[None][feat] perf_metrics endpoint functionality improvement #8005
Conversation
nv-yilinf
force-pushed
the
ttft-tracer
branch
from
355db6d to
bc06b8b
Compare
nv-yilinf
changed the title
nv-yilinf
force-pushed
the
ttft-tracer
branch
from
bc06b8b to
6bd99cc
Compare
📝 WalkthroughWalkthroughAdds an optional global steady clock offset across C++ core, Python bindings, and servers. Propagates offset from server/executor to request objects, adjusts timing points, and extends metrics. Introduces executor-side distributed clock sync, server endpoints/middleware for offset and arrival timestamps, and threads metrics through postproc/result paths. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
participant R0 as Rank 0
participant Rn as Other Ranks
participant PE as PyExecutor
participant ERQ as ExecutorRequestQueue
Note over PE: Initialization
PE->>R0: barrier + capture steady_clock now
PE->>Rn: barrier + capture steady_clock now
R0-->>PE: t0
Rn-->>PE: t_n (per-rank)
PE->>PE: compute offsets (t0 - t_n)
PE->>ERQ: construct(..., global_steady_clock_offset=offset_n)
Note over ERQ: Store offset for requests
sequenceDiagram
autonumber
participant Client
participant S as OpenAI Server
participant DS as Disagg Server (optional)
participant PX as PyExecutor
participant Q as ExecutorRequestQueue
participant PyReq as pyexecutor LlmRequest
participant CppReq as C++ LlmRequest
Client->>S: HTTP request
activate S
Note over S: Middleware stamps server_arrival_time (monotonic)
S->>S: apply disagg_server_steady_clock_offset to timings
alt Disaggregated
S->>DS: request (raw_request passed)
DS->>DS: stamp arrival, apply per-server offsets
DS-->>S: streamed/merged responses (+metrics)
end
S->>PX: enqueue generation request (+raw_request metrics)
PX->>Q: push(..., global_steady_clock_offset)
Q->>PyReq: attach py_global_steady_clock_offset
PyReq->>CppReq: construct(arrival_time, globalSteadyClockOffset)
Note over CppReq: use maybeToGlobalSteadyClock/getSteadyClockNow for timing fields
deactivate S
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Pre-merge checks and finishing touches❌ Failed checks (2 warnings)
✅ Passed checks (1 passed)
✨ Finishing touches
🧪 Generate unit tests
Tip 👮 Agentic pre-merge checks are now available in preview!Pro plan users can now enable pre-merge checks in their settings to enforce checklists before merging PRs.
Please see the documentation for more information. Example: reviews:
pre_merge_checks:
custom_checks:
- name: "Undocumented Breaking Changes"
mode: "warning"
instructions: |
Pass/fail criteria: All breaking changes to public APIs, CLI flags, environment variables, configuration keys, database schemas, or HTTP/GraphQL endpoints must be documented in the "Breaking Change" section of the PR description and in CHANGELOG.md. Exclude purely internal or private changes (e.g., code not exported from package entry points or explicitly marked as internal).Please share your feedback with us on this Discord post. Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 8
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (4)
cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.cpp (1)
79-132: Don’t always engage arrivalTime: pass nullopt when unsetYou pass mPerfMetrics.timingMetrics.arrivalTime (a default-constructed TimePoint) unconditionally, which forces arrivalTime to epoch when none was provided. That breaks metrics (e.g., huge E2E). Build an optional and pass nullopt for the default value.
Apply this diff:
- return std::make_shared<tb::LlmRequest>( // + std::optional<tb::LlmRequest::TimePoint> arrivalTimeOpt = std::nullopt; + if (mPerfMetrics.timingMetrics.arrivalTime != tb::LlmRequest::TimePoint{}) { + arrivalTimeOpt = mPerfMetrics.timingMetrics.arrivalTime; + } + return std::make_shared<tb::LlmRequest>( // mRequestId, // mMaxNewTokens, // std::make_shared<std::vector<TokenIdType>>(mTokens.at(0)), // mSamplingConfig, // mIsStreaming, // mEndId, // mPadId, // from_torch(mEmbeddingBias), // from_torch(mBadWordsList), // from_torch(mStopWordsList), // mPositionIds, // from_torch(mPromptEmbeddingTable), // mPromptVocabSize, // mMultimodalHashes, // mMultimodalPositions, // mMultimodalLengths, // from_torch(mMultimodalEmbedding), // from_torch(mMropeRotaryCosSin), // mMropePositionDeltas, // mLoraTaskId, // from_torch(mLoraWeights), // from_torch(mLoraConfig), // mLookaheadConfig, // mKvCacheRetentionConfig, // mReturnLogProbs, // mReturnContextLogits, // mReturnGenerationLogits, // optDraftTokens, // from_torch(mDraftLogits), // mExcludeInputFromOutput, // callbackAdapter(mLogitsPostProcessor), // mApplyLogitsPostProcessorBatched, // optEncoderInputTokens, // mReturnEncoderOutput, // mClientId, // mPriority, // from_torch(mEncoderInputFeatures), // mEncoderOutputLength, // from_torch(mCrossAttentionMask), // getLlmRequestType(), // std::nullopt, // inputTokenExtraIds mNumReturnSequences, // mEagleConfig, // from_torch(mSkipCrossAttnBlocks), // false, // returnPerfMetrics mGuidedDecodingParams, // mLanguageAdapterUid, // mAllottedTimeMs, // mContextPhaseParams, // mCacheSaltID, // - mPerfMetrics.timingMetrics.arrivalTime, // + arrivalTimeOpt, // mGlobalSteadyClockOffset // );cpp/tensorrt_llm/pybind/batch_manager/llmRequest.cpp (1)
79-132: Mirror nanobind fix: guard arrivalTime with std::optionalSame issue here: passing a default-constructed TimePoint always engages arrivalTime. Pass nullopt unless it was explicitly set.
Apply this diff:
- return std::make_shared<tb::LlmRequest>( // + std::optional<tb::LlmRequest::TimePoint> arrivalTimeOpt = std::nullopt; + if (mPerfMetrics.timingMetrics.arrivalTime != tb::LlmRequest::TimePoint{}) { + arrivalTimeOpt = mPerfMetrics.timingMetrics.arrivalTime; + } + return std::make_shared<tb::LlmRequest>( // mRequestId, // mMaxNewTokens, // std::make_shared<std::vector<TokenIdType>>(mTokens.at(0)), // mSamplingConfig, // mIsStreaming, // mEndId, // mPadId, // from_torch(mEmbeddingBias), // from_torch(mBadWordsList), // from_torch(mStopWordsList), // mPositionIds, // from_torch(mPromptEmbeddingTable), // mPromptVocabSize, // mMultimodalHashes, // mMultimodalPositions, // mMultimodalLengths, // from_torch(mMultimodalEmbedding), // from_torch(mMropeRotaryCosSin), // mMropePositionDeltas, // mLoraTaskId, // from_torch(mLoraWeights), // from_torch(mLoraConfig), // mLookaheadConfig, // mKvCacheRetentionConfig, // mReturnLogProbs, // mReturnContextLogits, // mReturnGenerationLogits, // optDraftTokens, // from_torch(mDraftLogits), // mExcludeInputFromOutput, // callbackAdapter(mLogitsPostProcessor), // mApplyLogitsPostProcessorBatched, // optEncoderInputTokens, // mReturnEncoderOutput, // mClientId, // mPriority, // from_torch(mEncoderInputFeatures), // mEncoderOutputLength, // from_torch(mCrossAttentionMask), // getLlmRequestType(), // std::nullopt, // inputTokenExtraIds mNumReturnSequences, // mEagleConfig, // from_torch(mSkipCrossAttnBlocks), // false, // returnPerfMetrics mGuidedDecodingParams, // mLanguageAdapterUid, // mAllottedTimeMs, // mContextPhaseParams, // mCacheSaltID, // - mPerfMetrics.timingMetrics.arrivalTime, // + arrivalTimeOpt, // mGlobalSteadyClockOffset // );cpp/tensorrt_llm/pybind/batch_manager/bindings.cpp (1)
299-369: Fix pybind constructor wiring for clock offsetThe new constructor branch doesn’t compile:
TimePoint::durationis unqualified.- You pass
global_steady_clock_offset, which isn’t defined—the parameter isglobalSteadyClockOffset.- The exposed kwargs use
nb::arg, but this translation unit only haspy::arg.These regressions block the build and the binding can’t expose the new argument. Please wire it up with the qualified type, the correct variable name, and
py::arg.Apply this diff:
- std::optional<TimePoint::duration> globalSteadyClockOffset = std::nullopt) + std::optional<tb::LlmRequest::TimePoint::duration> globalSteadyClockOffset = std::nullopt) ... - language_adapter_uid, allotted_time_ms, context_phase_params, cache_salt_id, arrival_time, global_steady_clock_offset}; + language_adapter_uid, allotted_time_ms, context_phase_params, cache_salt_id, arrival_time, + globalSteadyClockOffset}; ... - nb::arg("arrival_time") = std::nullopt, nb::arg("global_steady_clock_offset") = std::nullopt) + py::arg("arrival_time") = std::nullopt, py::arg("global_steady_clock_offset") = std::nullopt)cpp/tensorrt_llm/pybind/batch_manager/llmRequest.h (1)
89-154: Forward the global offset to the base ctorThe new
globalSteadyClockOffsetparameter stops at this wrapper—theBaseconstructor never receives it, so the pybind path can’t participate in the global clock sync you’re adding elsewhere. Please forward it alongsidearrivalTime.Apply this diff:
- cacheSaltID, // - arrivalTime // + cacheSaltID, // + arrivalTime, // + globalSteadyClockOffset //
🧹 Nitpick comments (6)
cpp/include/tensorrt_llm/batch_manager/llmRequest.h (3)
1697-1715: Make KvCache transfer getters const-correctgetKvCacheTransferStart/End can be const; they just read metrics. Improves API clarity.
Apply this diff:
- TimePoint getKvCacheTransferStart() + TimePoint getKvCacheTransferStart() const { return mPerfMetrics.timingMetrics.kvCacheTransferStart; } - TimePoint getKvCacheTransferEnd() + TimePoint getKvCacheTransferEnd() const { return mPerfMetrics.timingMetrics.kvCacheTransferEnd; }
104-104: Prefer steady_clock::duration aliasSimplify the alias; both are equivalent but this is clearer.
- using Duration = std::chrono::time_point<std::chrono::steady_clock>::duration; + using Duration = std::chrono::steady_clock::duration;
1679-1681: Remove redundant duration_castcurrentTime - mStartTime already yields steady_clock::duration. Minor readability win.
- auto const elapsed = (std::chrono::duration_cast<Duration>(currentTime - mStartTime)); + auto const elapsed = currentTime - mStartTime;tensorrt_llm/executor/postproc_worker.py (1)
72-73: Consider using explicit types instead of Any for new fields.The new fields
request_perf_metricsanddisaggregated_paramsuseAnytype hints, which reduces type safety and makes the code harder to understand.Define more specific types or use
Optional[Dict]if the structure is dictionary-based:class Output(NamedTuple): client_id: int res: Any is_final: bool error: str = "" metrics: Optional[dict[str, float]] = None - request_perf_metrics: Any = None - disaggregated_params: Any = None + request_perf_metrics: Optional[Dict[str, Any]] = None + disaggregated_params: Optional[Dict[str, Any]] = Nonecpp/tensorrt_llm/nanobind/batch_manager/bindings.cpp (1)
294-336: Validateglobal_steady_clock_offsetwithin ±60 s
In cpp/tensorrt_llm/nanobind/batch_manager/bindings.cpp (after line 295), add:if (global_steady_clock_offset) { auto ms = global_steady_clock_offset->count(); if (ms > 60000 || ms < -60000) { // Log warning for large clock offset } }tensorrt_llm/_torch/pyexecutor/py_executor.py (1)
368-381: Ensure error handling for distributed sync in_get_global_steady_clock_offset
Wrapbarrier()andallgather()in atry/exceptto catch and log communication failures and provide a safe fallback, for example:def _get_global_steady_clock_offset(self): assert self.global_rank >= 0, "rank should be >= 0" - # Sync all ranks - self.dist.barrier() - local_timestamp = time.monotonic() - all_rank_timestamps = self.dist.allgather(local_timestamp) - if self.global_rank == 0: - logger.info( - f"global_steady_clock_offset at each rank: {[local_timestamp - ts for ts in all_rank_timestamps]}" - ) - return all_rank_timestamps[0] - local_timestamp + try: + # Sync all ranks + self.dist.barrier() + local_timestamp = time.monotonic() + all_rank_timestamps = self.dist.allgather(local_timestamp) + if self.global_rank == 0: + logger.info( + f"global_steady_clock_offset at each rank: {[local_timestamp - ts for ts in all_rank_timestamps]}" + ) + return all_rank_timestamps[0] - local_timestamp + except Exception as e: + logger.error(f"Failed to compute global steady clock offset: {e}") + return 0.0
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (14)
cpp/include/tensorrt_llm/batch_manager/llmRequest.h(16 hunks)cpp/tensorrt_llm/nanobind/batch_manager/bindings.cpp(3 hunks)cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.cpp(1 hunks)cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.h(2 hunks)cpp/tensorrt_llm/pybind/batch_manager/bindings.cpp(3 hunks)cpp/tensorrt_llm/pybind/batch_manager/llmRequest.cpp(1 hunks)cpp/tensorrt_llm/pybind/batch_manager/llmRequest.h(1 hunks)tensorrt_llm/_torch/pyexecutor/executor_request_queue.py(3 hunks)tensorrt_llm/_torch/pyexecutor/llm_request.py(1 hunks)tensorrt_llm/_torch/pyexecutor/py_executor.py(4 hunks)tensorrt_llm/executor/postproc_worker.py(4 hunks)tensorrt_llm/executor/result.py(1 hunks)tensorrt_llm/serve/openai_disagg_server.py(12 hunks)tensorrt_llm/serve/openai_server.py(12 hunks)
🧰 Additional context used
📓 Path-based instructions (8)
**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh}: Namespace closing braces must include a trailing comment with the namespace name (e.g., '} // namespace foo').
Prefer const or constexpr variables over #define for constants.
Declare variables that are not modified after initialization as const.
Avoid magic literals in code; except for 0, nullptr, true, false. Use named constants for comparisons and logic.
Use Allman brace style for formatting.
Place the semicolon of an empty for/while loop on a new line.
Bodies of switch/while/do-while/for must be compound statements (brace-delimited), and if/else must always be followed by brace-delimited statements.
Type names (e.g., classes) must be CamelCase starting with an uppercase letter (e.g., FooBar).
Local variables, methods, and namespaces use lowerCamelCase (e.g., localFooBar).
Non-magic-number global variables that are non-static and not in an anonymous namespace must be lowerCamelCase prefixed with 'g' (e.g., gDontUseGlobalFoos).
Non-magic-number globals that are static or in an anonymous namespace use lowerCamelCase prefixed with 's' (e.g., sMutableStaticGlobal).
Locally visible static variables use lowerCamelCase with 's' prefix (e.g., static std::once_flag sFlag).
Private/protected member variables use 'm' prefix with CamelCase (e.g., mNbFooValues). Public members may omit, but 'm' is encouraged for clarity.
Constants (enums, global constants, static constants, and function-scope magic/literal constants) use uppercase SNAKE_CASE with 'k' prefix (e.g., kDIGIT_NUM).
Function-scope constants that are not magic numbers or literals are named like non-constant variables (e.g., bool const pass = a && b).
If macros are necessary, name them in UPPER_SNAKE_CASE (e.g., FOO_VERSION) and prefer constants over #define.
Use LLVM clang-format; wrap lines at a maximum of 120 columns; use '// clang-format off/on' sparingly with justification.
Use smart pointers for heap allocations; prefer unique_ptr for sole ownership, shared_ptr for shared...
Files:
cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.hcpp/tensorrt_llm/nanobind/batch_manager/bindings.cppcpp/tensorrt_llm/pybind/batch_manager/bindings.cppcpp/tensorrt_llm/pybind/batch_manager/llmRequest.cppcpp/tensorrt_llm/nanobind/batch_manager/llmRequest.cppcpp/include/tensorrt_llm/batch_manager/llmRequest.hcpp/tensorrt_llm/pybind/batch_manager/llmRequest.h
**/*.{cpp,cxx,cc,cu,h,hpp,hh,hxx,cuh}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
C++ filenames should be lowerCamelCase (first letter lowercase) and must be case-insensitive unique within a compilation target.
Files:
cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.hcpp/tensorrt_llm/nanobind/batch_manager/bindings.cppcpp/tensorrt_llm/pybind/batch_manager/bindings.cppcpp/tensorrt_llm/pybind/batch_manager/llmRequest.cppcpp/tensorrt_llm/nanobind/batch_manager/llmRequest.cppcpp/include/tensorrt_llm/batch_manager/llmRequest.hcpp/tensorrt_llm/pybind/batch_manager/llmRequest.h
**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Use only spaces, no tabs; indent with 4 spaces.
Files:
cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.htensorrt_llm/_torch/pyexecutor/llm_request.pycpp/tensorrt_llm/nanobind/batch_manager/bindings.cppcpp/tensorrt_llm/pybind/batch_manager/bindings.cpptensorrt_llm/executor/result.pytensorrt_llm/executor/postproc_worker.pytensorrt_llm/_torch/pyexecutor/executor_request_queue.pycpp/tensorrt_llm/pybind/batch_manager/llmRequest.cppcpp/tensorrt_llm/nanobind/batch_manager/llmRequest.cpptensorrt_llm/serve/openai_disagg_server.pycpp/include/tensorrt_llm/batch_manager/llmRequest.htensorrt_llm/serve/openai_server.pytensorrt_llm/_torch/pyexecutor/py_executor.pycpp/tensorrt_llm/pybind/batch_manager/llmRequest.h
**/*.{h,hpp,hh,hxx}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Document new class interfaces and function prototypes with Doxygen; use //! for single-line and //!< for members.
Files:
cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.hcpp/include/tensorrt_llm/batch_manager/llmRequest.hcpp/tensorrt_llm/pybind/batch_manager/llmRequest.h
**/*.{h,hpp,hh,hxx,cpp,cxx,cc}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.{h,hpp,hh,hxx,cpp,cxx,cc}: Prefer anonymous namespaces over 'static' for internal linkage of functions.
All templates (class/function/member/static) must be instantiated at least once; non-POD classes should have private data members.
Files:
cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.hcpp/tensorrt_llm/nanobind/batch_manager/bindings.cppcpp/tensorrt_llm/pybind/batch_manager/bindings.cppcpp/tensorrt_llm/pybind/batch_manager/llmRequest.cppcpp/tensorrt_llm/nanobind/batch_manager/llmRequest.cppcpp/include/tensorrt_llm/batch_manager/llmRequest.hcpp/tensorrt_llm/pybind/batch_manager/llmRequest.h
**/*.{h,hpp,hh,hxx,cuh}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Use include guards named 'TRTLLM_<FILE_NAME_IN_CAPS_WITH_UNDERSCORES>_H' (no leading or trailing underscore; directory names excluded).
Files:
cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.hcpp/include/tensorrt_llm/batch_manager/llmRequest.hcpp/tensorrt_llm/pybind/batch_manager/llmRequest.h
**/*.{cpp,cxx,cc,h,hpp,hh,hxx,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend the NVIDIA Apache-2.0 copyright header with current year to the top of all source files (e.g., .cpp, .h, .cu, .py).
Files:
cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.htensorrt_llm/_torch/pyexecutor/llm_request.pycpp/tensorrt_llm/nanobind/batch_manager/bindings.cppcpp/tensorrt_llm/pybind/batch_manager/bindings.cpptensorrt_llm/executor/result.pytensorrt_llm/executor/postproc_worker.pytensorrt_llm/_torch/pyexecutor/executor_request_queue.pycpp/tensorrt_llm/pybind/batch_manager/llmRequest.cppcpp/tensorrt_llm/nanobind/batch_manager/llmRequest.cpptensorrt_llm/serve/openai_disagg_server.pycpp/include/tensorrt_llm/batch_manager/llmRequest.htensorrt_llm/serve/openai_server.pytensorrt_llm/_torch/pyexecutor/py_executor.pycpp/tensorrt_llm/pybind/batch_manager/llmRequest.h
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+.
Indent Python code with 4 spaces; do not use tabs.
Maintain module namespace when importing; prefer 'from package.subpackage import foo' then 'foo.SomeClass()' instead of importing the class directly.
Python filenames should be snake_case (e.g., some_file.py).
Python classes use PascalCase names.
Functions and methods use snake_case names.
Local variables use snake_case; prefix 'k' for variables that start with a number (e.g., k_99th_percentile).
Global variables use upper SNAKE_CASE prefixed with 'G' (e.g., G_MY_GLOBAL).
Constants use upper SNAKE_CASE (e.g., MY_CONSTANT).
Avoid shadowing variables from an outer scope.
Initialize all externally visible members of a class in the constructor.
Prefer docstrings for interfaces that may be used outside a file; comments for in-function or file-local interfaces.
Use Google-style docstrings for classes and functions (Sphinx-parsable).
Document attributes and variables inline so they render under the class/function docstring.
Avoid reflection when a simpler, explicit approach suffices (e.g., avoid dict(**locals()) patterns).
In try/except, catch the most specific exceptions possible.
For duck-typing try/except, keep the try body minimal and use else for the main logic.
Files:
tensorrt_llm/_torch/pyexecutor/llm_request.pytensorrt_llm/executor/result.pytensorrt_llm/executor/postproc_worker.pytensorrt_llm/_torch/pyexecutor/executor_request_queue.pytensorrt_llm/serve/openai_disagg_server.pytensorrt_llm/serve/openai_server.pytensorrt_llm/_torch/pyexecutor/py_executor.py
🧠 Learnings (1)
📚 Learning: 2025-08-20T06:56:02.889Z
Learnt from: eopXD
PR: NVIDIA/TensorRT-LLM#6768
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:577-579
Timestamp: 2025-08-20T06:56:02.889Z
Learning: In cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp, maxSequenceLength is now enforced as a non-optional argument in the BlockManager constructor, so concerns about std::nullopt defaulting to 0 are not applicable. When windowSize > maxSequenceLength, a warning should be added instead of handling optional parameter cases.
Applied to files:
cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.hcpp/tensorrt_llm/pybind/batch_manager/llmRequest.cpp
🧬 Code graph analysis (4)
tensorrt_llm/executor/postproc_worker.py (1)
tensorrt_llm/serve/openai_disagg_server.py (1)
perf_metrics(204-255)
tensorrt_llm/serve/openai_disagg_server.py (2)
tensorrt_llm/serve/openai_server.py (3)
add_process_time_header(167-171)openai_completion(603-756)set_steady_clock_offset(323-326)tensorrt_llm/serve/router.py (4)
finish_request(172-173)finish_request(423-424)finish_request(496-502)finish_request(601-608)
tensorrt_llm/serve/openai_server.py (3)
tensorrt_llm/serve/openai_disagg_server.py (2)
add_process_time_header(141-145)set_steady_clock_offset(529-533)triton_backend/all_models/inflight_batcher_llm/tensorrt_llm_bls/1/lib/decode.py (1)
Response(200-245)tensorrt_llm/llmapi/llm.py (1)
RequestOutput(49-89)
tensorrt_llm/_torch/pyexecutor/py_executor.py (1)
tensorrt_llm/_torch/distributed/communicator.py (3)
barrier(339-340)allgather(101-102)allgather(336-337)
🪛 Ruff (0.13.1)
tensorrt_llm/serve/openai_disagg_server.py
269-269: Do not catch blind exception: Exception
(BLE001)
327-327: Avoid specifying long messages outside the exception class
(TRY003)
340-340: Avoid specifying long messages outside the exception class
(TRY003)
342-342: Undefined name anext. Consider specifying requires-python = ">= 3.10" or tool.ruff.target-version = "py310" in your pyproject.toml file.
(F821)
349-350: Store a reference to the return value of asyncio.create_task
(RUF006)
527-527: Do not catch blind exception: Exception
(BLE001)
tensorrt_llm/serve/openai_server.py
681-681: Undefined name anext. Consider specifying requires-python = ">= 3.10" or tool.ruff.target-version = "py310" in your pyproject.toml file.
(F821)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (5)
cpp/include/tensorrt_llm/batch_manager/llmRequest.h (1)
2192-2211: Nice: centralized clock-domain normalizationmaybeToGlobalSteadyClock and getSteadyClockNow encapsulate the offset logic well. This is the right approach to avoid scattered conversions.
tensorrt_llm/executor/result.py (1)
340-350: LGTM: propagate request_perf_metrics and disaggregated_paramsCorrectly attaches request_perf_metrics and ensures disaggregated_params is populated for the first output, aligning with postproc worker changes.
tensorrt_llm/_torch/pyexecutor/llm_request.py (1)
589-591: Confirm global_steady_clock_offset units in Python binding
The C++ constructor takes a std::chrono::steady_clock::duration (nanosecond ticks by default); ensure py_global_steady_clock_offset is supplied as an integer count of nanoseconds (or convert seconds→nanoseconds) to avoid silent timing skew.cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.h (1)
89-90: LGTM! Appropriate extension for clock synchronization.The addition of the
globalSteadyClockOffsetparameter maintains backward compatibility with the defaultstd::nulloptvalue, properly forwards to the base class, and follows the existing pattern of optional parameters.cpp/tensorrt_llm/nanobind/batch_manager/bindings.cpp (1)
294-295: Consistent parameter naming and documentation.The new parameter
global_steady_clock_offsetis well-integrated with appropriate default value. The parameter placement at the end maintains backward compatibility.Also applies to: 336-336, 362-362
nv-yilinf
requested review from
kaiyux,
pcastonguay,
reasonsolo and
zhengd-nv
|
/bot run --disable-fail-fast |
nv-yilinf
force-pushed
the
ttft-tracer
branch
from
188924b to
dce681e
Compare
|
PR_Github #20130 [ run ] triggered by Bot |
nv-yilinf
force-pushed
the
ttft-tracer
branch
from
dce681e to
e6d7027
Compare
|
/bot run --disable-fail-fast |
|
PR_Github #20131 [ run ] triggered by Bot |
|
PR_Github #20130 [ run ] completed with state |
|
PR_Github #20131 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #20441 [ run ] triggered by Bot |
|
PR_Github #20441 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #20470 [ run ] triggered by Bot |
|
PR_Github #20470 [ run ] completed with state |
Signed-off-by: nv-yilinf <[email protected]>
|
/bot run --disable-fail-fast |
|
PR_Github #20484 [ run ] triggered by Bot |
|
/bot run --disable-fail-fast |
|
PR_Github #20493 [ run ] triggered by Bot |
|
PR_Github #20484 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #20502 [ run ] triggered by Bot |
|
PR_Github #20493 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #20509 [ run ] triggered by Bot |
|
PR_Github #20502 [ run ] completed with state |
|
PR_Github #20509 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #20539 [ run ] triggered by Bot |
|
PR_Github #20539 [ run ] completed with state |
…8005) Signed-off-by: Yilin Fan <[email protected]> Signed-off-by: nv-yilinf <[email protected]>
…8005) Signed-off-by: Yilin Fan <[email protected]> Signed-off-by: nv-yilinf <[email protected]> Signed-off-by: Faradawn Yang <[email protected]>
6 participants
@coderabbitai summary
Description
Improve /perf_metrics endpoints in the following aspects:
server_start_time/server_first_token_timeanddisagg_server_start_time/disagg_server_first_token_timeto help us better understand server overhead.example output of /perf_metrics :
Test Coverage
PR Checklist
Please review the following before submitting your PR:
PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.
GitHub Bot Help
/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...Provide a user friendly way for developers to interact with a Jenkins server.
Run
/bot [-h|--help]to print this help message.See details below for each supported subcommand.
run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]Launch build/test pipelines. All previously running jobs will be killed.
--reuse-test (optional)pipeline-id(OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.--disable-reuse-test(OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.--disable-fail-fast(OPTIONAL) : Disable fail fast on build/tests/infra failures.--skip-test(OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.--stage-list "A10-PyTorch-1, xxx"(OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.--gpu-type "A30, H100_PCIe"(OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.--test-backend "pytorch, cpp"(OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.--only-multi-gpu-test(OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.--disable-multi-gpu-test(OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.--add-multi-gpu-test(OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.--post-merge(OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx"(OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".--detailed-log(OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.--debug(OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in thestage-listparameter to access the appropriate container environment. Note: Does NOT update GitHub check status.For guidance on mapping tests to stage names, see
docs/source/reference/ci-overview.mdand the
scripts/test_to_stage_mapping.pyhelper.kill
killKill all running builds associated with pull request.
skip
skip --comment COMMENTSkip testing for latest commit on pull request.
--comment "Reason for skipping build/test"is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.reuse-pipeline
reuse-pipelineReuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.