sequence-base 0.8.0rc2__tar.gz → 0.8.1__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.
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/PKG-INFO +1 -2
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/pyproject.toml +2 -2
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/__init__.py +1 -3
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/by_name.py +8 -28
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/codec/blocks.py +2 -7
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/codec/compressed.py +1 -2
- sequence_base-0.8.1/sequence_base/constants/serving.py +52 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/constants/version.py +2 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/__init__.py +8 -8
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/accounts.py +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/adapters.py +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/benchmarks.py +1 -2
- sequence_base-0.8.1/sequence_base/contract/connect.py +32 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/deployments.py +6 -16
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/observe.py +6 -24
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/own_io.py +1 -7
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/planner.py +1 -2
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/policy.py +7 -47
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/settings.py +8 -5
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/templates.py +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/volumes.py +3 -4
- sequence_base-0.8.1/sequence_base/contract/wire.py +60 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/msgpack_io.py +2 -8
- sequence_base-0.8.1/sequence_base/presets/prompts.py +11 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schema_export.py +1 -4
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ActRequest.json +19 -7
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ActResponse.json +1 -9
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ActionChunk.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ActionContract.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ActionStep.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/AdapterPair.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/AddLocal.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/Autoscaler.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/BenchmarkRegisterRequest.json +2 -2
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/BenchmarkSpec.json +3 -3
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/BenchmarkStatusResponse.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/BootstrapResponse.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/Camera.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/CodeBundle.json +1 -1
- sequence_base-0.8.1/sequence_base/schemas/Concurrency.json +57 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/DeploymentAccepted.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/DeploymentPolicy.json +1 -6
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/DeploymentRequest.json +3 -64
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/DeploymentStatusResponse.json +1 -18
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ErrorBody.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ErrorCodes.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/Estimate.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/EvalDryRunResponse.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/EvalReport.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/EvalRequest.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/EvalResponse.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ImageFrame.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ImageSpec.json +2 -2
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ImageStepAddLocal.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ImageStepApt.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ImageStepEnv.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ImageStepProject.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ImageStepRun.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ImageStepTemplateFiles.json +2 -2
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ImageStepUvPip.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/InferParam.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/Lease.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/MetricAgg.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/Observation.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ObservationContract.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ObserveRequest.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ObserveResponse.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlanAppend.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlanAppended.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerMember.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerObservation.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerObserve.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerObserved.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerPlanIn.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerPlans.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerPlansWritten.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerRobot.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerRound.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerRoundRequest.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PlannerSpec.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PolicyRegistration.json +4 -65
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/PolicySpec.json +3 -64
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/Project.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/Proprioception.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/ReplaySpec.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/Resources.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/RolloutResult.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/SecretCreateRequest.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/SecretInfo.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/SecretRef.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/SeqFlags.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/SettingsChange.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/SettingsChanged.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/SettingsView.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/TaskAgg.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/TemplateFiles.json +2 -2
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/Tensors.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/Timeouts.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/UploadTarget.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/VolumeFile.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/VolumeInfo.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/VolumeMount.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/VolumeRef.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/VolumeUploadRequest.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/VolumeUploadResponse.json +1 -1
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/schemas/WarmingError.json +1 -1
- sequence_base-0.8.0rc2/sequence_base/contract/connect.py +0 -96
- sequence_base-0.8.0rc2/sequence_base/contract/wire.py +0 -47
- sequence_base-0.8.0rc2/sequence_base/presets/frames.py +0 -180
- sequence_base-0.8.0rc2/sequence_base/schemas/Concurrency.json +0 -113
- sequence_base-0.8.0rc2/sequence_base/schemas/ConnectRequest.json +0 -28
- sequence_base-0.8.0rc2/sequence_base/schemas/ConnectResponse.json +0 -204
- sequence_base-0.8.0rc2/sequence_base/schemas/Standby.json +0 -92
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/.gitignore +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/README.md +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/auth.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/codec/__init__.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/codec/image.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/codec/marked.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/constants/__init__.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/constants/cards.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/constants/headers.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/constants/limits.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/construct.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/errors.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/eval.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/contract/flags.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/episode.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/errors/__init__.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/errors/transport.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/obs/__init__.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/obs/log.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/obs/rid.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/presets/__init__.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/presets/robot_arm.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/telemetry/__init__.py +0 -0
- {sequence_base-0.8.0rc2 → sequence_base-0.8.1}/sequence_base/telemetry/events.py +0 -0
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.5
|
|
2
2
|
Name: sequence-base
|
|
3
|
-
Version: 0.8.
|
|
3
|
+
Version: 0.8.1
|
|
4
4
|
Summary: Shared foundation for the Sequences platform — wire contract, correlation, error taxonomy, constants. Used by the robot SDK, gateway, and container harness.
|
|
5
5
|
Project-URL: Changelog, https://generalsequences.com/docs/changelog/
|
|
6
6
|
Requires-Python: >=3.10
|
|
@@ -13,7 +13,6 @@ Requires-Dist: av>=17; extra == 'observe'
|
|
|
13
13
|
Requires-Dist: numpy>=1.26; extra == 'observe'
|
|
14
14
|
Requires-Dist: zstandard>=0.22; extra == 'observe'
|
|
15
15
|
Provides-Extra: test
|
|
16
|
-
Requires-Dist: hypothesis>=6; extra == 'test'
|
|
17
16
|
Requires-Dist: jsonschema>=4; extra == 'test'
|
|
18
17
|
Requires-Dist: pytest; extra == 'test'
|
|
19
18
|
Requires-Dist: pyyaml; extra == 'test'
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
[project]
|
|
2
2
|
name = "sequence-base"
|
|
3
|
-
version = "0.8.
|
|
3
|
+
version = "0.8.1"
|
|
4
4
|
description = "Shared foundation for the Sequences platform — wire contract, correlation, error taxonomy, constants. Used by the robot SDK, gateway, and container harness."
|
|
5
5
|
requires-python = ">=3.10"
|
|
6
6
|
dependencies = ["pydantic>=2.0"]
|
|
@@ -11,7 +11,7 @@ Changelog = "https://generalsequences.com/docs/changelog/"
|
|
|
11
11
|
[project.optional-dependencies]
|
|
12
12
|
codec = ["pillow>=10", "numpy>=1.26"]
|
|
13
13
|
observe = ["numpy>=1.26", "av>=17", "zstandard>=0.22"]
|
|
14
|
-
test = ["pytest", "jsonschema>=4", "
|
|
14
|
+
test = ["pytest", "jsonschema>=4", "pyyaml"]
|
|
15
15
|
|
|
16
16
|
[build-system]
|
|
17
17
|
requires = ["hatchling"]
|
|
@@ -55,8 +55,6 @@ from sequence_base.contract import (
|
|
|
55
55
|
BootstrapResponse,
|
|
56
56
|
CodeBundle,
|
|
57
57
|
Concurrency,
|
|
58
|
-
ConnectRequest,
|
|
59
|
-
ConnectResponse,
|
|
60
58
|
DeploymentAccepted,
|
|
61
59
|
DeploymentPolicy,
|
|
62
60
|
InferParam,
|
|
@@ -149,7 +147,7 @@ __all__ = [
|
|
|
149
147
|
"Resources", "Autoscaler", "Timeouts", "DeploymentPolicy", "InferParam", "DeploymentRequest",
|
|
150
148
|
"DeploymentAccepted", "DeploymentState", "DeploymentStatusResponse",
|
|
151
149
|
|
|
152
|
-
"Lease",
|
|
150
|
+
"Lease",
|
|
153
151
|
|
|
154
152
|
"ReplaySpec", "BenchmarkSpec", "BenchmarkRegisterRequest", "BenchmarkStatusResponse",
|
|
155
153
|
"EvalRequest", "Estimate", "EvalResponse", "EvalDryRunResponse", "RolloutResult",
|
|
@@ -16,10 +16,9 @@ declared (seq/named.py, registry.infer_params), the worker when it calls the met
|
|
|
16
16
|
rule on both sides.
|
|
17
17
|
|
|
18
18
|
What the platform hands a method itself is `episode` - an act function, an observe function and the method an
|
|
19
|
-
episode's end calls alike - and an eval's `seed`; `
|
|
20
|
-
|
|
21
|
-
|
|
22
|
-
declared, the worker when it loads one declared before.
|
|
19
|
+
episode's end calls alike - and an eval's `seed`; `ends(parameters, what)` refuses an end method that needs anything but
|
|
20
|
+
the episode, in words that say what to write instead, when the SDK declares a class.
|
|
21
|
+
|
|
23
22
|
|
|
24
23
|
|
|
25
24
|
A composite's planner, @seq.plan, is called on a clock with the episode alone and what it returns is kept in the
|
|
@@ -34,8 +33,8 @@ from collections.abc import Iterable, Mapping
|
|
|
34
33
|
from dataclasses import dataclass
|
|
35
34
|
from typing import Any
|
|
36
35
|
|
|
37
|
-
__all__ = ["INFER_PLATFORM", "OBSERVE_PLATFORM", "
|
|
38
|
-
"
|
|
36
|
+
__all__ = ["INFER_PLATFORM", "OBSERVE_PLATFORM", "WAKE_ON", "ByName", "ends", "plan_refused", "takes_by_name",
|
|
37
|
+
"wake_on_refused"]
|
|
39
38
|
|
|
40
39
|
|
|
41
40
|
|
|
@@ -43,15 +42,6 @@ INFER_PLATFORM = frozenset({"seed", "episode"})
|
|
|
43
42
|
OBSERVE_PLATFORM = frozenset({"step", "episode"})
|
|
44
43
|
|
|
45
44
|
|
|
46
|
-
RETIRED = {"episode_id": "take episode, a seq.Episode, and read episode.id",
|
|
47
|
-
"embodiment": "send what the robot is with its observations, in observe(x)",
|
|
48
|
-
"feedback": "send what the robot was playing with its observations, in observe(x)",
|
|
49
|
-
|
|
50
|
-
"plan": ('take episode and read the newest plan as episode.state["seq.plans"][-1]["value"], what '
|
|
51
|
-
'@seq.plan returned (before the first, episode.state.get("seq.plans") is None)'),
|
|
52
|
-
"robot": ("take episode: episode.state is the robot's own dict, which @seq.plan is handed too, and "
|
|
53
|
-
"episode.id names it")}
|
|
54
|
-
|
|
55
45
|
WAKE_ON = ('@seq.plan(wake_on=...) is a list of the observation\'s keys whose arrival plans at once, each a string - '
|
|
56
46
|
'wake_on=["say"]: an observation that holds "say" plans its episode then, its clock going on beside')
|
|
57
47
|
|
|
@@ -114,22 +104,12 @@ def takes_by_name(parameters: Iterable[inspect.Parameter], platform: frozenset[s
|
|
|
114
104
|
rest=any(p.kind is inspect.Parameter.VAR_KEYWORD for p in every), platform=frozenset(platform))
|
|
115
105
|
|
|
116
106
|
|
|
117
|
-
def retired(parameters: Iterable[inspect.Parameter], what: str) -> str | None:
|
|
118
|
-
"""Why `what` ("Model.infer"), a method the platform calls, cannot be deployed: it takes what the platform no longer
|
|
119
|
-
hands it (RETIRED), said with what to write instead - or None, when it takes none of it."""
|
|
120
|
-
names = [p.name for p in parameters if p.name in RETIRED]
|
|
121
|
-
if not names:
|
|
122
|
-
return None
|
|
123
|
-
return (f"{what} takes {' and '.join(names)}, which the platform no longer hands it: "
|
|
124
|
-
+ "; ".join(RETIRED[n] for n in names))
|
|
125
|
-
|
|
126
|
-
|
|
127
107
|
def ends(parameters: Iterable[inspect.Parameter], what: str) -> str | None:
|
|
128
108
|
"""Why `what` ("Model.end"), the method an episode's end calls (@seq.on_episode), cannot be deployed: an episode's
|
|
129
|
-
end hands it the episode alone, so it takes `episode` or nothing
|
|
130
|
-
|
|
109
|
+
end hands it the episode alone, so it takes `episode` or nothing, or any other argument with a default - or None,
|
|
110
|
+
when it does."""
|
|
131
111
|
taken = [p.name for p in parameters if p.kind not in _VARIADIC
|
|
132
|
-
and
|
|
112
|
+
and p.name != "episode" and p.default is inspect.Parameter.empty]
|
|
133
113
|
if not taken:
|
|
134
114
|
return None
|
|
135
115
|
name = what.rsplit(".", 1)[-1]
|
|
@@ -4,8 +4,7 @@ of one whose content its worker already holds, and what the worker turns back -
|
|
|
4
4
|
large input sent again with every observation, goes up once.
|
|
5
5
|
|
|
6
6
|
|
|
7
|
-
In one sentence: the SDK, packing an observation for
|
|
8
|
-
or str of MIN_BLOCK_BYTES or more into blocks - an array stacked along its first axis, 2 to SPLIT_MAX slices each of
|
|
7
|
+
In one sentence: the SDK, packing an observation for its worker, cuts each unmarked numpy array, bytes or str of MIN_BLOCK_BYTES or more into blocks - an array stacked along its first axis, 2 to SPLIT_MAX slices each of
|
|
9
8
|
MIN_BLOCK_BYTES or more, a block a slice - each named by the first ID_BYTES of its sha256 (block_id) → puts a reference
|
|
10
9
|
in its place (encode) → its post carries the content of each block its worker does not hold (ObserveRequest.blocks), as
|
|
11
10
|
both sides count what the worker holds (Retention) → the worker keeps each episode's blocks and, as each observation's
|
|
@@ -36,14 +35,10 @@ from typing import Any, Iterable, Mapping
|
|
|
36
35
|
|
|
37
36
|
from sequence_base.codec.compressed import is_compressed
|
|
38
37
|
from sequence_base.codec.marked import MARK, is_marked
|
|
39
|
-
from sequence_base.contract.observe import BLOCKS_FORMAT
|
|
40
38
|
|
|
41
|
-
__all__ = ["
|
|
39
|
+
__all__ = ["ID_BYTES", "KEEP_OBSERVATIONS", "KINDS", "MIN_BLOCK_BYTES", "SPLIT_MAX", "VERSION", "BlocksError",
|
|
42
40
|
"MissingBlock", "Retention", "block_id", "cut", "decode", "encode", "ids_of", "is_reference"]
|
|
43
41
|
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
FORMAT = BLOCKS_FORMAT
|
|
47
42
|
KINDS = ("array", "bytes", "str")
|
|
48
43
|
VERSION = 1
|
|
49
44
|
ID_BYTES = 16
|
|
@@ -6,8 +6,7 @@ after the worker is handed the bytes the robot built.
|
|
|
6
6
|
|
|
7
7
|
|
|
8
8
|
In one sentence: the robot's SDK compresses an unmarked numpy array or bytes of 16 KiB or more with zstd level 1 (zlib
|
|
9
|
-
level 1 where it has no zstd) and puts the map below in its place →
|
|
10
|
-
only it is sent such maps → as a post comes, before anything keeps it, expand() turns every map back into the array or
|
|
9
|
+
level 1 where it has no zstd) and puts the map below in its place → as a post comes, before anything keeps it, expand() turns every map back into the array or
|
|
11
10
|
bytes it was (np.frombuffer: read-only, as msgpack_io reads an array) and says how many bytes that added.
|
|
12
11
|
|
|
13
12
|
The map, version 1 (inside the observe body's msgpack; `data` is bin):
|
|
@@ -0,0 +1,52 @@
|
|
|
1
|
+
"""The platform's serving numbers, defined once here: how long an idle worker stays, how long a start, a line and an act
|
|
2
|
+
may take, how a robot's heartbeat and a console's teleop are held, how many robots a copy of a model seats, how many
|
|
3
|
+
cards one machine has and why a worker stopped serving - the gateway, the worker and the SDK each kept a copy of these.
|
|
4
|
+
|
|
5
|
+
"""
|
|
6
|
+
|
|
7
|
+
|
|
8
|
+
|
|
9
|
+
IDLE_S = 600
|
|
10
|
+
|
|
11
|
+
QUEUE_TIMEOUT_S = 120
|
|
12
|
+
|
|
13
|
+
|
|
14
|
+
START_WAIT_S = 600
|
|
15
|
+
|
|
16
|
+
|
|
17
|
+
|
|
18
|
+
|
|
19
|
+
STARTUP_TIMEOUT_S = 1800
|
|
20
|
+
REQUEST_TIMEOUT_S = 300
|
|
21
|
+
EPISODE_IDLE_S = 300
|
|
22
|
+
|
|
23
|
+
|
|
24
|
+
|
|
25
|
+
HEARTBEAT_WINDOW_S = 2.0
|
|
26
|
+
|
|
27
|
+
|
|
28
|
+
TELEOP_BEAT_S = 0.5
|
|
29
|
+
TELEOP_LOST_S = 3.0
|
|
30
|
+
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
|
|
34
|
+
|
|
35
|
+
MARGIN_MS = 50.0
|
|
36
|
+
LOAD_MAX = 0.9
|
|
37
|
+
ROUND_TRIP_MS = 100.0
|
|
38
|
+
MAX_ROBOTS_PER_COPY = 64
|
|
39
|
+
|
|
40
|
+
|
|
41
|
+
MACHINE_CARDS = 8
|
|
42
|
+
|
|
43
|
+
MAX_WARM_WORKERS = 16
|
|
44
|
+
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
MAX_SOURCE_TREE_BYTES = 64 * 1024 * 1024
|
|
48
|
+
|
|
49
|
+
|
|
50
|
+
|
|
51
|
+
|
|
52
|
+
DEATH_CAUSES = ("gpu_out_of_memory", "host_out_of_memory", "crashed", "hung", "group_member_lost", "load_failed")
|
|
@@ -46,7 +46,7 @@ from sequence_base.contract.policy import (
|
|
|
46
46
|
)
|
|
47
47
|
|
|
48
48
|
|
|
49
|
-
from sequence_base.contract.own_io import
|
|
49
|
+
from sequence_base.contract.own_io import Tensors
|
|
50
50
|
|
|
51
51
|
|
|
52
52
|
from sequence_base.contract.deployments import (
|
|
@@ -64,7 +64,7 @@ from sequence_base.contract.deployments import (
|
|
|
64
64
|
)
|
|
65
65
|
|
|
66
66
|
|
|
67
|
-
from sequence_base.contract.connect import
|
|
67
|
+
from sequence_base.contract.connect import Lease
|
|
68
68
|
|
|
69
69
|
|
|
70
70
|
from sequence_base.contract.benchmarks import (
|
|
@@ -102,7 +102,6 @@ from sequence_base.contract.adapters import AdapterPair, parse_pairs
|
|
|
102
102
|
|
|
103
103
|
from sequence_base.contract.volumes import (
|
|
104
104
|
MAX_FILES,
|
|
105
|
-
MAX_UPLOAD_BYTES,
|
|
106
105
|
UploadTarget,
|
|
107
106
|
VolumeFile,
|
|
108
107
|
VolumeInfo,
|
|
@@ -135,7 +134,7 @@ from sequence_base.contract.planner import (
|
|
|
135
134
|
PLANNER_ID,
|
|
136
135
|
POLICY_ID,
|
|
137
136
|
ROBOT_ID,
|
|
138
|
-
|
|
137
|
+
WAITING_FOR_PLAN,
|
|
139
138
|
PlanAppend,
|
|
140
139
|
PlanAppended,
|
|
141
140
|
PlannerMember,
|
|
@@ -165,13 +164,13 @@ __all__ = [
|
|
|
165
164
|
"TemplateFiles", "ImageStepTemplateFiles", "Project", "ImageStepProject", "FROZEN_PATH", "ImageStep", "ImageSpec",
|
|
166
165
|
"VolumeMount", "VolumeRef", "Concurrency", "ActionContract", "Camera", "ObservationContract",
|
|
167
166
|
"PolicySpec", "PolicyRegistration", "PolicyConstructor", "CHUNK_NAMES",
|
|
168
|
-
"
|
|
167
|
+
"Tensors",
|
|
169
168
|
|
|
170
169
|
"MAX_CODE_BUNDLE_BYTES", "CodeBundle",
|
|
171
170
|
"Resources", "Autoscaler", "Timeouts", "DeploymentPolicy", "InferParam", "DeploymentRequest",
|
|
172
171
|
"DeploymentAccepted", "DeploymentState", "DeploymentStatusResponse",
|
|
173
172
|
|
|
174
|
-
"Lease",
|
|
173
|
+
"Lease",
|
|
175
174
|
|
|
176
175
|
"ReplaySpec", "BenchmarkSpec", "BenchmarkRegisterRequest", "BenchmarkStatusResponse",
|
|
177
176
|
"EvalRequest", "Estimate", "EvalResponse", "EvalDryRunResponse", "RolloutResult",
|
|
@@ -179,13 +178,14 @@ __all__ = [
|
|
|
179
178
|
|
|
180
179
|
"SecretRef", "SecretCreateRequest", "SecretInfo", "BootstrapResponse",
|
|
181
180
|
|
|
182
|
-
"MAX_FILES", "
|
|
181
|
+
"MAX_FILES", "VolumeFile", "VolumeUploadRequest", "UploadTarget", "VolumeUploadResponse",
|
|
183
182
|
"VolumeInfo", "check_volume_name", "parse_vol",
|
|
184
183
|
|
|
185
184
|
"SETTINGS", "Setting", "SettingsChange", "Applied", "SettingState", "SettingsView", "SettingChanged",
|
|
186
185
|
"SettingsChanged",
|
|
187
186
|
|
|
188
|
-
"KEPT", "MOST_LAYERS", "PLANS_MOST", "STATE_MOST", "PLAN_ON_A_POLICY", "PLANNER_ID", "POLICY_ID", "ROBOT_ID", "
|
|
187
|
+
"KEPT", "MOST_LAYERS", "PLANS_MOST", "STATE_MOST", "PLAN_ON_A_POLICY", "PLANNER_ID", "POLICY_ID", "ROBOT_ID", "WAITING_FOR_PLAN",
|
|
188
|
+
"PlannerSpec",
|
|
189
189
|
"PlannerRoundRequest", "PlannerRobot", "PlannerMember", "PlannerObservation", "PlannerPlanIn", "PlannerRound",
|
|
190
190
|
"PlannerPlans", "PlannerPlansWritten", "PlanAppend", "PlanAppended", "PlannerObserve", "PlannerObserved",
|
|
191
191
|
|
|
@@ -3,7 +3,7 @@ Contract 6 — accounts / secret / billing wire types.
|
|
|
3
3
|
|
|
4
4
|
Secret-leak rule, mechanically checkable (goal 15). Two kinds of secret:
|
|
5
5
|
(a) platform-minted — `direct_secret`, TLS private key. May appear in the DOWN direction only:
|
|
6
|
-
|
|
6
|
+
a connect's answer, BootstrapResponse.direct_secret + tls_key.
|
|
7
7
|
(b) user Secret-Manager values. May appear in the WRITE direction only:
|
|
8
8
|
SecretCreateRequest.env, BootstrapResponse.env (injected into the worker subprocess).
|
|
9
9
|
These four fields are the entire allow-list for carrying a plaintext secret value. Everything returned
|
|
@@ -57,7 +57,7 @@ def _side(value: Any, which: str, i: int) -> str:
|
|
|
57
57
|
return side
|
|
58
58
|
if side.startswith(TEMPLATE):
|
|
59
59
|
if not _TEMPLATE_NAME.fullmatch(side[len(TEMPLATE):]):
|
|
60
|
-
raise ValueError(f"pair {i}: {which} {value!r}: a template is template:<name>,
|
|
60
|
+
raise ValueError(f"pair {i}: {which} {value!r}: a template is template:<name>, an official template's name "
|
|
61
61
|
"(lowercase letters, digits or '-'), e.g. template:libero-10")
|
|
62
62
|
return side
|
|
63
63
|
if not _NAME.fullmatch(side):
|
|
@@ -36,8 +36,7 @@ TASK_NAME = r"^[A-Za-z0-9][A-Za-z0-9._/:-]{0,127}$"
|
|
|
36
36
|
|
|
37
37
|
SETUP_GONE = ("@seq.setup is gone: a benchmark builds its worlds in its constructor, after the machine's snapshot - "
|
|
38
38
|
'init= its arguments, "{num_worlds}" how many worlds this eval builds and "{policy.<field>}" what the '
|
|
39
|
-
"evaluated policy declares; a benchmark registered with @seq.setup
|
|
40
|
-
"0.51.1 or later (seq deploy)")
|
|
39
|
+
"evaluated policy declares; a benchmark registered with @seq.setup has to be registered again")
|
|
41
40
|
|
|
42
41
|
|
|
43
42
|
def _names_worlds(constructor: PolicyConstructor | None) -> bool:
|
|
@@ -0,0 +1,32 @@
|
|
|
1
|
+
"""
|
|
2
|
+
Contract 3 — a worker's lease, and the cold-start answer of a connect (control plane).
|
|
3
|
+
|
|
4
|
+
A connect's own body and answer are not here: the gateway reads the one by its model (routes/connect.py ConnectRequest)
|
|
5
|
+
and sequence-ai sends and reads a dict, and the classes kept here matched neither - no `session`, which it requires - and
|
|
6
|
+
were read by nothing. A lease is read by the worker's bootstrap (contract/accounts.py). The cold-start fallback is a
|
|
7
|
+
defined wire body, not an HTTP header alone — a 503 returns `WarmingError` (contract/errors.py), so `503 + Retry-After`
|
|
8
|
+
and the warming ETA are typed and checkable.
|
|
9
|
+
"""
|
|
10
|
+
from __future__ import annotations
|
|
11
|
+
|
|
12
|
+
from datetime import datetime
|
|
13
|
+
from typing import Literal
|
|
14
|
+
|
|
15
|
+
from pydantic import BaseModel, ConfigDict, Field
|
|
16
|
+
|
|
17
|
+
|
|
18
|
+
from sequence_base.contract.errors import WarmingError
|
|
19
|
+
|
|
20
|
+
__all__ = ["Lease", "WarmingError"]
|
|
21
|
+
|
|
22
|
+
|
|
23
|
+
class Lease(BaseModel):
|
|
24
|
+
"""A time-boxed grant of the direct coordinates. Renewed before `expires_at`; `epoch` is the
|
|
25
|
+
worker/secret generation — a client holding a stale epoch against a new worker gets 401 and must
|
|
26
|
+
re-connect."""
|
|
27
|
+
|
|
28
|
+
model_config = ConfigDict(extra="forbid")
|
|
29
|
+
|
|
30
|
+
ttl_s: int
|
|
31
|
+
expires_at: datetime
|
|
32
|
+
epoch: int = Field(..., description="Worker/secret generation counter; rolls on worker change")
|
|
@@ -56,6 +56,8 @@ class Resources(BaseModel):
|
|
|
56
56
|
|
|
57
57
|
|
|
58
58
|
|
|
59
|
+
|
|
60
|
+
|
|
59
61
|
cpu: float | None = Field(default=None, ge=0.125, le=64)
|
|
60
62
|
cpu_limit: float | None = Field(default=None, ge=0.125, le=128)
|
|
61
63
|
memory_gb: float | None = Field(default=None, ge=0.125, le=1024)
|
|
@@ -207,9 +209,6 @@ class DeploymentPolicy(BaseModel):
|
|
|
207
209
|
|
|
208
210
|
|
|
209
211
|
episodes: Episodes | None = None
|
|
210
|
-
|
|
211
|
-
|
|
212
|
-
streaming: bool = Field(default=False, exclude=True)
|
|
213
212
|
secrets: list[SecretRef] = Field(default_factory=list)
|
|
214
213
|
|
|
215
214
|
|
|
@@ -247,17 +246,6 @@ class DeploymentPolicy(BaseModel):
|
|
|
247
246
|
raise ValueError(PLAN_ON_A_POLICY)
|
|
248
247
|
return self
|
|
249
248
|
|
|
250
|
-
@model_validator(mode="after")
|
|
251
|
-
def _episodes_of_streaming(self) -> "DeploymentPolicy":
|
|
252
|
-
|
|
253
|
-
if self.streaming and self.episodes not in (None, "own"):
|
|
254
|
-
raise ValueError(f"streaming is what a record from before contract 0.5.24 says; beside "
|
|
255
|
-
f"episodes={self.episodes!r}, leave it out")
|
|
256
|
-
if self.streaming:
|
|
257
|
-
self.episodes = "own"
|
|
258
|
-
self.streaming = self.episodes == "own"
|
|
259
|
-
return self
|
|
260
|
-
|
|
261
249
|
@model_validator(mode="after")
|
|
262
250
|
def _plays_what_it_holds(self) -> "DeploymentPolicy":
|
|
263
251
|
check_chunk_lens(self.action_chunk_valid_len, self.action_chunk_replan_at)
|
|
@@ -380,6 +368,10 @@ class DeploymentRequest(BaseModel):
|
|
|
380
368
|
|
|
381
369
|
|
|
382
370
|
|
|
371
|
+
|
|
372
|
+
|
|
373
|
+
|
|
374
|
+
|
|
383
375
|
_NAME = re.compile(r"[A-Za-z0-9][A-Za-z0-9._-]{0,63}")
|
|
384
376
|
|
|
385
377
|
|
|
@@ -409,5 +401,3 @@ class DeploymentStatusResponse(BaseModel):
|
|
|
409
401
|
status: DeploymentState
|
|
410
402
|
model: str
|
|
411
403
|
error: str | None = Field(default=None, description="Build-failure reason; set only when FAILED")
|
|
412
|
-
|
|
413
|
-
reason: Literal["gpu_out_of_memory", "host_out_of_memory", "crashed"] | None = None
|
|
@@ -1,12 +1,11 @@
|
|
|
1
1
|
"""
|
|
2
|
-
The observation stream:
|
|
3
|
-
|
|
2
|
+
The observation stream: every observation a robot makes is sent to its worker as it is made (POST /observe), and its
|
|
3
|
+
act carries none.
|
|
4
4
|
|
|
5
5
|
In one sentence: the robot's SDK numbers each observation of an episode from 0 → posts them as they come, each post the
|
|
6
|
-
ones that came while the last was on its way (ObserveRequest) → the worker keeps them by episode
|
|
7
|
-
|
|
8
|
-
worker waits for
|
|
9
|
-
the act's own last - what an eval's since_last_act sends in one act today.
|
|
6
|
+
ones that came while the last was on its way (ObserveRequest) → the worker keeps them by episode, answers how far it has
|
|
7
|
+
every one (ObserveResponse.through) and hands each to the policy in order → an act names how many came before it
|
|
8
|
+
(ActRequest.seq) → the worker waits for those to reach the policy, STREAM_WAIT_S at most, then calls its act function.
|
|
10
9
|
|
|
11
10
|
WHY: an act was the only upload, so every frame between two acts waited for the next one - 24 frames, 1304 KB and a read
|
|
12
11
|
p50 of 425 ms from a robot at home. Sent as they are made, they
|
|
@@ -23,23 +22,7 @@ from typing import Any
|
|
|
23
22
|
|
|
24
23
|
from pydantic import BaseModel, ConfigDict, Field
|
|
25
24
|
|
|
26
|
-
__all__ = ["
|
|
27
|
-
"ObserveResponse", "ZSTD_FORMAT"]
|
|
28
|
-
|
|
29
|
-
|
|
30
|
-
|
|
31
|
-
FORMAT = "x-seq-observe/1"
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
|
|
35
|
-
|
|
36
|
-
ZSTD_FORMAT = "x-seq-zstd/1"
|
|
37
|
-
|
|
38
|
-
|
|
39
|
-
|
|
40
|
-
|
|
41
|
-
|
|
42
|
-
BLOCKS_FORMAT = "x-seq-blocks/1"
|
|
25
|
+
__all__ = ["FRAMES_MISSING", "FRAMES_OVER_LIMIT", "ObserveFrame", "ObserveRequest", "ObserveResponse"]
|
|
43
26
|
|
|
44
27
|
|
|
45
28
|
|
|
@@ -68,7 +51,6 @@ class ObserveRequest(BaseModel):
|
|
|
68
51
|
skipped: list[int] = Field(default_factory=list, description="Numbers left out on purpose, never to come")
|
|
69
52
|
|
|
70
53
|
|
|
71
|
-
|
|
72
54
|
blocks: dict[bytes, Any] = Field(default_factory=dict,
|
|
73
55
|
description="Block contents by id, for the references in the frames")
|
|
74
56
|
|
|
@@ -16,13 +16,7 @@ from typing import Any
|
|
|
16
16
|
|
|
17
17
|
from pydantic import BaseModel, ConfigDict
|
|
18
18
|
|
|
19
|
-
__all__ = ["
|
|
20
|
-
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
OBSERVATIONS_GONE = ("observations= is gone: a policy is handed every observation as its robot sends it - keep the ones "
|
|
24
|
-
"an act reads in an @seq.observe method (on self, or in episode.state), and deploy again with the "
|
|
25
|
-
"latest sequence-ai")
|
|
19
|
+
__all__ = ["Tensors", "rows_of"]
|
|
26
20
|
|
|
27
21
|
|
|
28
22
|
class Tensors(BaseModel):
|
|
@@ -49,9 +49,8 @@ STATE_MOST = 256 * 1024
|
|
|
49
49
|
|
|
50
50
|
|
|
51
51
|
|
|
52
|
+
WAITING_FOR_PLAN = "waiting_for_plan"
|
|
52
53
|
|
|
53
|
-
TEAM_GONE = ("team is gone: a team is a planner now - @seq.planner(members=[...]) names the policies policies= named, "
|
|
54
|
-
"and planners too; a robot connects to its own policy as ever, and deploy again with the latest sequence-ai")
|
|
55
54
|
|
|
56
55
|
|
|
57
56
|
PLAN_ON_A_POLICY = ("@seq.plan is a planner's now: a policy is planned by a @seq.planner that names it in members= - "
|
|
@@ -115,7 +115,7 @@ class TemplateFiles(BaseModel):
|
|
|
115
115
|
|
|
116
116
|
model_config = ConfigDict(extra="forbid")
|
|
117
117
|
|
|
118
|
-
template: str = Field(..., pattern=r"^[a-z0-9][a-z0-9-]{0,63}$", description="the template's name
|
|
118
|
+
template: str = Field(..., pattern=r"^[a-z0-9][a-z0-9-]{0,63}$", description="the official template's name")
|
|
119
119
|
version: int = Field(..., ge=1, le=999_999_999, description="the template version whose files these are")
|
|
120
120
|
|
|
121
121
|
|
|
@@ -279,11 +279,8 @@ class VolumeRef(BaseModel):
|
|
|
279
279
|
|
|
280
280
|
|
|
281
281
|
class Concurrency(BaseModel):
|
|
282
|
-
"""How one worker serves robots, every field optional. `
|
|
283
|
-
|
|
284
|
-
worker). A worker's slots = cards // gpus_per_replica × replicas_per_gpu. `target_inputs` is how many robot sessions
|
|
285
|
-
a worker holds before another starts, `max_inputs` the most it takes while a new one boots — both default to the
|
|
286
|
-
slots.
|
|
282
|
+
"""How one worker serves robots, every field optional. `gpus_per_replica` cards make one copy (None: every card of
|
|
283
|
+
the worker makes its own copy).
|
|
287
284
|
|
|
288
285
|
`sizes` (contract 0.5.31) - from `@seq.concurrent(sizes=[1, 2, 4, 8])` on the policy's infer method - are the batch
|
|
289
286
|
sizes its infer is called with, ascending, the largest the most acts one forward takes (None: one robot a copy,
|
|
@@ -292,25 +289,14 @@ class Concurrency(BaseModel):
|
|
|
292
289
|
robot's act being a batch of one. A copy seats as many robots
|
|
293
290
|
as the largest size whose batch answers their control loops in time, and its worker holds an early act a little so
|
|
294
291
|
robots fall into one batch.
|
|
295
|
-
|
|
296
|
-
alone, or max_concurrent a copy however late - is gone. The SDK takes `gpus_per_replica` and `replica_start` in
|
|
292
|
+
The SDK takes `gpus_per_replica` and `replica_start` in
|
|
297
293
|
`@seq.policy(...)`.
|
|
298
|
-
|
|
299
|
-
|
|
300
|
-
`max_concurrent` (contract 0.5.5: one number, the largest batch, every size up to it possible), `max_inputs`,
|
|
301
|
-
`target_inputs` and `replicas_per_gpu` are older ways': kept so a stored spec and an older SDK's request still
|
|
302
|
-
read, and refused on a new deploy by the platform; the SDK from contract 0.5.31 sends none of them.
|
|
303
|
-
|
|
304
|
-
"""
|
|
294
|
+
"""
|
|
305
295
|
|
|
306
296
|
model_config = ConfigDict(extra="forbid")
|
|
307
297
|
|
|
308
298
|
sizes: list[int] | None = Field(default=None, min_length=1, max_length=256,
|
|
309
299
|
description="The batch sizes infer is called with, ascending; the largest the most")
|
|
310
|
-
max_concurrent: int | None = Field(default=None, ge=2, le=256)
|
|
311
|
-
max_inputs: int | None = Field(default=None, ge=1, le=1024)
|
|
312
|
-
target_inputs: int | None = Field(default=None, ge=1, le=1024)
|
|
313
|
-
replicas_per_gpu: int | None = Field(default=None, ge=1, le=64)
|
|
314
300
|
gpus_per_replica: int | None = Field(default=None, ge=1, le=256)
|
|
315
301
|
|
|
316
302
|
|
|
@@ -331,17 +317,6 @@ class Concurrency(BaseModel):
|
|
|
331
317
|
raise ValueError("the batch sizes begin with 1: a lone robot's act is a batch of one")
|
|
332
318
|
return v
|
|
333
319
|
|
|
334
|
-
@model_validator(mode="after")
|
|
335
|
-
def _consistent(self) -> "Concurrency":
|
|
336
|
-
if self.sizes is not None and self.max_concurrent is not None:
|
|
337
|
-
raise ValueError("a policy declares its batch sizes (sizes) or, as an older SDK did, one max_concurrent - "
|
|
338
|
-
"not both")
|
|
339
|
-
if self.max_inputs is not None and self.target_inputs is not None and self.max_inputs < self.target_inputs:
|
|
340
|
-
raise ValueError("max_inputs must be at least target_inputs")
|
|
341
|
-
if (self.gpus_per_replica or 1) > 1 and (self.replicas_per_gpu or 1) > 1:
|
|
342
|
-
raise ValueError("a copy that spans several cards is one copy per card group: replicas_per_gpu must be 1")
|
|
343
|
-
return self
|
|
344
|
-
|
|
345
320
|
|
|
346
321
|
|
|
347
322
|
|
|
@@ -385,10 +360,6 @@ class PolicySpec(BaseModel):
|
|
|
385
360
|
|
|
386
361
|
|
|
387
362
|
|
|
388
|
-
streaming: bool = Field(default=False, exclude=True)
|
|
389
|
-
|
|
390
|
-
|
|
391
|
-
|
|
392
363
|
|
|
393
364
|
|
|
394
365
|
|
|
@@ -428,17 +399,6 @@ class PolicySpec(BaseModel):
|
|
|
428
399
|
check_weights_uri(uri, f"weights[{name!r}]")
|
|
429
400
|
return self
|
|
430
401
|
|
|
431
|
-
@model_validator(mode="after")
|
|
432
|
-
def _episodes_of_streaming(self) -> "PolicySpec":
|
|
433
|
-
|
|
434
|
-
if self.streaming and self.episodes not in (None, "own"):
|
|
435
|
-
raise ValueError(f"streaming is what a record from before contract 0.5.24 says; beside "
|
|
436
|
-
f"episodes={self.episodes!r}, leave it out")
|
|
437
|
-
if self.streaming:
|
|
438
|
-
self.episodes = "own"
|
|
439
|
-
self.streaming = self.episodes == "own"
|
|
440
|
-
return self
|
|
441
|
-
|
|
442
402
|
@model_validator(mode="after")
|
|
443
403
|
def _plays_what_it_holds(self) -> "PolicySpec":
|
|
444
404
|
check_chunk_lens(self.action_chunk_valid_len, self.action_chunk_replan_at)
|
|
@@ -451,7 +411,7 @@ class PolicySpec(BaseModel):
|
|
|
451
411
|
if self.planner is not None:
|
|
452
412
|
said = [name for name in ("hz", "action_chunk_valid_len", "action_chunk_replan_at", "tensors")
|
|
453
413
|
if getattr(self, name) is not None]
|
|
454
|
-
said += [f"concurrency.{name}" for name in ("sizes",
|
|
414
|
+
said += [f"concurrency.{name}" for name in ("sizes",)
|
|
455
415
|
if self.concurrency is not None and getattr(self.concurrency, name) is not None]
|
|
456
416
|
if said:
|
|
457
417
|
raise ValueError(f"a planner answers no act, so it declares no {', '.join(said)}")
|
|
@@ -496,7 +456,7 @@ def check_weights_uri(v: str, what: str = "weights") -> str:
|
|
|
496
456
|
if v.startswith("hf://") and len(v) > len("hf://"):
|
|
497
457
|
return v
|
|
498
458
|
raise ValueError(f"{what} must be a template:// official template, an hf:// repo or a vol:// volume: upload files "
|
|
499
|
-
"of your own once as a volume
|
|
459
|
+
"of your own once as a volume and use vol://<name>")
|
|
500
460
|
|
|
501
461
|
|
|
502
462
|
def check_weights_parts(weights_name: str | None, extra: dict) -> None:
|
|
@@ -25,6 +25,8 @@ from typing import Literal
|
|
|
25
25
|
|
|
26
26
|
from pydantic import BaseModel, ConfigDict, Field
|
|
27
27
|
|
|
28
|
+
from sequence_base.constants.serving import EPISODE_IDLE_S, MAX_WARM_WORKERS, STARTUP_TIMEOUT_S
|
|
29
|
+
|
|
28
30
|
__all__ = [
|
|
29
31
|
"Reads",
|
|
30
32
|
"Setting",
|
|
@@ -90,14 +92,15 @@ SETTINGS: tuple[Setting, ...] = (
|
|
|
90
92
|
minimum=1, maximum=256, unit="GPUs", code="max_gpus"),
|
|
91
93
|
_s("min_workers", "int", "gateway", "scaling", "Workers kept running, and billed, whether or not a robot is "
|
|
92
94
|
"connected.",
|
|
93
|
-
minimum=0, maximum=
|
|
95
|
+
minimum=0, maximum=MAX_WARM_WORKERS, unit="workers", code="min_workers", default=0),
|
|
94
96
|
_s("buffer_workers", "int", "gateway", "scaling", "Spare running workers, billed, beside the busy ones, so a robot "
|
|
95
|
-
"that connects or loses its worker starts at once.", minimum=0, maximum=
|
|
97
|
+
"that connects or loses its worker starts at once.", minimum=0, maximum=MAX_WARM_WORKERS, unit="workers",
|
|
98
|
+
code="buffer_workers",
|
|
96
99
|
default=0),
|
|
97
100
|
_s("schedule.hours", "str", "gateway", "scaling", "Days and hours it keeps schedule.min_workers running, as "
|
|
98
101
|
"'Mon-Fri 08:00-18:00'.", max_length=80),
|
|
99
102
|
_s("schedule.min_workers", "int", "gateway", "scaling", "Workers kept running in schedule.hours.",
|
|
100
|
-
minimum=0, maximum=
|
|
103
|
+
minimum=0, maximum=MAX_WARM_WORKERS, unit="workers"),
|
|
101
104
|
_s("schedule.tz", "str", "gateway", "scaling", "The time zone of schedule.hours, as 'America/Los_Angeles'.",
|
|
102
105
|
max_length=64, default="UTC"),
|
|
103
106
|
|
|
@@ -124,9 +127,9 @@ SETTINGS: tuple[Setting, ...] = (
|
|
|
124
127
|
"that it timed out.", minimum=1, maximum=3600, unit="s", code="seq.Timeouts(request=)"),
|
|
125
128
|
_s("timeouts.episode_idle", "int", "workers", "timeouts", "How long an episode may go without an act before its "
|
|
126
129
|
"place goes to a new one.", minimum=10, maximum=86400, unit="s", code="seq.Timeouts(episode_idle=)",
|
|
127
|
-
default=
|
|
130
|
+
default=EPISODE_IDLE_S),
|
|
128
131
|
_s("timeouts.startup", "int", "new_workers", "timeouts", "How long a worker may take to load.",
|
|
129
|
-
minimum=60, maximum=18000, unit="s", code="seq.Timeouts(startup=)", default=
|
|
132
|
+
minimum=60, maximum=18000, unit="s", code="seq.Timeouts(startup=)", default=STARTUP_TIMEOUT_S),
|
|
130
133
|
|
|
131
134
|
_s("plan.every_s", "float", "workers", "planner", "Seconds between calls of its @seq.plan method.",
|
|
132
135
|
above=0, unit="s", code="@seq.plan(every_s=)"),
|