[TRTLLM-12842][feat] Maximal LLMAPI capture in usage telemetry

.github/CODEOWNERS

@@ -298,3 +298,8 @@ docs/source/performance/perf-benchmarking.md @NVIDIA/trtllm-bench-reviewers
 # of the NVIDIA/trt-llm-release-branch-approval team, regardless of who else approves the PR.
 # Without approval from a member of this team, PRs cannot be merged to release branches.
 # * @NVIDIA/trt-llm-release-branch-approval
+
+### Telemetry / privacy review
+# Golden manifest is the privacy-review artifact; route it and the usage package to the privacy owner.
+/tensorrt_llm/usage/llm_args_golden_manifest.json @NVIDIA/trt-llm-oss-compliance
+/tensorrt_llm/usage/ @NVIDIA/trt-llm-oss-compliance

README.md

@@ -298,9 +298,10 @@ Deprecation is used to inform developers that some APIs and tools are no longer
 TensorRT-LLM collects anonymous telemetry data by default. This data is used
 in aggregate to understand usage patterns and prioritize engineering efforts.
 **This data cannot be traced back to any individual user.** No prompts,
-user-identifying information, or persistent identifiers are collected. Any
-deployment identifiers are ephemeral, randomly generated per deployment, and
-not linked to users. The data we collect includes:
+outputs, model weights, model paths, tokenizer paths, user-identifying
+information, raw free-form configuration strings, or persistent identifiers are
+collected. Any deployment identifiers are ephemeral, randomly generated per
+deployment, and not linked to users. The data we collect includes:
 
 - Ingress point (e.g., LLM API, CLI, serve command)
 - Deployment duration (via periodic heartbeats)
@@ -309,8 +310,10 @@ not linked to users. The data we collect includes:
 - Parallelism configuration (TP/PP/CP/MoE-EP/MoE-TP sizes), quantization algorithm, dtype, KV cache dtype
 - System information (OS platform, Python version, CPU architecture, CPU count)
 - TRT-LLM version and backend
-- Feature flags (LoRA, speculative decoding, prefix caching, CUDA graphs, chunked context, data parallelism)
+- Feature summary flags (LoRA, speculative decoding, prefix caching, CUDA graphs, chunked context, data parallelism)
 - Disaggregated serving metadata (role and deployment ID)
+- Selected LLM API configuration values: parallelism, dtype, KV cache, scheduler, CUDA graph, and compile settings
+- Capture diagnostics for that payload: a schema checksum (for provenance), the count of captured fields, and whether any free-form value was skipped
 
 Telemetry is automatically disabled in CI and test environments.
 

docs/source/_ext/llmapi_config_telemetry.py

@@ -0,0 +1,102 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+_GOLDEN_REL = "tensorrt_llm/usage/llm_args_golden_manifest.json"
+
+_REFERENCE_PREAMBLE = """\
+# Telemetry
+
+This page documents TensorRT-LLM usage telemetry. It is generated during the
+Sphinx docs build by rendering the committed telemetry manifest
+(`tensorrt_llm/usage/llm_args_golden_manifest.json`).
+
+Start with the
+[Telemetry Data Collection section in the root README](source:README.md#telemetry-data-collection)
+for the user-facing collection and opt-out overview, and the
+[telemetry schema reference](source:tensorrt_llm/usage/schemas/README.md)
+for the wire schema.
+
+**No PII or free-form fields are captured.** LLM API configuration capture is
+*type-driven*: fields whose type is categorical (`Literal`/`Enum`/`bool`) or
+numeric (`int`/`float`), plus safe collections of those, are captured
+automatically. Free-form `str`/`Any`/`Path`/`dict`/`Callable` are never captured
+unless a field carries an explicit allowlist (`TelemetryField.categorical(...)`),
+and any field may opt out with `telemetry=False`. Every captured field is listed
+below; the runtime can capture nothing absent from this list.
+
+## LLM API Configuration Fields
+
+A field can still be absent from a specific payload when its parent config is
+unset or when the safety sanitizer rejects the runtime value.
+"""
+
+
+def _escape(text: str) -> str:
+    return text.replace("|", "\\|").replace("\n", " ")
+
+
+def _format_values(values: list[str]) -> str:
+    return ", ".join(f"`{_escape(v)}`" for v in values) if values else ""
+
+
+def _table(rows: list[dict]) -> str:
+    lines = [
+        "| Captured key | Annotation | Kind | Converter | Allowed values |",
+        "|--------------|------------|------|-----------|----------------|",
+    ]
+    for row in rows:
+        lines.append(
+            f"| `{_escape(row['path'])}` | `{_escape(row['annotation'])}` | "
+            f"`{_escape(row['kind'])}` | {_escape(row['converter'])} | "
+            f"{_format_values(row['allowed_values'])} |"
+        )
+    return "\n".join(lines)
+
+
+def generate_telemetry_reference(repo_root: Path | str, output_path: Path | str) -> None:
+    repo_root = Path(repo_root)
+    golden = json.loads((repo_root / _GOLDEN_REL).read_text())
+    content = [_REFERENCE_PREAMBLE]
+    for args_class in ("TorchLlmArgs", "TrtLlmArgs"):
+        rows = golden.get(args_class, [])
+        content.extend(
+            [
+                f"### `{args_class}`",
+                "",
+                f"{len(rows)} captured fields.",
+                "",
+                _table(rows),
+                "",
+            ]
+        )
+    output = Path(output_path)
+    output.parent.mkdir(parents=True, exist_ok=True)
+    output.write_text("\n".join(content))
+
+
+def _on_builder_inited(app) -> None:
+    docs_source = Path(app.confdir)
+    repo_root = docs_source.parents[1]
+    generate_telemetry_reference(repo_root, docs_source / "developer-guide/telemetry.md")
+
+
+def setup(app) -> dict[str, object]:
+    app.connect("builder-inited", _on_builder_inited)
+    return {"version": "0.2", "parallel_read_safe": True, "parallel_write_safe": True}

docs/source/conf.py

@@ -1,3 +1,6 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
 # Configuration file for the Sphinx documentation builder.
 #
 # For the full list of built-in configuration values, see the documentation:
@@ -68,6 +71,7 @@
     'sphinx_togglebutton',
     'sphinxcontrib.mermaid',
     'trtllm_auto_deploy',
+    'llmapi_config_telemetry',
     'trtllm_config_selector',
 ]