Add versioned submission validation to inference_endpoint by nv-alicheng · Pull Request #532 · mlcommons/endpoints

nv-alicheng · 2026-10-08T12:28:18Z

What does this PR do?

Add inference_endpoint.validation to validate submission artifacts within the endpoints project. Validation first classifies the submission, systems, Pareto curves, and measurement points, then selects and executes the applicable checks. Missing or invalid prerequisites produce blocking findings rather than a passing result.

Load the 2026-10-C1 revision 1 policy from five files: catalog.yaml, submission_checks.yaml, system_checks.yaml, curve_checks.yaml, and point_checks.yaml. Cohort/revision parser intervals support schema evolution without a redundant schema-version field.
Retain typed requirement models through loading, exceptions, planning, and evaluation. Callable evaluator classes register by enum kind; YAML supplies conditions, exclusions, evidence references, thresholds, and model-specific exceptions.
Cover artifact structure and disclosures, system-name consistency, seeds and submission flags, minimum runtime and sample accounting, accuracy and point coverage, client/checkpoint approvals, steady-state diagnostics, and provisioned/per-point power.
Expose SubmissionChecker, validate_submission, and inference-endpoint-validation --submission DIRECTORY, with structured reports and failure exit codes. Online database submission is outside this package.
Support speculative decoding identities by approved repository/revision or weight checksum, and by target checksum plus exact configuration. Configuration comparison preserves scalar types recursively: true, 1, and 1.0 are distinct; non-finite values cannot match. Approval lead time applies to both forms.
Normalize steady-state status/verdict values with enums and document the architecture, evaluator families, and policy rules. Submission fixtures are synthetic and use neutral system names.

Policy status

The bundled policy is a proposed catalog, not an official MLCommons publication. approved_client_revisions is intentionally empty until published client SHAs are supplied, so client approval currently blocks. The catalog includes six weight-based speculative-head approvals; no configuration-based heads have been officially approved. Seeds use the published cohort catalog, and approval catalogs cannot be replaced through CLI or environment overrides.

Type of change

Bug fix
New feature
Documentation update
Refactor/cleanup

Related issues

No linked issue.

Testing

Tests added/updated
All tests pass locally
Manual testing completed
Standalone Linux pre-commit hooks passed for all files changed by this PR, including mypy across 396 source files. Host commit hooks were disabled after the standalone check.
All 471 validation and steady-state tests passed in Linux.
Wheel and source distribution built successfully; the wheel contains all five policy YAML files and the new identity-comparison modules, with local helpers excluded.
The full macOS suite stops at the existing Linux-only CPU-affinity integration test (pin_loadgen rejects Darwin); the all-tests checkbox remains unchecked. Repository-wide lint also has unrelated baseline failures outside this PR.
Review covered behavioral mapping against the submission checker and targeted malformed-artifact/identity probes.

Checklist

Code follows project style
Pre-commit hooks pass
Documentation updated (if needed)

…identities

github-actions · 2026-10-08T12:28:28Z

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

nvzhihanj · 2026-10-08T23:17:42Z

+uv run inference-endpoint-validation path/to/2026-10-C1 --submission submissions/acme/submission-1
+```
+
+The API accepts a loaded `Policy` or policy directory through its `policy` keyword.


The flow described here seems to be no connection to the flowchart above. Consider either using what's here to rewrite the flowchart or the other way (depending on how the code is written)

nvzhihanj · 2026-10-08T23:18:48Z

+failed derivations cannot satisfy dependent checks. Check selection and compliance
+evaluation are separate stages.
+
+Checks register by `CheckKind` through `Evaluator.__init_subclass__`. An evaluator


I think it's really hard to read this README and I would recommend a bit human review here. E.g. you just need to walk through the checking logics and components, using e.g. bullet list and some psuedo code. Right now everything is bundled in a paragraph

nvzhihanj · 2026-10-08T23:19:26Z

+| point_checks.yaml      | Measurement point checks                                                     |
+
+All files declare the same cohort `version` and positive integer `revision`.
+The directory name matches the cohort. There is no top-level `kind` or


What does this mean - There is no top-level kindorschema_version``?

nvzhihanj · 2026-10-08T23:20:11Z

+
+`BundleParser` is a callable abstract class registered through `__init_subclass__`.
+Each parser declares a cohort and inclusive revision interval.
+`schemas/requirements_v1.py` defines a typed contract for each check kind,


schema is something new here again - maybe define a directory structure above will help understand

nvzhihanj · 2026-10-08T23:20:44Z

+planning and execution; evaluators read typed attributes. Overrides are revalidated
+against the same contract. Override patches contain only changed fields; each fully
+merged rule is validated when loading and when planning. `operations.py` shares
+operation and mode enums between


Using examples would be much better. It's really hard for me to visualize what's happening here

nvzhihanj · 2026-10-08T23:21:35Z

+Each input file is read and parsed once. `ParsedArtifact.from_json` retains the
+typed value, supplied field names, and structural errors; it discards the input
+document. Supplied field names distinguish omissions from schema defaults.
+Reported metric aliases remain separate from calculated metrics. Accuracy scores
+accept finite numbers and finite numeric strings; booleans, malformed values,
+and explicitly empty scores produce structural errors. Supplied non-null decode
+head declarations select approval checks even when identity is incomplete.
+Cooling uses the shared `Cooling` enum, including `mixed`, and requires a
+matching policy overhead. Artifact metrics must be finite. Calculated metrics are
+checked for finiteness even when no stored metric is available. Cyclic aliases and
+structures deeper than 100 levels produce artifact errors; shared aliases are valid.
+
+Comparisons accept typed field, constant, and sum operands. A field operand may
+specify a numeric default for an absent value. Sample accounting compares the sum
+of completed and failed samples with issued samples; omitted failed counts default
+to zero. System names must agree within each submitted system.


I really couldn't digest this... please help re-write and self review a little bit.

nvzhihanj · 2026-10-08T23:23:00Z

@@ -0,0 +1,280 @@
+# Cohort policy coverage


Consider merge the Design.md and readme.md

nvzhihanj · 2026-10-08T23:23:28Z

+The artifact validation API loads the five-file cohort policy, classifies the
+submission, plans applicable checks, and executes registered evaluators. Its
+regression suite covers artifact parsing, disclosures, seeds, accuracy, metric
+consistency, coverage, and provisioned/per-point power.
+
+Check selection uses typed classifications and prerequisites. Ready means selected
+with available prerequisites; evaluators determine compliance. Requirement field
+names and references are validated by the versioned policy parser.


Use bullet points or table would be easier to read/understand

nvzhihanj · 2026-10-08T23:25:53Z

+    INSUFFICIENT = "insufficient"
+
+
+_LEGACY_TREND_VERDICTS = {


Can we use the enum instead of using this (since you are already refactoring)?

nvzhihanj · 2026-10-08T23:26:12Z

+    if not isinstance(global_trend, dict):
+        return raw
+    normalized = {
+        key: _LEGACY_TREND_VERDICTS.get(value, value)


You seem to be only using it here

nvzhihanj · 2026-10-08T23:27:19Z

+
+
+@dataclass(frozen=True)
+class SeedSet:


Is this the right place to define this? @arekay-nv

I would think it was defined elsewhere in the repo already

nvzhihanj · 2026-10-08T23:28:19Z

+# derived rather than listed.
+#
+
+cohort:


I am okay with putting it here, @arekay-nv are we okay with making this the only source of truth of the repo?

nvzhihanj · 2026-10-08T23:29:12Z

+from ..types import SampleUnit
+from ..vocabulary import CheckKind
+
+


Some top level docstring would be helpful

nvzhihanj · 2026-10-09T04:06:52Z

+from .helpers import finding, unsupported, warning
+
+
+class SeedBindingEvaluator(Evaluator, kind=CheckKind.SEED_BINDING):


I felt this is a bit overly defensive and complicated coding

nvzhihanj · 2026-10-09T04:08:54Z

+from pydantic import JsonValue, RootModel, model_validator
+
+
+class AccuracyResult(RootModel[dict[str, dict[str, JsonValue]]]):


Do we need modularized checker for in-line accuracy, OSL, swebench accuracy and other group normalized accuracies?

Slightly hard to read at the current moment, @nv-alicheng please help use your Python judgement to help agent design

nvzhihanj · 2026-10-09T04:10:04Z

+from typing import cast
+
+from pydantic import JsonValue, RootModel, model_validator
+


Docstring

And btw, is evidence the right name for the directory? I felt it's more like report or format

nvzhihanj · 2026-10-09T04:11:48Z

+        inline_minimum_percent: 52.36
+        swebench_mean_minimum_percent: 96.4
+        full_run_osl_range: [793, 970]
+    llama3_1-405b:


This is not a valid model, delete

nvzhihanj · 2026-10-09T04:11:53Z

+              minimum_multiplier: 0.9,
+              maximum_multiplier: 1.1,
+            }
+    llama2-70b:


Remove this model

nvzhihanj · 2026-10-09T04:11:59Z

+              minimum_multiplier: 0.9,
+              maximum_multiplier: 1.1,
+            }
+    mixtral-8x7b:


Stale hallucination

nvzhihanj · 2026-10-09T04:12:21Z

+    - model: kimi-k3
+      repository: RadixArk/Kimi-K3-DSpark
+      revision: 3c5bac301d9cf392706189d82ed947feca6c2f0f
+      approved_cohort: 2026-09-C1


When you say cohort here, this is the earliest cohort that we can use it right?

nvzhihanj · 2026-10-09T04:12:49Z

+      - repository: nvidia/Qwen3.6-35B-A3B-NVFP4
+        revision: 1355db6a052410cfd62085d94b58866fd0f2c3c5
+      - repository: nvidia/Qwen3.6-35B-A3B-NVFP4
+        revision: 491c2f1ea524c639598bf8fa787a93fed5a6fbce


We need 2 commits for the same repo? @leopck

nvzhihanj · 2026-10-09T04:13:19Z

+    verdicts: [steady_state, drifting_up, drifting_down, anomaly, not_found]
+    metric_states: [plateau, drifting_up, drifting_down]
+    minimum_super_passes: 4
+    minimum_duration_ms:


Agentic will be 10 min for low, 40min for everything else (new rule just discussed this week)
cc: @arekay-nv

nvzhihanj · 2026-10-09T04:14:01Z

+      accelerators:
+        {
+          gb300: 1400,
+          gb200: 1200,
+          b300: 1100,
+          b200: 1000,
+          mi455x: 2500,
+          mi355x: 1400,
+          mi350x: 1000,
+          tpuv7: 1000,
+          tn3: 700,
+        }
+      switches:
+        sn6810ld: { passive: 1960, active_optical: 1960 }
+        sn5610: { passive: 900, active_optical: 2080 }
+        sn5400: { passive: 670 }
+        sn4700: { passive: 630 }
+      nic_watts: 75


These shouldn't be here. It's not a check
cc: @arekay-nv

nvzhihanj · 2026-10-09T04:14:11Z

+  metrics:
+    {
+      relative_tolerance: 0.01,
+      relative_denominator_floor: 1.0e-09,
+      utilization_absolute_tolerance: 0.1,


What's this? Need docstring/comment

nvzhihanj · 2026-10-09T04:16:12Z

+  model-name-valid:
+    kind: membership
+    target: point.model_name
+    catalog: enrollment.models
+    matching: exact
+    skip_missing: true
+    deduplicate: declared_name
+  model-name-consistency:


What is this checking?

nvzhihanj · 2026-10-09T04:16:44Z

+    right: curve.first_declared_model_name
+    comparison: exact
+    on_missing: warning
+  config-consistency-model:


I think you should docstring these checks and let it auto generate documentation (not AI generated)

nvzhihanj · 2026-10-09T04:17:04Z

+    minimum: 7
+    overrides:
+      - when:
+          curve_type: single_turn


Both agentic and single-turn

It should apply to LLM in general

nvzhihanj · 2026-10-09T04:17:26Z

+    kind: offline
+    operation: declaration_count
+    single_turn_count: 1
+    agentic_count: 0


@arekay-nv should this be 0 or 1?

nvzhihanj · 2026-10-09T04:17:48Z

+    single_turn_count: 1
+    agentic_count: 0
+    elected_must_equal: curve.max_supported_concurrency
+  offline-ordering:


What does this mean?
(Maybe it's good to link to the rule doc)

nvzhihanj · 2026-10-09T04:30:57Z

@@ -0,0 +1,192 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Wire models and normalization for 2026-10-C1 revision 1."""


Just to verify - are we saying this v1 only applies to 1 cohort and we would rewrite for every cohort?

nvzhihanj · 2026-10-09T04:31:23Z

@@ -0,0 +1,687 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Strict check requirement contracts for cohort revision 1."""


Might need better docstring and file name

nv-alicheng added 14 commits October 8, 2026 01:31

chore(tooling): align lint hooks and preserve copyright notices

19a3495

refactor(metrics): normalize steady-state verdicts with enums

d286fe1

feat(validation): load versioned cohort policies and plan typed checks

65996db

feat(validation): calculate policy-driven system power

2744459

feat(validation): parse owned artifacts and classify submission runs

a820fc6

feat(validation): execute submission checks with callable evaluators

920d57f

feat(validation): expose submission-check command and documentation

2ef463f

docs(validation): clarify conditions and evidence as bullet points

a2b9701

docs(validation): explain evaluator kinds and group rules by scope

aeaa68f

docs(validation): group evaluator kinds and their shared operations

7eab5c1

refactor(validation): consolidate count checks and clarify rule names

ff433f6

docs(validation): rename implementation document to DESIGN.md

8b9538f

test(validation): use neutral names for synthetic submission systems

4f9e35b

refactor(validation): retain typed requirements and enforce approval …

131fe1b

…identities

nv-alicheng requested a review from a team October 8, 2026 12:28

github-actions Bot requested review from arekay-nv and nvzhihanj October 8, 2026 12:28

github-actions Bot added the size/very-large PR Review Policy: >1500 lines or >50 files label Oct 8, 2026

test(validation): generate submission artifacts from compact cases

f4250d7

github-code-quality Bot found potential problems Oct 8, 2026

View reviewed changes

nv-alicheng added 3 commits October 8, 2026 14:00

test(validation): consolidate repeated setup and seed cases

0a469ab

refactor(validation): remove unused helpers and consolidate registration

67536ba

style(validation): use pass for empty overload bodies

615ea76

nvzhihanj reviewed Oct 8, 2026

View reviewed changes

nvzhihanj reviewed Oct 9, 2026

View reviewed changes

nvzhihanj requested changes Oct 9, 2026

View reviewed changes

		from ..types import SampleUnit
		from ..vocabulary import CheckKind

		from .helpers import finding, unsupported, warning


		class SeedBindingEvaluator(Evaluator, kind=CheckKind.SEED_BINDING):

		from pydantic import JsonValue, RootModel, model_validator


		class AccuracyResult(RootModel[dict[str, dict[str, JsonValue]]]):

		from typing import cast

		from pydantic import JsonValue, RootModel, model_validator

Conversation

nv-alicheng commented Oct 8, 2026

What does this PR do?

Policy status

Type of change

Related issues

Testing

Checklist

Uh oh!

github-actions Bot commented Oct 8, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment