[TRTLLM-12842][feat] Maximal LLMAPI capture in usage telemetry#14398
Conversation
|
/bot kill |
|
/bot run |
|
PR_Github #49757 [ kill ] triggered by Bot. Commit: |
|
PR_Github #49757 [ kill ] completed with state |
|
PR_Github #49758 [ run ] triggered by Bot. Commit: |
|
PR_Github #49758 [ run ] completed with state
|
|
/bot kill |
|
/bot run |
|
PR_Github #49763 [ kill ] triggered by Bot. Commit: |
|
PR_Github #49763 [ kill ] completed with state |
|
PR_Github #49764 [ run ] triggered by Bot. Commit: |
|
/bot kill |
|
/bot run |
|
PR_Github #49771 [ kill ] triggered by Bot. Commit: |
|
PR_Github #49772 [ run ] triggered by Bot. Commit: |
|
PR_Github #49771 [ kill ] completed with state |
|
PR_Github #49764 [ run ] completed with state |
|
/bot kill |
|
/bot run |
|
PR_Github #49778 [ run ] triggered by Bot. Commit: |
|
PR_Github #49779 [ kill ] triggered by Bot. Commit: |
|
PR_Github #49778 [ run ] completed with state |
|
PR_Github #49772 [ run ] completed with state |
|
PR_Github #49779 [ kill ] completed with state |
|
/bot kill |
|
/bot run |
|
PR_Github #49781 [ run ] triggered by Bot. Commit: |
|
PR_Github #49782 [ kill ] triggered by Bot. Commit: |
|
PR_Github #49781 [ run ] completed with state |
|
/bot run |
|
PR_Github #53714 [ run ] triggered by Bot. Commit: |
|
PR_Github #53714 [ run ] completed with state
|
|
/bot run |
|
PR_Github #53908 [ run ] triggered by Bot. Commit: |
|
PR_Github #53908 [ run ] completed with state
|
Capture schema-stable LLM API config telemetry from the type-driven manifest. Add the generated manifest golden, docs updates, and runtime/config capture test coverage. Signed-off-by: venkywonka <23023424+venkywonka@users.noreply.github.com>
Signed-off-by: venkywonka <23023424+venkywonka@users.noreply.github.com>
b9b771d to
010df91
Compare
|
/bot run |
|
PR_Github #54010 [ run ] triggered by Bot. Commit: |
|
PR_Github #54010 [ run ] completed with state
|
|
/bot run |
|
PR_Github #54031 [ run ] triggered by Bot. Commit: |
|
PR_Github #54031 [ run ] completed with state
|
|
/bot run |
|
PR_Github #54034 [ run ] triggered by Bot. Commit: |
|
PR_Github #54034 [ run ] completed with state
|
|
/bot run |
|
PR_Github #54039 [ run ] triggered by Bot. Commit: |
|
PR_Github #54039 [ run ] completed with state
|
|
/bot run |
|
PR_Github #54040 [ run ] triggered by Bot. Commit: |
|
PR_Github #54040 [ run ] completed with state
|
|
/bot run |
|
PR_Github #54044 [ run ] triggered by Bot. Commit: |
|
PR_Github #54044 [ run ] completed with state |
Motivation
Usage telemetry had coarse feature flags, not effective LLM API runtime-shape config. Needed: know how deployed configs are shaped without collecting private or user-controlled data.
Hard rule: no PII, no prompts, no outputs, no model/tokenizer paths, no env vars, no raw YAML, no arbitrary fields, no arbitrary strings.
Design
Full list and related TRTLLM documentation preview: https://venkywonka.github.io/TensorRT-LLM/developer-guide/telemetry.html
Field(..., telemetry=...)opt-in metadata for approved LLM args fields.llmApiConfigJsonfor sanitized config values.llmApiConfigMetaJsonwith capture version, policy version, args class, digests, counts, source, and unsafe-exclusion status.Any, callables, unsafe unions, and heterogeneous tuples.featuresJsonfor legacy summary consumers. Add TODO to dedupe withllmApiConfigJsonlater.llm_args.pyannotations so exhaustive captured-key docs do not drift manually.Testing
python3 -m json.tool tensorrt_llm/usage/schemas/trtllm_usage_event_schema.jsonpython3 -m py_compileon telemetry usage modules and docs generatorruff check/ruff format --checkon touched Python filesgit diff --check199 passed, 5 skipped2 passed2 passed54 passedstatus=readySummary by CodeRabbit
Documentation
New Features
Tests