Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 6 additions & 2 deletions packages/learning/environments/brand-check/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -200,11 +200,15 @@ export REPLICA_TRUSTED_MODEL_DIRS="$(pwd)/models/d1-omni-600M"
.venv/bin/replica-decision-runtime \
--environment development \
--tenant local --repository github/example \
--taskset /path/to/taskset --state /path/to/runtime-state \
--state /path/to/runtime-state \
--d1-model "$(pwd)/models/d1-omni-600M" --port 8010
```

Keep the `.cache/huggingface/download` metadata when moving this directory. The runtime verifies every pinned artifact's metadata and SHA-256 digest before loading model code or weights. Changing the base-model revision invalidates existing adapters; retrain and explicitly promote a candidate built against the new pin.
Upload tasksets through the authenticated route before submitting a D1 job. Starting with `--taskset /path/to/taskset` pins D1 training and inference to that taskset's digest and learning text; it is not only a default for jobs that omit `tasksetDigest`.

New D1 adapters carry the taskset digest and learning-text digests needed for upload-only inference. An adapter created before this contract needs its original `--taskset` mount for the learning-text check. Before removing that mount, train and explicitly promote a new adapter from the uploaded taskset.

Keep the `.cache/huggingface/download` metadata when moving the model directory. The runtime verifies every pinned artifact's metadata and SHA-256 digest before loading model code or weights. Changing the base-model revision invalidates existing adapters; retrain and explicitly promote a candidate built against the new pin.

Submit a real job with `"backend":"d1-omni"`. The runtime applies the same update and wall-time bounds, operation scope, taskset pin, durable lease, idempotency, cancellation, candidate evidence, promotion, and usage reporting as the fixture path. The local model backend needs no external credential, so the runtime does not mint one for its training or inference calls.

Expand Down
129 changes: 110 additions & 19 deletions packages/learning/environments/brand-check/d1_connector.py
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@
import io
import os
import random
import re
import time
import unicodedata
from pathlib import Path
Expand Down Expand Up @@ -182,7 +183,7 @@ def __init__(
self,
model_directory: str | Path,
*,
expected_taskset_digest: str,
expected_taskset_digest: str | None = None,
expected_learning_texts: Mapping[str, str] | None = None,
part: str = "readout",
tasks_per_update: int = 16,
Expand Down Expand Up @@ -216,6 +217,14 @@ def readiness(self) -> dict[str, Any]:

def project_request(self, request: Mapping[str, Any]) -> dict[str, Any]:
"""Keep only the model input and retain learning text inside its data field."""
return self._project_request(request, self.expected_learning_texts)

def _project_request(
self,
request: Mapping[str, Any],
expected_learning_texts: Mapping[str, str] | None,
expected_learning_digests: Mapping[str, str] | None = None,
) -> dict[str, Any]:
if len(request["images"]) > MAX_IMAGES:
raise ConnectorError(
"image_count_exceeded",
Expand Down Expand Up @@ -245,19 +254,29 @@ def project_request(self, request: Mapping[str, Any]) -> dict[str, Any]:
"Each d1-omni yes-or-no question needs non-empty learning text.",
category="invalid_input",
)
if self.expected_learning_texts is None:
raise ConnectorError(
"learning_contract_unavailable",
"The d1-omni backend has no configured learning contract.",
category="unavailable",
)
expected_learning = self.expected_learning_texts.get(key)
if expected_learning is None or projected["learning_text"] != expected_learning:
raise ConnectorError(
"learning_text_mismatch",
"The d1-omni question does not match the configured taskset learning.",
category="invalid_input",
)
if expected_learning_texts is not None:
expected_learning = expected_learning_texts.get(key)
if (
expected_learning is None
or projected["learning_text"] != expected_learning
):
raise ConnectorError(
"learning_text_mismatch",
"The d1-omni question does not match the trained taskset learning.",
category="invalid_input",
)
elif expected_learning_digests is not None:
expected_digest = expected_learning_digests.get(key)
if (
expected_digest is None
or _learning_text_digest(projected["learning_text"])
!= expected_digest
):
raise ConnectorError(
"learning_text_mismatch",
"The d1-omni question does not match the trained taskset learning.",
category="invalid_input",
)
projected["instructions"] = NOUL_INSTRUCTIONS
questions[key] = projected
projected_request = {
Expand Down Expand Up @@ -289,7 +308,10 @@ def train_with_context(
operation = context
self._check_ready(operation, training=True)
actual_taskset_digest = str(taskset_digest(taskset))
if actual_taskset_digest != self.expected_taskset_digest:
if (
self.expected_taskset_digest is not None
and actual_taskset_digest != self.expected_taskset_digest
):
raise ConnectorError(
"taskset_scope_mismatch",
"The configured d1-omni backend belongs to another taskset.",
Expand All @@ -307,10 +329,15 @@ def train_with_context(
str(error),
category="invalid_input",
) from None
learning_texts = taskset_learning_texts(taskset)
project_request = lambda request: self._project_request(
request,
learning_texts,
)
_validate_task_requests(
taskset,
[*train_ids, *dev_ids, *test_ids],
self.project_request,
project_request,
)
_validate_training_questions(taskset, train_ids)
work_directory.mkdir(parents=True, exist_ok=True, mode=0o700)
Expand All @@ -319,7 +346,7 @@ def train_with_context(
torch = d1.torch
torch.manual_seed(self.seed)
rng = random.Random(self.seed)
environment = _LocalEnvironment(taskset, self.project_request)
environment = _LocalEnvironment(taskset, project_request)
policy = d1.Policy(self.model_directory, self.part)
baseline_predictions = _predictions(
d1, policy, environment, test_ids, operation
Expand Down Expand Up @@ -431,6 +458,7 @@ def train_with_context(
best_state,
part=self.part,
taskset_digest_value=actual_taskset_digest,
learning_text_digests=_learning_text_digests(learning_texts),
)
capabilities = self.readiness()
return TrainingResult(
Expand Down Expand Up @@ -478,6 +506,18 @@ def predict_with_context(
category="integrity",
) from None
_validate_adapter(saved, self.expected_taskset_digest)
learning_text_digests = _adapter_learning_text_digests(saved)
if learning_text_digests is None and self.expected_learning_texts is None:
raise ConnectorError(
"candidate_adapter_mismatch",
"The promoted adapter does not contain its trained learning contract.",
category="integrity",
)
request = self._project_request(
request,
self.expected_learning_texts,
learning_text_digests,
)
_check_operation(context)
policy = d1.Policy(self.model_directory, saved["part"])
policy.load(saved["state"])
Expand Down Expand Up @@ -843,6 +883,7 @@ def _adapter_bytes(
*,
part: str,
taskset_digest_value: str,
learning_text_digests: Mapping[str, str],
) -> bytes:
output = io.BytesIO()
torch.save(
Expand All @@ -851,6 +892,7 @@ def _adapter_bytes(
"baseModelId": BASE_MODEL_ID,
"baseModelRevision": BASE_MODEL_REVISION,
"tasksetDigest": taskset_digest_value,
"learningTextDigests": dict(sorted(learning_text_digests.items())),
"part": part,
"state": dict(state),
},
Expand All @@ -861,16 +903,31 @@ def _adapter_bytes(

def _validate_adapter(
saved: Any,
expected_taskset_digest: str,
expected_taskset_digest: str | None,
) -> None:
taskset_digest_value = saved.get("tasksetDigest") if isinstance(saved, dict) else None
if (
not isinstance(saved, dict)
or saved.get("schemaVersion") != ADAPTER_SCHEMA
or saved.get("baseModelId") != BASE_MODEL_ID
or saved.get("baseModelRevision") != BASE_MODEL_REVISION
or saved.get("tasksetDigest") != expected_taskset_digest
or not isinstance(taskset_digest_value, str)
or (
(
expected_taskset_digest is not None
and taskset_digest_value != expected_taskset_digest
)
or (
expected_taskset_digest is None
and not re.fullmatch(r"sha256:[0-9a-f]{64}", taskset_digest_value)
)
)
or saved.get("part") not in ("readout", "head")
or not isinstance(saved.get("state"), dict)
or (
"learningTextDigests" in saved
and not _valid_learning_text_digests(saved["learningTextDigests"])
)
):
raise ConnectorError(
"candidate_adapter_mismatch",
Expand All @@ -879,6 +936,40 @@ def _validate_adapter(
)


def _adapter_learning_text_digests(
saved: Mapping[str, Any],
) -> Mapping[str, str] | None:
stored = saved.get("learningTextDigests")
return stored if _valid_learning_text_digests(stored) else None


def _valid_learning_text_digests(value: Any) -> bool:
return (
isinstance(value, dict)
and bool(value)
and all(
isinstance(key, str)
and bool(key)
and isinstance(digest, str)
and bool(re.fullmatch(r"sha256:[0-9a-f]{64}", digest))
for key, digest in value.items()
)
)


def _learning_text_digests(
learning_texts: Mapping[str, str],
) -> dict[str, str]:
return {
key: _learning_text_digest(text)
for key, text in learning_texts.items()
}


def _learning_text_digest(text: str) -> str:
return f"sha256:{hashlib.sha256(text.encode()).hexdigest()}"


def _uniform_answer(question: Mapping[str, Any]) -> dict[str, Any]:
if question["type"] == "choice":
criteria = question["criteria"]
Expand Down
19 changes: 11 additions & 8 deletions packages/learning/environments/brand-check/runtime_server.py
Original file line number Diff line number Diff line change
Expand Up @@ -121,11 +121,11 @@ def do_GET(self) -> None:
return
if path == "/ready":
try:
self.server.runtime.ready()
readiness = self.server.runtime.ready()
except RuntimeFailure as error:
self._send(503, {"status": "not_ready", "reason": error.code})
else:
self._send(200, {"status": "ready"})
self._send(200, readiness)
return
self._authorize()
if path == "/v1/promotion":
Expand Down Expand Up @@ -360,12 +360,15 @@ def main() -> None:
os.environ.pop(TOKEN_VARIABLE, None)
backends: dict[str, Any] = {"fixture": FixtureBackend()}
if args.d1_model is not None:
taskset = load_taskset(args.taskset)
backends["d1-omni"] = D1OmniBackend(
args.d1_model,
expected_taskset_digest=str(taskset_digest(taskset)),
expected_learning_texts=taskset_learning_texts(taskset),
)
if args.taskset is None:
backends["d1-omni"] = D1OmniBackend(args.d1_model)
else:
taskset = load_taskset(args.taskset)
backends["d1-omni"] = D1OmniBackend(
args.d1_model,
expected_taskset_digest=str(taskset_digest(taskset)),
expected_learning_texts=taskset_learning_texts(taskset),
)
runtime = LocalRuntime(
args.state,
args.taskset,
Expand Down
Loading
Loading