Dingo router upstream sync

.ai/ais-perf-model-guidelines.md

@@ -0,0 +1,39 @@
+<!--
+SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+SPDX-License-Identifier: Apache-2.0
+-->
+
+# AISimulate performance-model integration
+
+All new performance-model features must enter through AISimulate's canonical
+`ForwardPassPerfModelConfig` and `ForwardPassPerfModel::best_available(config)`.
+Python uses `aisimulate_core.sdk.RustForwardPassPerfModel.best_available(config)`.
+Do not add parallel constructors, revive removed `from_native` / `from_regression`
+APIs, or construct a private native engine to bypass this interface.
+
+- Use `ais` / `AIS` for Dynamo integration names. Accept old `--aic-*` spellings
+  only in CLI parsers and immediately lower them to the canonical configuration.
+  Do not restore retired SDK aliases, modules, environment variables or config
+  formats. Actual AIConfigurator task-v2/interpolation APIs and upstream-owned
+  wire protocols retain their upstream names at their boundaries.
+- AISimulate owns schema, defaults, validation, selection, tuning and prediction.
+  Pass complete configurations, including ordered roots and nested estimator
+  controls. Reject unknown fields and conflicting old/new input names.
+- New configs default to `estimation_mode: auto`, `fallback_policy: deny`.
+  Auto still searches `op_level`, `fpm_interpolation`, then `fpm_regression`.
+  Explicit selections obey their fallback policy; do not silently relax it.
+- Bind `worker_type` to the deployed worker role, not the current query phase.
+  Preserve model, backend/version, parallelism, quantization and speculation
+  identity across CLI, SDK, Rust, replay, diagnostics, and saved configuration.
+- Planner organizes full configurations under `ais_perf_model.roles` for
+  `prefill`, `decode`, or `aggregated`. Omitted configuration explicitly creates
+  cold regression with traceable model identity. Consumers without a training
+  source must reject an unready regression model.
+- Use upstream migration helpers for legacy controls: 16 total buckets means a
+  `[4, 4]` grid, not `[16, 16]`. Preserve independent regression/correction controls.
+- Include complete configuration and limits in model reuse decisions. Keep
+  online training state local to the worker group/consumer. Rebuild and replay
+  retained observations when identity or limits change.
+- Upgrade Python wheel and Rust crate together from the same released source and
+  rebuild Dynamo bindings. Validate the installed wheel and actual Rust callback;
+  permissive test fakes cannot detect removed APIs or EngineSpec schema mismatches.

.codex/agents/hypothesis-challenger.toml

@@ -0,0 +1,13 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+name = "hypothesis_challenger"
+description = "Adversarially review one generated Dynamo optimization candidate before GPU spend."
+
+developer_instructions = """
+Before acting, read agents/hypothesis-challenger/AGENTS.md from the repository root and follow that role contract,
+including the skills, documentation, rules, inputs, outputs, and stopping conditions it names.
+
+Operate only as the hypothesis challenger. Do not edit the consultation or candidate draft, deploy, run AIPerf,
+generate a competing hypothesis, or approve while a blocking objection remains.
+"""

.codex/agents/hypothesis-generator.toml

@@ -0,0 +1,13 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+name = "hypothesis_generator"
+description = "Generate one evidence-backed Dynamo optimization proposal and challenger-ready DGD draft."
+
+developer_instructions = """
+Before acting, read agents/hypothesis-generator/AGENTS.md from the repository root and follow that role contract,
+including the skills, documentation, rules, inputs, outputs, and stopping conditions it names.
+
+Operate only as the hypothesis generator. Do not deploy, benchmark, approve candidates, change frozen benchmark
+semantics, or present an untested internal mechanism as fact.
+"""

.codex/agents/perf-analyzer.toml

@@ -0,0 +1,13 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+name = "perf_analyzer"
+description = "Configure, run, validate, and analyze AIPerf for one successfully deployed Dynamo candidate."
+
+developer_instructions = """
+Before acting, read agents/perf-analyzer/AGENTS.md from the repository root and follow that role contract, including
+the skills, documentation, rules, inputs, outputs, and stopping conditions it names.
+
+Operate only as the performance analyzer. Do not modify the deployed DGD, change frozen benchmark semantics, generate
+the next hypothesis, or approve a candidate.
+"""

.codex/agents/recipe-deployer.toml

@@ -0,0 +1,13 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+name = "recipe_deployer"
+description = "Deploy one assigned Dynamo DGD, wait for readiness, and produce smoke-test evidence."
+
+developer_instructions = """
+Before acting, read agents/recipe-deployer/AGENTS.md from the repository root and follow that role contract, including
+the skills, documentation, rules, inputs, outputs, and stopping conditions it names.
+
+Operate only as the recipe deployer. Do not search for or substitute a DGD, tune deployment settings, run performance
+benchmarks, generate hypotheses, or approve candidates.
+"""

.codex/agents/user-interviewer.toml

@@ -0,0 +1,17 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+name = "user_interviewer"
+description = "Act first, establish the baseline DGD via the baseline-source ladder, and synthesize the canonical workload contract."
+
+developer_instructions = """
+Before acting, read agents/user-interviewer/AGENTS.md from the repository root and follow that role contract, including
+the skill, documentation, rules, inputs, outputs, and stopping conditions it names.
+
+Operate only as the user interviewer. Use the initial user request and minimal follow-up questions to produce the
+canonical user_workload.yaml and immutable inputs/user_provided_dgd.yaml. The baseline comes from the baseline-source
+ladder in AGENTS.md: capture the DGD the user supplies; or propose an exact or close catalog recipe (a close match
+with an explicit adaptation diff); or invoke author-baseline-dgd to draft one. Every rung requires the user's explicit
+confirmation, and only the confirmed manifest is captured, with its origin recorded in the contract. Never deploy,
+benchmark, or generate optimization hypotheses, and never record an unconfirmed manifest as the baseline.
+"""

.codex/config.toml

@@ -0,0 +1,25 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+[features]
+multi_agent = true
+
+[agents.user_interviewer]
+description = "First optimization role: capture the user's DGD and synthesize the canonical workload contract."
+config_file = "agents/user-interviewer.toml"
+
+[agents.recipe_deployer]
+description = "Deploy one assigned Dynamo DGD and produce smoke-test evidence."
+config_file = "agents/recipe-deployer.toml"
+
+[agents.perf_analyzer]
+description = "Configure, run, validate, and analyze AIPerf for one deployed candidate."
+config_file = "agents/perf-analyzer.toml"
+
+[agents.hypothesis_generator]
+description = "Generate one evidence-backed Dynamo optimization proposal and challenger-ready DGD draft."
+config_file = "agents/hypothesis-generator.toml"
+
+[agents.hypothesis_challenger]
+description = "Adversarially review one generated Dynamo optimization candidate before GPU spend."
+config_file = "agents/hypothesis-challenger.toml"

.dockerignore

@@ -25,12 +25,24 @@
 **/*.libtorch_model_store/*
 **/.git
 **/.github
+# Test images exercise these scripts without copying the rest of the GitHub
+# metadata into every Docker build context.
+!.github
+.github/*
+!.github/scripts
+.github/scripts/*
+!.github/scripts/apply_dev_version.py
+!.github/scripts/retry_kubectl.sh
+!.github/scripts/collect_vcluster_diagnostics.sh
 **/*backup*/
 .dockerignore
 **/target/*
 **/*safetensors
 container/Dockerfile*
 container/**/*.Dockerfile
+# The AISimulate dependency contract test reads this template from runtime test
+# images; keep every other generated/template Dockerfile excluded.
+!container/templates/wheel_builder.Dockerfile
 container/render.py
 container/context.yaml
 # container/compliance/ is now COPY'd into each image's `licenses` Dockerfile

.gitattributes

@@ -1,5 +1,6 @@
 *                   text  eol=lf
 *.sln               text  eol=crlf
+*.[Gg][Ii][Ff]      binary
 *.[Ii][Cc][Oo]      binary
 *.[Jj][Pp][Ee][Gg]  binary
 *.[Jj][Pp][Gg]      binary
@@ -11,7 +12,11 @@
 *.[Gg][Zz]          binary
 *.fatbin            binary
 *.mp4               binary
+*.woff2             binary -eol
 
 # Exclude test data files from linguist language detection
 lib/llm/tests/data/** linguist-vendored
 lib/llm/tests/snapshots/** linguist-vendored
+
+# Preserve the released vLLM proto byte-for-byte, including its trailing blank line.
+lib/backend/vllm-sidecar/proto/vllm_grpc.proto -text diff whitespace=-blank-at-eof

.github/ISSUE_TEMPLATE/agent-reported.yml

@@ -0,0 +1,35 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+name: Agent-reported instruction gap
+description: For AI agents using the optimization skills — report instructions that misled, blocked, or contradicted live verification.
+title: "[AGENT]: "
+# NOTE: the "agent-reported" label must exist in the repository before this template is used —
+# GitHub silently drops unknown labels at issue creation.
+labels: ["agent-reported"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        For AI agents. Search existing agent-reported issues before filing; one issue per
+        optimization session; sanitize all user/engagement specifics.
+  - type: input
+    id: agent-identity
+    attributes:
+      label: Agent identity
+      description: Driver model and the skills commit you were running.
+      placeholder: "claude-opus-4-8, skills @ abc1234"
+    validations:
+      required: true
+  - type: textarea
+    id: gap
+    attributes:
+      label: What the instructions said vs what you verified
+      description: Describe the instruction gap only. No user workload details, traffic numbers, cluster/namespace names, company names, or credentials.
+    validations:
+      required: true
+  - type: textarea
+    id: suggestion
+    attributes:
+      label: Suggested correction
+      description: Optional. The wording you would have needed.

.github/ISSUE_TEMPLATE/cherry_pick_request.yml

@@ -0,0 +1,58 @@
+name: Cherry-Pick Request
+description: Request a merged fix on a release branch
+title: '[Cherry-Pick Request] [release/X.Y.Z]: '
+labels: ["cherry-pick:requested"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Use this form to request a fix on a release branch after code freeze. Get the fix merged to `main` first. Release-only fixes skip this step. The maintainer who merged your PR opens the release PR. For a release-only fix, the release approver assigns a maintainer to open it.
+
+        In the title, replace `X.Y.Z` with the target release, for example `[release/1.6.0]`.
+
+  - type: input
+    id: main-pr
+    attributes:
+      label: Merged Main PR
+      description: Link to the merged PR on `main`. Leave empty only for a release-only fix.
+      placeholder: https://github.com/ai-dynamo/dynamo/pull/12345
+    validations:
+      required: false
+
+  - type: input
+    id: target-release
+    attributes:
+      label: Target Release
+      description: The release this fix ships in, for example 1.6.0.
+      placeholder: "1.6.0"
+    validations:
+      required: true
+
+  - type: dropdown
+    id: type
+    attributes:
+      label: Type
+      options:
+        - Fix
+        - Test
+        - Infra
+        - CI
+        - Dependency
+    validations:
+      required: true
+
+  - type: textarea
+    id: reason
+    attributes:
+      label: Reason
+      description: What breaks in the release without it?
+    validations:
+      required: true
+
+  - type: textarea
+    id: dependencies
+    attributes:
+      label: Dependencies
+      description: Other PRs that must land first, if any.
+    validations:
+      required: false

.github/actionlint.yaml

@@ -0,0 +1,16 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+paths:
+  .github/workflows/**/*.{yml,yaml}:
+    ignore:
+      # actionlint 1.7.12 does not recognize GitHub's self-repository syntax yet.
+      - '^specifying action "\$/\.github/(actions|workflows)/[^"]+" in invalid format because ref is missing\.'
+      - '^reusable workflow call "\$/\.github/workflows/[^"]+" at "uses" is not following the format '
+  .github/actions/**/action.{yml,yaml}:
+    ignore:
+      - '^specifying action "\$/\.github/actions/[^"]+" in invalid format because ref is missing\.'
+
+self-hosted-runner:
+  labels:
+    - prod-builder-v3

.github/actions/changed-files/report.py

@@ -0,0 +1,53 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Report tj-actions JSON file lists and reject files without a CI filter."""
+
+import json
+import os
+from pathlib import Path
+
+
+def load_files(path: Path) -> set[str]:
+    """Read an unescaped JSON array written by tj-actions/changed-files."""
+    filenames = json.loads(path.read_text())
+    if not isinstance(filenames, list) or not all(
+        isinstance(name, str) for name in filenames
+    ):
+        raise ValueError(f"Expected a JSON array of filenames in {path.name}")
+    return set(filenames)
+
+
+def report(output_dir: Path, base_sha: str = "") -> int:
+    """Log filenames as JSON data and check the union of all explicit filters."""
+    all_path = output_dir / "all_all_modified_files.json"
+    all_files = load_files(all_path)
+    print("Base SHA:", json.dumps(base_sha or "default (previous commit)"))
+    print(
+        f"All modified files ({len(all_files)} total):", json.dumps(sorted(all_files))
+    )
+    print("Files matching each filter:")
+    covered: set[str] = set()
+    for path in sorted(output_dir.glob("*_all_modified_files.json")):
+        if path == all_path:
+            continue
+        filenames = load_files(path)
+        filter_name = path.name.removesuffix("_all_modified_files.json")
+        print(f"  {json.dumps(filter_name)}: {json.dumps(sorted(filenames))}")
+        covered.update(filenames)
+
+    uncovered = all_files - covered
+    if uncovered:
+        print("::error::The following files are not covered by any CI filter:")
+        for filename in sorted(uncovered):
+            print(json.dumps(filename))
+        print("Add these paths to .github/filters.yaml. See .github/FILTERS.md.")
+        return 1
+    print("All modified files are covered by CI filters.")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(
+        report(Path(os.environ["CHANGED_FILES_DIR"]), os.environ["BASE_SHA"])
+    )