sequence-base 0.3.5__tar.gz → 0.3.7__tar.gz

{sequence_base-0.3.5 → sequence_base-0.3.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: sequence-base
-Version: 0.3.5
+Version: 0.3.7
 Summary: Shared foundation for the Sequences platform — wire contract, correlation, error taxonomy, constants. Used by the robot SDK, gateway, and container harness.
 Requires-Python: >=3.11
 Requires-Dist: pydantic>=2.0

{sequence_base-0.3.5 → sequence_base-0.3.7}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sequence-base"
-version = "0.3.5"
+version = "0.3.7"
 description = "Shared foundation for the Sequences platform — wire contract, correlation, error taxonomy, constants. Used by the robot SDK, gateway, and container harness."
 requires-python = ">=3.11"
 dependencies = ["pydantic>=2.0"]

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/__init__.py

@@ -31,6 +31,7 @@ from sequence_base.contract import (  # noqa: F401  (re-export surface)
     AdapterKind,
     AdapterTransform,
     Autoscaler,
+    BenchmarkConditionSpec,
     BenchmarkRegisterRequest,
     BenchmarkStatusResponse,
     BenchmarkSpec,
@@ -43,6 +44,7 @@ from sequence_base.contract import (  # noqa: F401  (re-export surface)
     ConnectResponse,
     DeploymentAccepted,
     DeploymentPolicy,
+    InferParam,
     DeploymentRequest,
     DeploymentState,
     DeploymentStatusResponse,
@@ -121,12 +123,12 @@ __all__ = [
     "PolicySpec", "PolicyRegistration",
     # contract 2 — deployments
     "MAX_CODE_BUNDLE_BYTES", "CodeBundle",
-    "Resources", "Autoscaler", "Timeouts", "VolumeAttach", "DeploymentPolicy", "DeploymentRequest",
+    "Resources", "Autoscaler", "Timeouts", "VolumeAttach", "DeploymentPolicy", "InferParam", "DeploymentRequest",
     "DeploymentAccepted", "DeploymentState", "DeploymentStatusResponse",
     # contract 3 — connect / lease
     "Lease", "ConnectRequest", "ConnectResponse",
     # contract 5 — benchmark + eval
-    "Variant", "Conditions", "RenderSpec", "Step", "BenchmarkSpec", "BenchmarkRegisterRequest", "BenchmarkStatusResponse",
+    "Variant", "Conditions", "RenderSpec", "Step", "BenchmarkSpec", "BenchmarkConditionSpec", "BenchmarkRegisterRequest", "BenchmarkStatusResponse",
     "EvalRequest", "ClampInfo", "Estimate", "EvalResponse", "EvalDryRunResponse", "RolloutResult",
     "EvalStatus", "SuiteAgg", "TaskAgg", "MetricAgg", "LossyTransform", "EvalReport",
     "AdapterKind", "AdapterTransform",
@@ -144,4 +146,4 @@ __all__ = [
     "auth", "WorkerLogEvent",
 ]
 
-__version__ = "0.3.5"
+__version__ = "0.3.6"

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/constants/version.py

@@ -6,7 +6,9 @@ PROTOCOL_VERSION = "1"
 # across deployments / repos can tell which version of a shape it is validating against. MUST change
 # whenever any shared wire shape changes (a field added / removed / retyped) — that is what makes
 # contract evolution mechanically detectable instead of eyeballed. Bump on any shape change.
-CONTRACT_VERSION = "0.4.4"  # 0.4.3 -> 0.4.4: BenchmarkRegisterRequest.secrets (by name) + score (has @seq.score)
+CONTRACT_VERSION = "0.4.6"  # 0.4.5 -> 0.4.6: InferParam + DeploymentPolicy.params; EvalRequest.params / variants / record; BenchmarkRegisterRequest.gpu
+# previous: 0.4.5  # 0.4.4 -> 0.4.5: BenchmarkRegisterRequest.condition_list / timeout_s / default_adapter_ref (+ BenchmarkConditionSpec); EvalRequest.tasks / max_usd / rerun_of
+# previous: 0.4.4  # 0.4.3 -> 0.4.4: BenchmarkRegisterRequest.secrets (by name) + score (has @seq.score)
 # previous: 0.4.3  # 0.4.2 -> 0.4.3: I/O contract — ObservationContract/Camera/Embodiment on PolicySpec + DeploymentPolicy, Observation.state, ConnectRequest.embodiment, ConnectResponse.action/observation/embodiment
 # previous: 0.4.2  # 0.4.1 -> 0.4.2: 契约7 volumes (uploads: VolumeFile/VolumeUploadRequest/UploadTarget/VolumeUploadResponse/VolumeInfo), weights accept vol://<name>[@v<n>]; DeploymentRequest name is a slug and weights need an origin
 # previous: 0.4.1  # 0.4.0 -> 0.4.1: 契约2 deploy body completed during S3/G4 — DeploymentPolicy.secrets (name-only passthrough), DeploymentRequest.volumes: list[VolumeAttach{uri,mount,create_if_missing}], concurrency default materialized; telemetry events add load_start stage + emitted_at/rid/dedup_key (O4)

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/contract/__init__.py

@@ -39,6 +39,7 @@ from sequence_base.contract.deployments import (
     CodeBundle,
     DeploymentAccepted,
     DeploymentPolicy,
+    InferParam,
     DeploymentRequest,
     DeploymentState,
     DeploymentStatusResponse,
@@ -52,6 +53,7 @@ from sequence_base.contract.connect import ConnectRequest, ConnectResponse, Leas
 
 # contract 5 — benchmark registration + eval run / report / adapter
 from sequence_base.contract.benchmarks import (
+    BenchmarkConditionSpec,
     BenchmarkRegisterRequest,
     BenchmarkStatusResponse,
     BenchmarkSpec,
@@ -112,12 +114,12 @@ __all__ = [
     "PolicySpec", "PolicyRegistration",
     # contract 2
     "MAX_CODE_BUNDLE_BYTES", "CodeBundle",
-    "Resources", "Autoscaler", "Timeouts", "VolumeAttach", "DeploymentPolicy", "DeploymentRequest",
+    "Resources", "Autoscaler", "Timeouts", "VolumeAttach", "DeploymentPolicy", "InferParam", "DeploymentRequest",
     "DeploymentAccepted", "DeploymentState", "DeploymentStatusResponse",
     # contract 3
     "Lease", "ConnectRequest", "ConnectResponse",
     # contract 5
-    "Variant", "Conditions", "RenderSpec", "Step", "BenchmarkSpec", "BenchmarkRegisterRequest", "BenchmarkStatusResponse",
+    "Variant", "Conditions", "RenderSpec", "Step", "BenchmarkSpec", "BenchmarkConditionSpec", "BenchmarkRegisterRequest", "BenchmarkStatusResponse",
     "EvalRequest", "ClampInfo", "Estimate", "EvalResponse", "EvalDryRunResponse", "RolloutResult",
     "EvalStatus", "SuiteAgg", "TaskAgg", "MetricAgg", "LossyTransform", "EvalReport",
     "AdapterKind", "AdapterTransform",

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/contract/benchmarks.py

@@ -114,6 +114,20 @@ class BenchmarkSpec(BaseModel):
         return self
 
 
+class BenchmarkConditionSpec(BaseModel):
+    """One episode setting exactly as the author declared it — a sparse list or a cross product, with its own step
+    cap, wall-clock limit and suite. {gs-plans/customization/benchmark.md "6. 其它在 SDK 里收下、在生产里丢掉的声明"}"""
+
+    model_config = ConfigDict(extra="forbid")
+
+    task: str
+    seed: int
+    variant: dict[str, str] = Field(default_factory=dict)
+    max_steps: int | None = Field(default=None, gt=0)
+    timeout_s: float | None = Field(default=None, gt=0)
+    suite: str | None = None
+
+
 class BenchmarkRegisterRequest(BaseModel):
     """POST /v1/benchmarks — register body, same types as the spec (ImageSpec / Conditions / VolumeRef /
     RenderSpec). Carries assets / render / cpu / memory_gb / max_steps so the gateway can build the image,
@@ -145,6 +159,15 @@ class BenchmarkRegisterRequest(BaseModel):
     # Whether the benchmark defines @seq.score — the platform then calls it per group (overall, each suite, each
     # task) after every episode has run, and its keys override those groups' means. {benchmark.md "A.2"}
     score: bool = False
+    # Every condition as declared (a sparse list stays sparse; each keeps its own max_steps / timeout_s / suite),
+    # where `conditions` can only say tasks x seeds x variants. The eval runs these when present.
+    condition_list: list[BenchmarkConditionSpec] | None = Field(default=None, max_length=5000)
+    # An episode's wall-clock limit (seconds) when a condition sets none; past it the episode ends truncated.
+    timeout_s: float | None = Field(default=None, gt=0)
+    # The adapter (module:Class in the benchmark's own code) an eval uses when it names none.
+    default_adapter_ref: str | None = None
+    # A GPU for each sim sandbox (a renderer or a GPU physics engine): a card name the platform offers. None: CPU.
+    gpu: str | None = Field(default=None, max_length=32)
 
 
 class BenchmarkStatusResponse(BaseModel):

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/contract/deployments.py

@@ -88,6 +88,17 @@ class VolumeAttach(BaseModel):
     create_if_missing: bool
 
 
+class InferParam(BaseModel):
+    """One inference parameter the policy's `@seq.infer` declares — `def infer(self, obs, *, num_steps: int = 10)`
+    — with its type and default. An /act or an eval may set it; the worker passes only declared names, cast to
+    this type, and refuses any other. {gs-plans/customization/policy.md "四、推理行为"}"""
+
+    model_config = ConfigDict(extra="forbid")
+
+    type: Literal["int", "float", "bool", "str"]
+    default: int | float | bool | str | None = None
+
+
 class DeploymentPolicy(BaseModel):
     """The policy block of the deployments body. `handler_ref` and `action` are required; `action` is the
     fixed obs→ActionChunk contract, not a bare dict.
@@ -114,6 +125,16 @@ class DeploymentPolicy(BaseModel):
     # robot. None (an older SDK) is treated as undeclared, so an older policy is never refused for the
     # placeholder it never chose. {io-contract.md "旧 SDK 写的 POLICY 照样能 BUILD"}
     action_declared: bool | None = None
+    # The inference parameters @seq.infer takes beyond the observation (name -> type, default). None: an older
+    # SDK that did not look; {}: the policy declares none.
+    params: dict[str, InferParam] | None = Field(default=None, max_length=32)
+
+    @model_validator(mode="after")
+    def _param_names(self) -> "DeploymentPolicy":
+        for name in self.params or {}:
+            if not name.isidentifier() or name.startswith("_"):
+                raise ValueError(f"inference parameter {name!r} is not a public Python identifier")
+        return self
 
 
 MAX_CODE_BUNDLE_BYTES = 8 * 1024 * 1024

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/contract/eval.py

@@ -10,6 +10,7 @@ not a bare dict or free text.
 from __future__ import annotations
 
 from enum import Enum
+from typing import Literal
 
 from pydantic import BaseModel, ConfigDict, Field, model_validator
 
@@ -47,6 +48,19 @@ class EvalRequest(BaseModel):
     adapter: str | None = None
     compare: str | None = Field(default=None, description="Another model id to compare against")
     dry_run: bool = False
+    # Only these tasks (of the selected suite / trials). {gs-plans/customization/benchmark.md "子集"}
+    tasks: list[str] | None = Field(default=None, max_length=1000)
+    # The most this eval may spend on GPU time: it stops there and reports what ran (incomplete).
+    max_usd: float | None = Field(default=None, gt=0)
+    # Run again only the episodes of that earlier eval that errored; the report covers both runs.
+    rerun_of: str | None = None
+    # Inference parameters the model's @seq.infer declares, set for every /act of this eval.
+    params: dict[str, int | float | bool | str] | None = Field(default=None, max_length=32)
+    # Only the conditions whose variant takes one of these values on each named axis.
+    variants: dict[str, list[str]] | None = Field(default=None, max_length=32)
+    # Which episodes are recorded when the benchmark declares render=: the first seed of each task (default),
+    # every episode, or none.
+    record: Literal["per-task", "all", "none"] | None = None
 
 
 class ClampInfo(BaseModel):

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schema_export.py

@@ -35,6 +35,7 @@ from sequence_base.contract.act import (
     Proprioception,
 )
 from sequence_base.contract.benchmarks import (
+    BenchmarkConditionSpec,
     BenchmarkRegisterRequest,
     BenchmarkStatusResponse,
     BenchmarkSpec,
@@ -49,6 +50,7 @@ from sequence_base.contract.deployments import (
     CodeBundle,
     DeploymentAccepted,
     DeploymentPolicy,
+    InferParam,
     DeploymentRequest,
     DeploymentStatusResponse,
     Resources,
@@ -128,6 +130,7 @@ WIRE_MODELS: dict[str, type[BaseModel]] = {
     "VolumeAttach": VolumeAttach,
     "CodeBundle": CodeBundle,
     "DeploymentPolicy": DeploymentPolicy,
+    "InferParam": InferParam,
     "DeploymentRequest": DeploymentRequest,
     "DeploymentAccepted": DeploymentAccepted,
     "DeploymentStatusResponse": DeploymentStatusResponse,
@@ -141,6 +144,7 @@ WIRE_MODELS: dict[str, type[BaseModel]] = {
     "RenderSpec": RenderSpec,
     "Step": Step,
     "BenchmarkSpec": BenchmarkSpec,
+    "BenchmarkConditionSpec": BenchmarkConditionSpec,
     "BenchmarkRegisterRequest": BenchmarkRegisterRequest,
     "BenchmarkStatusResponse": BenchmarkStatusResponse,
     # contract 5 — eval run / report / adapter

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/ActionChunk.json

@@ -98,5 +98,5 @@
   ],
   "title": "ActionChunk",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/ActionContract.json

@@ -39,5 +39,5 @@
   ],
   "title": "ActionContract",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/ActionStep.json

@@ -19,5 +19,5 @@
   ],
   "title": "ActionStep",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/AdapterTransform.json

@@ -38,5 +38,5 @@
   ],
   "title": "AdapterTransform",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/Autoscaler.json

@@ -39,5 +39,5 @@
   },
   "title": "Autoscaler",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

sequence_base-0.3.7/sequence_base/schemas/BenchmarkConditionSpec.json

@@ -0,0 +1,66 @@
+{
+  "additionalProperties": false,
+  "description": "One episode setting exactly as the author declared it — a sparse list or a cross product, with its own step\ncap, wall-clock limit and suite. {gs-plans/customization/benchmark.md \"6. 其它在 SDK 里收下、在生产里丢掉的声明\"}",
+  "properties": {
+    "max_steps": {
+      "anyOf": [
+        {
+          "exclusiveMinimum": 0,
+          "type": "integer"
+        },
+        {
+          "type": "null"
+        }
+      ],
+      "default": null,
+      "title": "Max Steps"
+    },
+    "seed": {
+      "title": "Seed",
+      "type": "integer"
+    },
+    "suite": {
+      "anyOf": [
+        {
+          "type": "string"
+        },
+        {
+          "type": "null"
+        }
+      ],
+      "default": null,
+      "title": "Suite"
+    },
+    "task": {
+      "title": "Task",
+      "type": "string"
+    },
+    "timeout_s": {
+      "anyOf": [
+        {
+          "exclusiveMinimum": 0,
+          "type": "number"
+        },
+        {
+          "type": "null"
+        }
+      ],
+      "default": null,
+      "title": "Timeout S"
+    },
+    "variant": {
+      "additionalProperties": {
+        "type": "string"
+      },
+      "title": "Variant",
+      "type": "object"
+    }
+  },
+  "required": [
+    "task",
+    "seed"
+  ],
+  "title": "BenchmarkConditionSpec",
+  "type": "object",
+  "x-contract-version": "0.4.6"
+}

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/BenchmarkRegisterRequest.json

@@ -40,6 +40,71 @@
       "title": "ActionSpace",
       "type": "string"
     },
+    "BenchmarkConditionSpec": {
+      "additionalProperties": false,
+      "description": "One episode setting exactly as the author declared it — a sparse list or a cross product, with its own step\ncap, wall-clock limit and suite. {gs-plans/customization/benchmark.md \"6. 其它在 SDK 里收下、在生产里丢掉的声明\"}",
+      "properties": {
+        "max_steps": {
+          "anyOf": [
+            {
+              "exclusiveMinimum": 0,
+              "type": "integer"
+            },
+            {
+              "type": "null"
+            }
+          ],
+          "default": null,
+          "title": "Max Steps"
+        },
+        "seed": {
+          "title": "Seed",
+          "type": "integer"
+        },
+        "suite": {
+          "anyOf": [
+            {
+              "type": "string"
+            },
+            {
+              "type": "null"
+            }
+          ],
+          "default": null,
+          "title": "Suite"
+        },
+        "task": {
+          "title": "Task",
+          "type": "string"
+        },
+        "timeout_s": {
+          "anyOf": [
+            {
+              "exclusiveMinimum": 0,
+              "type": "number"
+            },
+            {
+              "type": "null"
+            }
+          ],
+          "default": null,
+          "title": "Timeout S"
+        },
+        "variant": {
+          "additionalProperties": {
+            "type": "string"
+          },
+          "title": "Variant",
+          "type": "object"
+        }
+      },
+      "required": [
+        "task",
+        "seed"
+      ],
+      "title": "BenchmarkConditionSpec",
+      "type": "object"
+    },
     "CodeBundle": {
       "additionalProperties": false,
       "description": "The policy's own source, shipped with the deploy so the worker can import `policy.handler_ref`\n(what Modal's automount does for the entrypoint module). A gzip'd tar of the handler's top-level module\nor package, base64 on the wire, with the sha256 of the decoded bytes so a truncated or altered upload\nis refused at the edge instead of failing as an ImportError on a GPU.",
@@ -338,6 +403,22 @@
       ],
       "default": null
     },
+    "condition_list": {
+      "anyOf": [
+        {
+          "items": {
+            "$ref": "#/$defs/BenchmarkConditionSpec"
+          },
+          "maxItems": 5000,
+          "type": "array"
+        },
+        {
+          "type": "null"
+        }
+      ],
+      "default": null,
+      "title": "Condition List"
+    },
     "conditions": {
       "$ref": "#/$defs/Conditions"
     },
@@ -353,6 +434,31 @@
       "default": null,
       "title": "Cpu"
     },
+    "default_adapter_ref": {
+      "anyOf": [
+        {
+          "type": "string"
+        },
+        {
+          "type": "null"
+        }
+      ],
+      "default": null,
+      "title": "Default Adapter Ref"
+    },
+    "gpu": {
+      "anyOf": [
+        {
+          "maxLength": 32,
+          "type": "string"
+        },
+        {
+          "type": "null"
+        }
+      ],
+      "default": null,
+      "title": "Gpu"
+    },
     "handler_ref": {
       "title": "Handler Ref",
       "type": "string"
@@ -409,6 +515,19 @@
       },
       "title": "Suite Of",
       "type": "object"
+    },
+    "timeout_s": {
+      "anyOf": [
+        {
+          "exclusiveMinimum": 0,
+          "type": "number"
+        },
+        {
+          "type": "null"
+        }
+      ],
+      "default": null,
+      "title": "Timeout S"
     }
   },
   "required": [
@@ -420,5 +539,5 @@
   ],
   "title": "BenchmarkRegisterRequest",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/BenchmarkSpec.json

@@ -207,7 +207,7 @@
       ],
       "title": "SeqFlags",
       "type": "integer",
-      "x-contract-version": "0.4.4",
+      "x-contract-version": "0.4.6",
       "x-flag-names": {
         "CHECK": 65536,
         "INFER": 16,
@@ -397,5 +397,5 @@
   ],
   "title": "BenchmarkSpec",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/BenchmarkStatusResponse.json

@@ -40,5 +40,5 @@
   ],
   "title": "BenchmarkStatusResponse",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/BootstrapResponse.json

@@ -72,5 +72,5 @@
   ],
   "title": "BootstrapResponse",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/Camera.json

@@ -21,5 +21,5 @@
   ],
   "title": "Camera",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/ClampInfo.json

@@ -23,5 +23,5 @@
   ],
   "title": "ClampInfo",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/CodeBundle.json

@@ -23,5 +23,5 @@
   ],
   "title": "CodeBundle",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/Concurrency.json

@@ -17,5 +17,5 @@
   ],
   "title": "Concurrency",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/Conditions.json

@@ -55,5 +55,5 @@
   ],
   "title": "Conditions",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/ConnectRequest.json

@@ -25,5 +25,5 @@
   ],
   "title": "ConnectRequest",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/ConnectResponse.json

@@ -221,5 +221,5 @@
   ],
   "title": "ConnectResponse",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }

{sequence_base-0.3.5 → sequence_base-0.3.7}/sequence_base/schemas/DeploymentAccepted.json

@@ -18,5 +18,5 @@
   ],
   "title": "DeploymentAccepted",
   "type": "object",
-  "x-contract-version": "0.4.4"
+  "x-contract-version": "0.4.6"
 }