
"OptimizerRegistration",
|
|
96
|
+
"Privacy",
|
|
97
|
+
"PrivacyError",
|
|
98
|
+
"PrivacyLedger",
|
|
99
|
+
"RunError",
|
|
100
|
+
"RunResult",
|
|
101
|
+
"TaskError",
|
|
102
|
+
"TensorBatchAdapter",
|
|
103
|
+
"TrainingTask",
|
|
104
|
+
"TransportLimits",
|
|
105
|
+
"__version__",
|
|
106
|
+
"check_draft",
|
|
107
|
+
"code_hash",
|
|
108
|
+
"make_draft",
|
|
109
|
+
]
|
dromeus/__main__.py
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
1
|
+
"""Training algorithm adapters."""
|
|
2
|
+
|
|
3
|
+
from dromeus.algorithms.base import (
|
|
4
|
+
AlgorithmObservations,
|
|
5
|
+
AlgorithmSnapshot,
|
|
6
|
+
MaterializedArtifact,
|
|
7
|
+
SerializableState,
|
|
8
|
+
UpdateBundle,
|
|
9
|
+
ValidatedUpdate,
|
|
10
|
+
)
|
|
11
|
+
from dromeus.algorithms.codec import (
|
|
12
|
+
IdentityCodec,
|
|
13
|
+
NamedSafetensorsUpdateBundleCodec,
|
|
14
|
+
NamedUpdateBundleCodec,
|
|
15
|
+
UpdateCodec,
|
|
16
|
+
)
|
|
17
|
+
from dromeus.manifests.models import UpdateCodecBinding
|
|
18
|
+
|
|
19
|
+
__all__ = [
|
|
20
|
+
"AlgorithmObservations",
|
|
21
|
+
"AlgorithmSnapshot",
|
|
22
|
+
"IdentityCodec",
|
|
23
|
+
"MaterializedArtifact",
|
|
24
|
+
"NamedSafetensorsUpdateBundleCodec",
|
|
25
|
+
"NamedUpdateBundleCodec",
|
|
26
|
+
"SerializableState",
|
|
27
|
+
"UpdateBundle",
|
|
28
|
+
"UpdateCodec",
|
|
29
|
+
"UpdateCodecBinding",
|
|
30
|
+
"ValidatedUpdate",
|
|
31
|
+
]
|
|
@@ -0,0 +1,188 @@
|
|
|
1
|
+
"""Training algorithm interface."""
|
|
2
|
+
|
|
3
|
+
from __future__ import annotations
|
|
4
|
+
|
|
5
|
+
import hashlib
|
|
6
|
+
import math
|
|
7
|
+
from collections.abc import Mapping
|
|
8
|
+
from dataclasses import dataclass
|
|
9
|
+
from pathlib import Path
|
|
10
|
+
from types import MappingProxyType
|
|
11
|
+
from typing import Protocol
|
|
12
|
+
|
|
13
|
+
import numpy as np
|
|
14
|
+
|
|
15
|
+
from dromeus.manifests.canonical import file_sha256, update_bundle_digest
|
|
16
|
+
from dromeus.manifests.models import (
|
|
17
|
+
Identifier,
|
|
18
|
+
OpaqueUpdateBundleMetadata,
|
|
19
|
+
RoundId,
|
|
20
|
+
TensorSchema,
|
|
21
|
+
)
|
|
22
|
+
|
|
23
|
+
|
|
24
|
+
class SerializableState(Protocol):
|
|
25
|
+
"""State seam shared by algorithm and codec implementations."""
|
|
26
|
+
|
|
27
|
+
def state_dict(self) -> dict[str, object]: ...
|
|
28
|
+
|
|
29
|
+
def load_state_dict(self, state: Mapping[str, object]) -> None: ...
|
|
30
|
+
|
|
31
|
+
|
|
32
|
+
@dataclass(frozen=True, slots=True)
|
|
33
|
+
class AlgorithmObservations:
|
|
34
|
+
"""Immutable algorithm observations captured for one round."""
|
|
35
|
+
|
|
36
|
+
local_loss: float | None = None
|
|
37
|
+
error_feedback_residual_l2_norm: float | None = None
|
|
38
|
+
error_feedback_signal_l2_norm: float | None = None
|
|
39
|
+
error_feedback_residual_to_signal_ratio: float | None = None
|
|
40
|
+
|
|
41
|
+
def __post_init__(self) -> None:
|
|
42
|
+
if self.local_loss is not None and not math.isfinite(self.local_loss):
|
|
43
|
+
raise ValueError("local loss must be finite")
|
|
44
|
+
values = (
|
|
45
|
+
("error-feedback residual norm", self.error_feedback_residual_l2_norm),
|
|
46
|
+
("error-feedback signal norm", self.error_feedback_signal_l2_norm),
|
|
47
|
+
(
|
|
48
|
+
"error-feedback residual-to-signal ratio",
|
|
49
|
+
self.error_feedback_residual_to_signal_ratio,
|
|
50
|
+
),
|
|
51
|
+
)
|
|
52
|
+
for label, value in values:
|
|
53
|
+
if value is not None and (not math.isfinite(value) or value < 0):
|
|
54
|
+
raise ValueError(f"{label} must be finite and non-negative")
|
|
55
|
+
|
|
56
|
+
|
|
57
|
+
def checksum_tensors(tensors: Mapping[str, np.ndarray]) -> str:
|
|
58
|
+
"""Hash named tensor values independently of the algorithm implementation."""
|
|
59
|
+
digest = hashlib.sha256()
|
|
60
|
+
for name in sorted(tensors):
|
|
61
|
+
tensor = np.ascontiguousarray(tensors[name])
|
|
62
|
+
digest.update(name.encode())
|
|
63
|
+
digest.update(str(tensor.dtype).encode())
|
|
64
|
+
digest.update(str(tensor.shape).encode())
|
|
65
|
+
digest.update(tensor.tobytes())
|
|
66
|
+
return digest.hexdigest()
|
|
67
|
+
|
|
68
|
+
|
|
69
|
+
def _immutable_tensors(
|
|
70
|
+
tensors: Mapping[str, np.ndarray],
|
|
71
|
+
) -> Mapping[str, np.ndarray]:
|
|
72
|
+
copied: dict[str, np.ndarray] = {}
|
|
73
|
+
for name, value in tensors.items():
|
|
74
|
+
tensor = np.ascontiguousarray(value).copy()
|
|
75
|
+
tensor.flags.writeable = False
|
|
76
|
+
copied[name] = tensor
|
|
77
|
+
return MappingProxyType(copied)
|
|
78
|
+
|
|
79
|
+
|
|
80
|
+
@dataclass(frozen=True, slots=True)
|
|
81
|
+
class AlgorithmSnapshot:
|
|
82
|
+
round_id: RoundId
|
|
83
|
+
phase: str
|
|
84
|
+
weights: Mapping[str, np.ndarray]
|
|
85
|
+
|
|
86
|
+
def __post_init__(self) -> None:
|
|
87
|
+
object.__setattr__(self, "weights", _immutable_tensors(self.weights))
|
|
88
|
+
|
|
89
|
+
|
|
90
|
+
@dataclass(frozen=True, slots=True)
|
|
91
|
+
class ValidatedUpdate:
|
|
92
|
+
"""Algorithm-owned peer update ready to apply."""
|
|
93
|
+
|
|
94
|
+
round_id: RoundId
|
|
95
|
+
tensors: Mapping[str, np.ndarray]
|
|
96
|
+
checksum: str
|
|
97
|
+
|
|
98
|
+
def __post_init__(self) -> None:
|
|
99
|
+
object.__setattr__(self, "tensors", _immutable_tensors(self.tensors))
|
|
100
|
+
|
|
101
|
+
|
|
102
|
+
@dataclass(frozen=True, slots=True)
|
|
103
|
+
class NamedValidatedUpdate:
|
|
104
|
+
"""Algorithm-owned named peer artifacts ready to apply."""
|
|
105
|
+
|
|
106
|
+
round_id: RoundId
|
|
107
|
+
artifacts: Mapping[str, Mapping[str, np.ndarray]]
|
|
108
|
+
checksum: str
|
|
109
|
+
|
|
110
|
+
def __post_init__(self) -> None:
|
|
111
|
+
object.__setattr__(
|
|
112
|
+
self,
|
|
113
|
+
"artifacts",
|
|
114
|
+
MappingProxyType(
|
|
115
|
+
{
|
|
116
|
+
name: _immutable_tensors(tensors)
|
|
117
|
+
for name, tensors in self.artifacts.items()
|
|
118
|
+
}
|
|
119
|
+
),
|
|
120
|
+
)
|
|
121
|
+
|
|
122
|
+
|
|
123
|
+
AlgorithmUpdate = ValidatedUpdate | NamedValidatedUpdate
|
|
124
|
+
|
|
125
|
+
|
|
126
|
+
def checksum_artifacts(
|
|
127
|
+
artifacts: Mapping[str, Mapping[str, np.ndarray]],
|
|
128
|
+
) -> str:
|
|
129
|
+
"""Hash named tensor artifacts without flattening their public shape."""
|
|
130
|
+
return checksum_tensors(
|
|
131
|
+
{
|
|
132
|
+
f"{artifact_name}.{tensor_name}": tensor
|
|
133
|
+
for artifact_name, tensors in artifacts.items()
|
|
134
|
+
for tensor_name, tensor in tensors.items()
|
|
135
|
+
}
|
|
136
|
+
)
|
|
137
|
+
|
|
138
|
+
|
|
139
|
+
@dataclass(frozen=True, slots=True)
|
|
140
|
+
class MaterializedArtifact:
|
|
141
|
+
"""Algorithm-produced bytes plus their transfer metadata."""
|
|
142
|
+
|
|
143
|
+
path: Path
|
|
144
|
+
transfer_codec_id: Identifier
|
|
145
|
+
transfer_schema: TensorSchema
|
|
146
|
+
|
|
147
|
+
|
|
148
|
+
@dataclass(frozen=True, slots=True)
|
|
149
|
+
class UpdateBundle:
|
|
150
|
+
"""Materialized algorithm update whose contents are opaque to gossip."""
|
|
151
|
+
|
|
152
|
+
metadata: OpaqueUpdateBundleMetadata
|
|
153
|
+
artifacts: tuple[MaterializedArtifact, ...]
|
|
154
|
+
|
|
155
|
+
def __post_init__(self) -> None:
|
|
156
|
+
if len(self.artifacts) != len(self.metadata.artifacts):
|
|
157
|
+
raise ValueError("materialized artifacts must match bundle metadata")
|
|
158
|
+
|
|
159
|
+
@property
|
|
160
|
+
def digest(self) -> str:
|
|
161
|
+
return update_bundle_digest(self.metadata)
|
|
162
|
+
|
|
163
|
+
def validate_materialized(self, max_bytes: int) -> None:
|
|
164
|
+
"""Validate materialized hashes plus aggregate encoded size."""
|
|
165
|
+
declared = sum(artifact.size_bytes for artifact in self.metadata.artifacts)
|
|
166
|
+
actual = sum(artifact.path.stat().st_size for artifact in self.artifacts)
|
|
167
|
+
if actual != declared:
|
|
168
|
+
raise ValueError("materialized update bundle size mismatch")
|
|
169
|
+
if actual > max_bytes:
|
|
170
|
+
raise ValueError("materialized update bundle exceeds payload limit")
|
|
171
|
+
for metadata, materialized in zip(
|
|
172
|
+
self.metadata.artifacts, self.artifacts, strict=True
|
|
173
|
+
):
|
|
174
|
+
if file_sha256(materialized.path) != metadata.sha256:
|
|
175
|
+
raise ValueError("materialized update bundle checksum mismatch")
|
|
176
|
+
|
|
177
|
+
|
|
178
|
+
__all__ = [
|
|
179
|
+
"AlgorithmSnapshot",
|
|
180
|
+
"AlgorithmUpdate",
|
|
181
|
+
"MaterializedArtifact",
|
|
182
|
+
"NamedValidatedUpdate",
|
|
183
|
+
"SerializableState",
|
|
184
|
+
"UpdateBundle",
|
|
185
|
+
"ValidatedUpdate",
|
|
186
|
+
"checksum_artifacts",
|
|
187
|
+
"checksum_tensors",
|
|
188
|
+
]
|