[https://nvbugs/5434320][bug] Fix disagg pp bug #7099
Conversation
📝 WalkthroughWalkthroughCentralizes per-request response progression in Changes
Sequence Diagram(s)sequenceDiagram
autonumber
participant Caller
participant DataTransceiver
participant Sender
participant LlmRequest
Caller->>DataTransceiver: response()
activate DataTransceiver
alt ready response present
DataTransceiver->>DataTransceiver: prepare reqId/counter if new
DataTransceiver->>DataTransceiver: call sendResponse(blockHashes,it)
else wait for ready response
loop wait/recheck
DataTransceiver-->>DataTransceiver: wait(mResponderCv) / recheck getCurrentResponse()
end
DataTransceiver->>DataTransceiver: call sendResponse(blockHashes,it)
end
rect rgba(200,230,255,0.18)
note right of DataTransceiver: sendResponse flow
DataTransceiver->>DataTransceiver: reqId = mCurrentRequest.value()\n--mRemainSendCount[reqId] (assert >=0)
DataTransceiver->>LlmRequest: setRequestedBlockHashes(move(blockHashes))
alt EnvParallelCacheSend == true
DataTransceiver->>Sender: spawn thread -> sendAndRemoveResponse(...)
else
DataTransceiver->>Sender: sendAndRemoveResponse(...) (inline)
end
DataTransceiver->>DataTransceiver: removeResponse(it)\nclear mCurrentRequest
end
deactivate DataTransceiver
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Possibly related PRs
Suggested reviewers
📜 Recent review detailsConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro 💡 Knowledge Base configuration:
You can enable these sources in your CodeRabbit configuration. 📒 Files selected for processing (2)
💤 Files with no reviewable changes (1)
🚧 Files skipped from review as they are similar to previous changes (1)
⏰ 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)
✨ Finishing Touches
🧪 Generate unit tests
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. 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. CodeRabbit Commands (Invoked using PR/Issue comments)Type Other keywords and placeholders
Status, Documentation and Community
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
tests/integration/defs/accuracy/test_disaggregated_serving.py (1)
144-148: Instance-specific NSYS output per ctx server.Create per-instance NSYS args so each server writes a distinct report.
Apply this diff:
- ctx_server_args = nsys_args + common_args + [ + nsys_args = nsys_args_base + [f"--output=ctx_server_trace_{port}"] + ctx_server_args = nsys_args + common_args + [ "--port", str(port), "--extra_llm_api_options", ctx_server_config_path, f"--tp_size={ctx_tp}", f"--pp_size={ctx_pp}" ]Optional: gate profiling to environments where NSYS is available (e.g., only when os.environ.get("TLLM_ENABLE_NSYS_CTX") == "1", or if shutil.which("nsys") is not None) to avoid CI failures.
🧹 Nitpick comments (4)
tensorrt_llm/_torch/pyexecutor/py_executor.py (2)
799-819: Block on isend only when you’re about to finalize the same microbatch (avoid unnecessary stall on last PP rank).The added wait() at Line 818 ensures memory lifetime before resource reclamation on non-last PP ranks (good to prevent UAF). However, on the last PP rank Stage 2 and Stage 3 operate on different prev_microbatch_ids, so this wait is not required and may reduce overlap. Restrict the post-isend wait to non-last PP ranks.
Apply this diff:
- self.send_handles[prev_microbatch_id].wait() + # Non-last PP ranks finalize the same microbatch in Stage 3, so ensure send completes. + if not self.dist.is_last_pp_rank: + self.send_handles[prev_microbatch_id].wait()Also, minor: needs_logits currently uses scheduled_batch, but the data being sent belongs to previous_batch. Consider basing the decision on previous_batch.sample_state.scheduled_requests to avoid mismatch if current/previous batches diverge. Not critical to this PR, but worth a follow-up.
1-10: Missing NVIDIA SPDX header for Python source.Per repo guidelines, prepend the NVIDIA copyright header to Python files.
I can provide a ready-to-paste header if you want.
tests/integration/defs/accuracy/test_disaggregated_serving.py (1)
1-18: Missing NVIDIA SPDX header in test file.Tests are also source files. Please prepend the NVIDIA header to comply with the repo standard.
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (1)
223-236: Avoid unnecessary copy and dangling lifetime risk; remove unused blockHashes copy and bind by value.The temporary RequestInfo from recvRequestInfo() is currently bound to a const reference, and you copy blockHashes into a local that isn’t used afterwards. Clean this up and move the RequestInfo into the map.
Apply this diff:
- std::vector<size_t> blockHashes; - auto const& requestInfo = mSender->recvRequestInfo(); - auto reqId = requestInfo.getRequestId(); - blockHashes = requestInfo.getBlockHashes(); + auto requestInfo = mSender->recvRequestInfo(); + auto reqId = requestInfo.getRequestId(); { std::unique_lock lk(mSendMutex); - mRequestInfoMap[reqId] = std::move(requestInfo); + mRequestInfoMap[reqId] = std::move(requestInfo); if (mRemainSendCount.find(reqId) == mRemainSendCount.end()) { mRemainSendCount[reqId] = mSender->getCounterpartsCount(reqId); } } sendResponse(reqId);
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (3)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp(4 hunks)tensorrt_llm/_torch/pyexecutor/py_executor.py(1 hunks)tests/integration/defs/accuracy/test_disaggregated_serving.py(2 hunks)
🧰 Additional context used
📓 Path-based instructions (5)
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+
Python indentation: 4 spaces, no tabs
Maintain module namespace in imports (from package.subpackage import foo; then use foo.SomeClass())
Python file names use snake_case
Python class names use PascalCase
Python functions/methods and local variables use snake_case; variables starting with a number get k_ prefix (e.g., k_99th_percentile)
Global variables use G_ prefixed UPPER_SNAKE_CASE (e.g., G_MY_GLOBAL)
Constants use UPPER_SNAKE_CASE in Python
Avoid shadowing variables from outer scopes in Python
Initialize all externally visible members of a Python class in init
Prefer docstrings for interfaces used outside a file; comments for local code
Use Google-style docstrings for classes and functions (Sphinx-parsable)
Document attributes/variables inline with short docstrings
Avoid reflection when simple alternatives exist (e.g., prefer explicit parameters over dict(**locals()))
In try/except, catch the narrowest exceptions possible
For duck-typing with try/except, keep try body minimal and put logic in else
Files:
tensorrt_llm/_torch/pyexecutor/py_executor.pytests/integration/defs/accuracy/test_disaggregated_serving.py
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Prepend NVIDIA copyright header (current year) to all source files
Files:
tensorrt_llm/_torch/pyexecutor/py_executor.pytests/integration/defs/accuracy/test_disaggregated_serving.pycpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh}: In C++, close namespaces with a comment naming the namespace (e.g., } // namespace foo)
Prefer const/constexpr variables over #define for constants
Declare variables const if not modified after initialization
Use Allman brace style in C++
C++ filenames use lowerCamelCase and must be case-insensitively unique within a build target
C++ type names use UpperCamelCase
Local variables, methods, and namespaces use lowerCamelCase
Global non-static variables not in anonymous namespace use gPrefix lowerCamelCase (e.g., gExample)
Static globals or globals in anonymous namespaces use sPrefix lowerCamelCase
Locally visible static variables start with 's' (e.g., static std::once_flag sFlag;)
Member variables use mPrefix lowerCamelCase; public members may omit but are encouraged to use 'm'
Constants (enums, global/static/function-scope magic numbers) use kPREFIXED_UPPER_SNAKE (e.g., kDIGIT_NUM)
If macros are unavoidable, use UPPER_SNAKE_CASE (prefer constants over #define)
Constructor parameter that conflicts with a public member name gets trailing underscore (foo_)
Literal suffixes should be uppercase (e.g., 1234L not 1234l)
C++: use spaces only; indent 4 spaces
Run clang-format (LLVM style) before submitting; wrap lines at 120 characters
If formatting must be bypassed, use // clang-format off/on around the section
Prefer smart pointers; use unique_ptr for sole ownership, shared_ptr for shared; weak_ptr only in exceptional cases
Do not use deprecated pre-C++11 smart pointers
Use C++ style comments; avoid C comments except special inline cases; prefer // single-line
Capitalize and punctuate full-sentence comments
Follow Doxygen rules: use //! for comments and //!< for members in C++
Disable code with #if/#endif and mnemonic conditions; avoid commented-out code; avoid dead code
Do not throw exceptions across library boundaries
Use least-forceful casts; avoid removing const/volatile; avoid C-style and functional casts (except constructors); p...
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{cpp,cxx,cc,cu}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.{cpp,cxx,cc,cu}: Avoid literal values except for 0, nullptr, true, false; use named constexpr for other literals
Place semicolon of empty for/while loop on a new line
Always use brace-delimited bodies for switch/while/do-for/if/else
Use inline C comments in argument lists when parameter meaning is unclear (e.g., /* checkForErrors = */ false)
Do not use assignment in subexpressions (e.g., if (x = y) ... is forbidden)
Switch on enums should enumerate all values and omit default to catch new values at compile time
Structure switch statements; prohibit fallthrough except between empty cases; each case ends with break or throw; return at end of case not allowed; put break inside braces for compound case
Prefer anonymous namespaces over static for internal linkage of functions
Every defined function must be called at least once (no unused methods)
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{h,hpp,hxx,hh,cuh,cpp,cxx,cc,cu}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Parameter names must be consistent between declarations and definitions
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
🧠 Learnings (1)
📚 Learning: 2025-08-06T08:18:28.669Z
Learnt from: zhengd-nv
PR: NVIDIA/TensorRT-LLM#6633
File: cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp:145-155
Timestamp: 2025-08-06T08:18:28.669Z
Learning: In cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp, the existing `mMtxForMap` mutex in DataSenderImpl is sufficient to synchronize measurement file operations in the `release` method, as all file operations occur within the same critical section that protects the `mRequestToSession` map access.
Applied to files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
🔇 Additional comments (1)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (1)
274-281: Lock ordering looks consistent, but please double-check for future changes.sendResponse() acquires mSendMutex then (temporarily) mResponderMutex and later removeResponse() acquires mResponderMutex (and mCondMutex). Elsewhere, respondAndSendAsync() acquires mResponderMutex first, but releases it before calling sendResponse(). This avoids an immediate inversion. Keep this ordering in mind for future edits.
If you want, I can scan the repository to ensure no other code paths acquire mResponderMutex then mSendMutex in that order, which could deadlock.
Tabrizian
force-pushed
the
user/imant/fixDisaggBug
branch
from
718b583 to
9b63bd7
Compare
There was a problem hiding this comment.
Actionable comments posted: 1
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (3)
tests/integration/defs/accuracy/test_disaggregated_serving.py (3)
53-53: Replace PEP 585 built-in generics with typing.List for Python 3.8+.
list[... ]requires Python 3.9+. UseList[...]which you already import.Apply this diff:
- self.futures: list[concurrent.futures.Future[RequestOutput]] = [] + self.futures: List[concurrent.futures.Future[RequestOutput]] = []
61-64: Avoidcancel_futuresin ThreadPoolExecutor.shutdown for Python 3.8.
cancel_futureswas added in Python 3.9. The manualfuture.cancel()loop above already handles pending futures.Apply this diff:
- self.shutdown(wait=True, cancel_futures=True) + self.shutdown(wait=True)
193-201: Add timeouts and a bounded wait for the health probe to avoid indefinite hangs.
requests.get()without a timeout can hang forever; the loop also lacks a global deadline even though the server is launched with--server_start_timeout 3600.Apply this diff:
]): - while True: + start = time.monotonic() + timeout_s = 3600 + while True: time.sleep(1) try: print("Checking health endpoint") - response = requests.get("http://localhost:8000/health") + response = requests.get("http://localhost:8000/health", timeout=5) if response.status_code == 200: break - except requests.exceptions.ConnectionError: + except requests.exceptions.RequestException: continue + if time.monotonic() - start > timeout_s: + raise RuntimeError("disaggregated server did not become healthy within 3600s")
🧹 Nitpick comments (3)
tests/integration/defs/accuracy/test_disaggregated_serving.py (3)
186-186: Makemax_workersconfigurable for CI tuning.Allow overriding via env var to ease A/B testing and CI tuning without touching code.
Apply this diff:
- with MyThreadPoolExecutor(max_workers=4) as thread_pool, temp_dir: + workers = int(os.getenv("TRTLLM_TEST_MAX_WORKERS", "4")) + with MyThreadPoolExecutor(max_workers=workers) as thread_pool, temp_dir:
1-1: Missing NVIDIA copyright header.Per project guidelines, prepend the current-year NVIDIA header.
Apply this diff at the top of the file:
+# Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
658-659: Remove stray no-op lines.These two lines are orphan expressions and do nothing; likely leftovers.
Apply this diff:
- ctx_server_config["cache_transceiver_config"] - ctx_server_config["cache_transceiver_config"]
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (4)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp(4 hunks)cpp/tensorrt_llm/batch_manager/llmRequest.cpp(0 hunks)tensorrt_llm/_torch/pyexecutor/py_executor.py(1 hunks)tests/integration/defs/accuracy/test_disaggregated_serving.py(1 hunks)
💤 Files with no reviewable changes (1)
- cpp/tensorrt_llm/batch_manager/llmRequest.cpp
🚧 Files skipped from review as they are similar to previous changes (2)
- tensorrt_llm/_torch/pyexecutor/py_executor.py
- cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+
Python indentation: 4 spaces, no tabs
Maintain module namespace in imports (from package.subpackage import foo; then use foo.SomeClass())
Python file names use snake_case
Python class names use PascalCase
Python functions/methods and local variables use snake_case; variables starting with a number get k_ prefix (e.g., k_99th_percentile)
Global variables use G_ prefixed UPPER_SNAKE_CASE (e.g., G_MY_GLOBAL)
Constants use UPPER_SNAKE_CASE in Python
Avoid shadowing variables from outer scopes in Python
Initialize all externally visible members of a Python class in init
Prefer docstrings for interfaces used outside a file; comments for local code
Use Google-style docstrings for classes and functions (Sphinx-parsable)
Document attributes/variables inline with short docstrings
Avoid reflection when simple alternatives exist (e.g., prefer explicit parameters over dict(**locals()))
In try/except, catch the narrowest exceptions possible
For duck-typing with try/except, keep try body minimal and put logic in else
Files:
tests/integration/defs/accuracy/test_disaggregated_serving.py
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Prepend NVIDIA copyright header (current year) to all source files
Files:
tests/integration/defs/accuracy/test_disaggregated_serving.py
🧠 Learnings (1)
📓 Common learnings
Learnt from: zhengd-nv
PR: NVIDIA/TensorRT-LLM#6633
File: cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp:145-155
Timestamp: 2025-08-06T08:18:28.669Z
Learning: In cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp, the existing `mMtxForMap` mutex in DataSenderImpl is sufficient to synchronize measurement file operations in the `release` method, as all file operations occur within the same critical section that protects the `mRequestToSession` map access.
⏰ 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 (2)
tests/integration/defs/accuracy/test_disaggregated_serving.py (2)
186-186: Lowering thread-pool concurrency to 4 looks reasonable for stabilizing disagg PP send/recv.This should reduce contention and flakiness when non-last PP ranks block on isend handles. No API changes, so low blast radius.
231-239: Streaming path isn’t handled; confirm it’s unused in this harness.
client.completions.create(..., stream=True)yields an event stream, not a single response. Sincesend_requestassumes a non-streaming object, ensurestreamingis always False in this test, or add streaming handling.Would you like me to add a guarded streaming implementation for completeness?
Tabrizian
force-pushed
the
user/imant/fixDisaggBug
branch
2 times, most recently
from
33978ba to
0653c4b
Compare
Tabrizian
changed the title
|
/bot run --add-multi-gpu-test --disable-fail-fast |
There was a problem hiding this comment.
Actionable comments posted: 0
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (1)
178-182: Incomplete cleanup in destructorThe destructor calls
terminate()but doesn't wait formResponseFutureto complete, which could lead to the response thread accessing destroyed members.~Impl() { terminate(); + if (mResponseFuture.valid()) + { + mResponseFuture.wait(); + } }
♻️ Duplicate comments (2)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (2)
107-149: Performance: Consider using move semantics for block hashesThe code copies the block hash vector on line 123 and then moves it on line 133. Since the
RequestInfois being erased immediately after, you can avoid the copy by adding a move-out accessor.Add a
takeBlockHashes()method toRequestInfo(in the header file):std::vector<size_t> takeBlockHashes() noexcept { return std::move(mBlockHashes); }Then update line 123:
- auto blockHashes = it->second.getBlockHashes(); + auto blockHashes = it->second.takeBlockHashes();
135-141: Thread safety concern: Detached thread capturesthisThe detached thread on lines 138-140 captures
this, which can lead to undefined behavior if theDataResponder::Implinstance is destroyed while the thread is still running.Consider one of these solutions:
- Track the threads and join them in the destructor
- Use a thread pool with proper lifecycle management
- Use
std::asyncwhich returns a future that can be trackedExample fix using futures:
+private: + std::vector<std::future<void>> mSendFutures; + public: + ~Impl() + { + terminate(); + // Wait for all pending sends to complete + for (auto& future : mSendFutures) + { + if (future.valid()) + { + future.wait(); + } + } + }And update the detached thread creation:
- std::thread( - &DataResponder::Impl::sendAndRemoveResponse, this, reqId, std::move(readyResponseIt->second)) - .detach(); + mSendFutures.push_back( + std::async(std::launch::async, + &DataResponder::Impl::sendAndRemoveResponse, this, reqId, std::move(readyResponseIt->second)));
🧹 Nitpick comments (2)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (2)
274-274: Naming convention: Inconsistent mutex orderingThe mutex declaration order
mCondMutex, mSendMutex, mResponderMutexdoesn't follow a clear pattern. Consider ordering them alphabetically or by usage frequency for better readability.- std::mutex mCondMutex, mSendMutex, mResponderMutex; + std::mutex mCondMutex, mResponderMutex, mSendMutex;
124-126: Add defensive check for negative countWhile the code checks that
count >= 0after decrement, it would be more defensive to also ensuremRemainSendCount[reqId]is positive before decrementing.+ TLLM_CHECK(mRemainSendCount[reqId] > 0); auto count = --mRemainSendCount[reqId]; TLLM_CHECK(count >= 0);
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (5)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp(4 hunks)cpp/tensorrt_llm/batch_manager/llmRequest.cpp(0 hunks)tensorrt_llm/_torch/pyexecutor/py_executor.py(1 hunks)tests/integration/defs/accuracy/test_disaggregated_serving.py(1 hunks)tests/integration/test_lists/waives.txt(0 hunks)
💤 Files with no reviewable changes (2)
- cpp/tensorrt_llm/batch_manager/llmRequest.cpp
- tests/integration/test_lists/waives.txt
🚧 Files skipped from review as they are similar to previous changes (2)
- tensorrt_llm/_torch/pyexecutor/py_executor.py
- tests/integration/defs/accuracy/test_disaggregated_serving.py
🧰 Additional context used
📓 Path-based instructions (4)
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh}: In C++, close namespaces with a comment naming the namespace (e.g., } // namespace foo)
Prefer const/constexpr variables over #define for constants
Declare variables const if not modified after initialization
Use Allman brace style in C++
C++ filenames use lowerCamelCase and must be case-insensitively unique within a build target
C++ type names use UpperCamelCase
Local variables, methods, and namespaces use lowerCamelCase
Global non-static variables not in anonymous namespace use gPrefix lowerCamelCase (e.g., gExample)
Static globals or globals in anonymous namespaces use sPrefix lowerCamelCase
Locally visible static variables start with 's' (e.g., static std::once_flag sFlag;)
Member variables use mPrefix lowerCamelCase; public members may omit but are encouraged to use 'm'
Constants (enums, global/static/function-scope magic numbers) use kPREFIXED_UPPER_SNAKE (e.g., kDIGIT_NUM)
If macros are unavoidable, use UPPER_SNAKE_CASE (prefer constants over #define)
Constructor parameter that conflicts with a public member name gets trailing underscore (foo_)
Literal suffixes should be uppercase (e.g., 1234L not 1234l)
C++: use spaces only; indent 4 spaces
Run clang-format (LLVM style) before submitting; wrap lines at 120 characters
If formatting must be bypassed, use // clang-format off/on around the section
Prefer smart pointers; use unique_ptr for sole ownership, shared_ptr for shared; weak_ptr only in exceptional cases
Do not use deprecated pre-C++11 smart pointers
Use C++ style comments; avoid C comments except special inline cases; prefer // single-line
Capitalize and punctuate full-sentence comments
Follow Doxygen rules: use //! for comments and //!< for members in C++
Disable code with #if/#endif and mnemonic conditions; avoid commented-out code; avoid dead code
Do not throw exceptions across library boundaries
Use least-forceful casts; avoid removing const/volatile; avoid C-style and functional casts (except constructors); p...
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{cpp,cxx,cc,cu}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.{cpp,cxx,cc,cu}: Avoid literal values except for 0, nullptr, true, false; use named constexpr for other literals
Place semicolon of empty for/while loop on a new line
Always use brace-delimited bodies for switch/while/do-for/if/else
Use inline C comments in argument lists when parameter meaning is unclear (e.g., /* checkForErrors = */ false)
Do not use assignment in subexpressions (e.g., if (x = y) ... is forbidden)
Switch on enums should enumerate all values and omit default to catch new values at compile time
Structure switch statements; prohibit fallthrough except between empty cases; each case ends with break or throw; return at end of case not allowed; put break inside braces for compound case
Prefer anonymous namespaces over static for internal linkage of functions
Every defined function must be called at least once (no unused methods)
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{h,hpp,hxx,hh,cuh,cpp,cxx,cc,cu}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Parameter names must be consistent between declarations and definitions
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Prepend NVIDIA copyright header (current year) to all source files
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
🧠 Learnings (2)
📓 Common learnings
Learnt from: zhengd-nv
PR: NVIDIA/TensorRT-LLM#6633
File: cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp:145-155
Timestamp: 2025-08-06T08:18:28.669Z
Learning: In cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp, the existing `mMtxForMap` mutex in DataSenderImpl is sufficient to synchronize measurement file operations in the `release` method, as all file operations occur within the same critical section that protects the `mRequestToSession` map access.
📚 Learning: 2025-08-06T08:18:28.669Z
Learnt from: zhengd-nv
PR: NVIDIA/TensorRT-LLM#6633
File: cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp:145-155
Timestamp: 2025-08-06T08:18:28.669Z
Learning: In cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp, the existing `mMtxForMap` mutex in DataSenderImpl is sufficient to synchronize measurement file operations in the `release` method, as all file operations occur within the same critical section that protects the `mRequestToSession` map access.
Applied to files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
🔇 Additional comments (2)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (2)
111-117: No deadlock risk between mSendMutex and mResponderMutex
A review of every lock invocation in dataTransceiver.cpp shows that the only time both mSendMutex and mResponderMutex are held together (lines 109–111) acquires them in the same order, and there are no code paths that acquire mResponderMutex first and then attempt mSendMutex. All other uses lock only one mutex at a time.• Lines 109–111:
unique_lock lk(mSendMutex);thenunique_lock lkResp(mResponderMutex);
• Lines 156 & 262: onlyunique_lock lkResp(mResponderMutex);
• Line 228: onlyunique_lock lk(mSendMutex);Because there is no reverse ordering anywhere, there is no inconsistent lock ordering and no deadlock risk here.
224-235: False positive: sendResponse ordering is correctThe call to
sendResponse(reqId)immediately after storing the newrequestInfois intentionally there to flush any previously queued responses for that request ID.
– In the receive loop,sendResponselocksmSendMutexthenmResponderMutex, checksmReadyResponsesforreqId, and simply returns if no response is pending.
– New responses are only ever added tomReadyResponsesinrespondAndSendAsync(lines 157–159), which then immediately callssendResponseto deliver them.Because
respondAndSendAsyncboth enqueues the response and invokessendResponseunder the proper locks, there is no race or ordering issue: the early-return in the receive path is expected and harmless.Likely an incorrect or invalid review comment.
|
PR_Github #15950 [ run ] triggered by Bot |
There was a problem hiding this comment.
Actionable comments posted: 3
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (1)
238-245: Concurrent access tomReadyResponsesin exception pathThe catch block iterates
mReadyResponseswithoutmResponderMutex, which races with other threads adding/removing entries.Guard the map while propagating exceptions:
catch (std::exception const& err) { TLLM_LOG_ERROR("Exception in DataResponder response: %s", err.what()); { std::unique_lock lk(mResponderMutex); for (auto& it : mReadyResponses) { it.second.mPromise.set_exception(std::current_exception()); } } }
♻️ Duplicate comments (1)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (1)
123-127: Build break + unnecessary copy: implement and use a move-out accessor for block hashes
RequestInfo::getBlockHashes()is called here but no such accessor is defined in this TU; additionally, you immediately move the vector intollmRequest->setRequestedBlockHashes(...), so copying first is wasteful. Use a move-taking accessor instead.Apply this minimal fix now (assuming you add
takeBlockHashes()on RequestInfo as shown below):- auto blockHashes = it->second.getBlockHashes(); + auto blockHashes = it->second.takeBlockHashes(); auto count = --mRemainSendCount[reqId];Add the accessors to RequestInfo (header; shown here for clarity):
// In class RequestInfo (public): [[nodiscard]] std::vector<size_t> const& getBlockHashes() const noexcept { return mBlockHashes; } // Move-out accessor to avoid copying when the entry is about to be erased: std::vector<size_t> takeBlockHashes() noexcept { return std::move(mBlockHashes); }
🧹 Nitpick comments (2)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (1)
111-118: Lock ordering and iterator lifetime — reduce risk of future deadlocks
sendResponsecurrently takesmSendMutexthen briefly takesmResponderMutex, releases it, and continues using the iterator. This is likely safe (std::map insertions don’t invalidate iterators; erasures are controlled). However, to avoid subtle lock-ordering hazards if future code paths acquire these locks in different orders, consider atomically acquiring both:- std::unique_lock lk(mSendMutex); - // Send context cache is not called for this request yet. - std::unique_lock<std::mutex> lkResp(mResponderMutex); + std::scoped_lock lk(mSendMutex, mResponderMutex); auto readyResponseIt = mReadyResponses.find(reqId); if (readyResponseIt == mReadyResponses.end()) { return; } - lkResp.unlock(); + // mResponderMutex remains held or release here if you preferIf you keep the current scheme, ensure no other code path locks
mResponderMutexand then tries to grabmSendMutex.tensorrt_llm/_torch/pyexecutor/py_executor.py (1)
818-821: Blocking onisend_objectremoves PP overlap; gate it and clear handle after waitThe immediate
wait()ensures correctness on problematic nodes but also serializes the PP communication path and can hurt throughput.Consider guarding with an env flag and clearing the handle after waiting to avoid a redundant wait later:
- self.send_handles[ - prev_microbatch_id] = self.dist.isend_object( + handle = self.dist.isend_object( ( serialized_logits, sample_state.host, ), dest=self.dist.next_pp_rank, tag=prev_microbatch_id) - # TODO: remove this wait, without this wait - # there is an intermittent hang on some nodes. - self.send_handles[prev_microbatch_id].wait() + # Keep for later completion checks + self.send_handles[prev_microbatch_id] = handle + # TODO: remove this guarded wait after root cause is fixed. + if os.environ.get("TLLM_FORCE_SYNC_PP_SEND", "0") == "1": + handle.wait() + # Prevent double-wait in the next iteration + self.send_handles[prev_microbatch_id] = NoneFollow-up:
- Please confirm this wait resolves the intermittent hang on the affected nodes and share repro details so we can pursue a targeted fix (e.g., NCCL PG settings, tags, or ordering).
- If hang persists, add a watchdog log if wait exceeds a threshold to aid triage.
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (5)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp(4 hunks)cpp/tensorrt_llm/batch_manager/llmRequest.cpp(0 hunks)tensorrt_llm/_torch/pyexecutor/py_executor.py(1 hunks)tests/integration/defs/accuracy/test_disaggregated_serving.py(1 hunks)tests/integration/test_lists/waives.txt(0 hunks)
💤 Files with no reviewable changes (2)
- cpp/tensorrt_llm/batch_manager/llmRequest.cpp
- tests/integration/test_lists/waives.txt
🚧 Files skipped from review as they are similar to previous changes (1)
- tests/integration/defs/accuracy/test_disaggregated_serving.py
🧰 Additional context used
📓 Path-based instructions (5)
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+
Python indentation: 4 spaces, no tabs
Maintain module namespace in imports (from package.subpackage import foo; then use foo.SomeClass())
Python file names use snake_case
Python class names use PascalCase
Python functions/methods and local variables use snake_case; variables starting with a number get k_ prefix (e.g., k_99th_percentile)
Global variables use G_ prefixed UPPER_SNAKE_CASE (e.g., G_MY_GLOBAL)
Constants use UPPER_SNAKE_CASE in Python
Avoid shadowing variables from outer scopes in Python
Initialize all externally visible members of a Python class in init
Prefer docstrings for interfaces used outside a file; comments for local code
Use Google-style docstrings for classes and functions (Sphinx-parsable)
Document attributes/variables inline with short docstrings
Avoid reflection when simple alternatives exist (e.g., prefer explicit parameters over dict(**locals()))
In try/except, catch the narrowest exceptions possible
For duck-typing with try/except, keep try body minimal and put logic in else
Files:
tensorrt_llm/_torch/pyexecutor/py_executor.py
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Prepend NVIDIA copyright header (current year) to all source files
Files:
tensorrt_llm/_torch/pyexecutor/py_executor.pycpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh}: In C++, close namespaces with a comment naming the namespace (e.g., } // namespace foo)
Prefer const/constexpr variables over #define for constants
Declare variables const if not modified after initialization
Use Allman brace style in C++
C++ filenames use lowerCamelCase and must be case-insensitively unique within a build target
C++ type names use UpperCamelCase
Local variables, methods, and namespaces use lowerCamelCase
Global non-static variables not in anonymous namespace use gPrefix lowerCamelCase (e.g., gExample)
Static globals or globals in anonymous namespaces use sPrefix lowerCamelCase
Locally visible static variables start with 's' (e.g., static std::once_flag sFlag;)
Member variables use mPrefix lowerCamelCase; public members may omit but are encouraged to use 'm'
Constants (enums, global/static/function-scope magic numbers) use kPREFIXED_UPPER_SNAKE (e.g., kDIGIT_NUM)
If macros are unavoidable, use UPPER_SNAKE_CASE (prefer constants over #define)
Constructor parameter that conflicts with a public member name gets trailing underscore (foo_)
Literal suffixes should be uppercase (e.g., 1234L not 1234l)
C++: use spaces only; indent 4 spaces
Run clang-format (LLVM style) before submitting; wrap lines at 120 characters
If formatting must be bypassed, use // clang-format off/on around the section
Prefer smart pointers; use unique_ptr for sole ownership, shared_ptr for shared; weak_ptr only in exceptional cases
Do not use deprecated pre-C++11 smart pointers
Use C++ style comments; avoid C comments except special inline cases; prefer // single-line
Capitalize and punctuate full-sentence comments
Follow Doxygen rules: use //! for comments and //!< for members in C++
Disable code with #if/#endif and mnemonic conditions; avoid commented-out code; avoid dead code
Do not throw exceptions across library boundaries
Use least-forceful casts; avoid removing const/volatile; avoid C-style and functional casts (except constructors); p...
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{cpp,cxx,cc,cu}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.{cpp,cxx,cc,cu}: Avoid literal values except for 0, nullptr, true, false; use named constexpr for other literals
Place semicolon of empty for/while loop on a new line
Always use brace-delimited bodies for switch/while/do-for/if/else
Use inline C comments in argument lists when parameter meaning is unclear (e.g., /* checkForErrors = */ false)
Do not use assignment in subexpressions (e.g., if (x = y) ... is forbidden)
Switch on enums should enumerate all values and omit default to catch new values at compile time
Structure switch statements; prohibit fallthrough except between empty cases; each case ends with break or throw; return at end of case not allowed; put break inside braces for compound case
Prefer anonymous namespaces over static for internal linkage of functions
Every defined function must be called at least once (no unused methods)
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{h,hpp,hxx,hh,cuh,cpp,cxx,cc,cu}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Parameter names must be consistent between declarations and definitions
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
🧠 Learnings (2)
📓 Common learnings
Learnt from: zhengd-nv
PR: NVIDIA/TensorRT-LLM#6633
File: cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp:145-155
Timestamp: 2025-08-06T08:18:28.669Z
Learning: In cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp, the existing `mMtxForMap` mutex in DataSenderImpl is sufficient to synchronize measurement file operations in the `release` method, as all file operations occur within the same critical section that protects the `mRequestToSession` map access.
📚 Learning: 2025-08-06T08:18:28.669Z
Learnt from: zhengd-nv
PR: NVIDIA/TensorRT-LLM#6633
File: cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp:145-155
Timestamp: 2025-08-06T08:18:28.669Z
Learning: In cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp, the existing `mMtxForMap` mutex in DataSenderImpl is sufficient to synchronize measurement file operations in the `release` method, as all file operations occur within the same critical section that protects the `mRequestToSession` map access.
Applied to files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
🧬 Code Graph Analysis (1)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (3)
cpp/tests/batch_manager/cacheTransceiverTest.cpp (8)
llmRequest(421-450)llmRequest(421-421)llmRequest(872-903)llmRequest(872-872)llmRequest(905-912)llmRequest(905-905)llmRequest(914-933)llmRequest(914-914)cpp/tensorrt_llm/common/envUtils.cpp (2)
getEnvParallelCacheSend(327-331)getEnvParallelCacheSend(327-327)cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp (1)
requestInfo(152-152)
🔇 Additional comments (2)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (2)
160-160: Proactive trigger looks goodCalling
sendResponse(llmRequest.mRequestId)immediately after enqueuing the ready response reduces latency and improves progress. No issues.
272-281: New synchronization members are reasonableIntroducing
mSendMutexand per-requestmRequestInfoMapis a sensible way to coordinate ready responses with received request-info. Names follow the style guide; types are appropriate.
|
PR_Github #15950 [ run ] completed with state |
|
@chuangz0 can you review the changes to the |
Tabrizian
force-pushed
the
user/imant/fixDisaggBug
branch
from
0653c4b to
72f97a3
Compare
|
/bot run --add-multi-gpu-test --disable-fail-fast |
|
PR_Github #16079 [ run ] triggered by Bot |
|
PR_Github #16079 [ run ] completed with state |
Tabrizian
force-pushed
the
user/imant/fixDisaggBug
branch
from
83babf0 to
4fd2f93
Compare
|
/bot run --add-multi-gpu-test --disable-fail-fast |
|
PR_Github #16216 [ run ] triggered by Bot |
There was a problem hiding this comment.
Actionable comments posted: 0
♻️ Duplicate comments (2)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (2)
138-144: Thread management approach addresses previous safety concernsUsing
mSendThreads.emplace_back()instead of detached threads properly addresses the use-after-free risk that was flagged in previous reviews. The threads are now tracked and joined in the destructor.
107-147: Critical:getBlockHashes()method missing - build will failLine 123 calls
it->second.getBlockHashes()but this method is not defined in theRequestInfoclass in this translation unit, causing a compilation error.Based on past review feedback, you need to implement
RequestInfo::getBlockHashes()and preferably add a move-taking accessor to avoid unnecessary copying since theRequestInfois immediately erased:Add to
RequestInfoclass in the header file:[[nodiscard]] const std::vector<size_t>& getBlockHashes() const noexcept { return mBlockHashes; } // Move-out accessor to avoid copying when the entry is about to be erased: std::vector<size_t> takeBlockHashes() noexcept { return std::move(mBlockHashes); }Then update the usage:
- auto blockHashes = it->second.getBlockHashes(); + auto blockHashes = it->second.takeBlockHashes();
🧹 Nitpick comments (2)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (2)
276-276: Mutex declaration alignment issueThe mutex declaration doesn't follow proper C++ formatting - all three mutexes are declared on a single line which hurts readability.
Apply this diff for better readability:
- std::mutex mCondMutex, mSendMutex, mResponderMutex; + std::mutex mCondMutex; + std::mutex mSendMutex; + std::mutex mResponderMutex;
179-184: Thread cleanup logic is correct but could be more robustThe destructor properly joins all threads before clearing the vector. However, consider adding exception safety around thread joining.
Apply this diff for more robust cleanup:
for (auto& thread : mSendThreads) { - thread.join(); + if (thread.joinable()) + { + thread.join(); + } }
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (4)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp(5 hunks)cpp/tensorrt_llm/batch_manager/llmRequest.cpp(0 hunks)tests/integration/defs/accuracy/test_disaggregated_serving.py(1 hunks)tests/integration/test_lists/waives.txt(0 hunks)
💤 Files with no reviewable changes (2)
- cpp/tensorrt_llm/batch_manager/llmRequest.cpp
- tests/integration/test_lists/waives.txt
🚧 Files skipped from review as they are similar to previous changes (1)
- tests/integration/defs/accuracy/test_disaggregated_serving.py
🧰 Additional context used
📓 Path-based instructions (4)
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh}: In C++, close namespaces with a comment naming the namespace (e.g., } // namespace foo)
Prefer const/constexpr variables over #define for constants
Declare variables const if not modified after initialization
Use Allman brace style in C++
C++ filenames use lowerCamelCase and must be case-insensitively unique within a build target
C++ type names use UpperCamelCase
Local variables, methods, and namespaces use lowerCamelCase
Global non-static variables not in anonymous namespace use gPrefix lowerCamelCase (e.g., gExample)
Static globals or globals in anonymous namespaces use sPrefix lowerCamelCase
Locally visible static variables start with 's' (e.g., static std::once_flag sFlag;)
Member variables use mPrefix lowerCamelCase; public members may omit but are encouraged to use 'm'
Constants (enums, global/static/function-scope magic numbers) use kPREFIXED_UPPER_SNAKE (e.g., kDIGIT_NUM)
If macros are unavoidable, use UPPER_SNAKE_CASE (prefer constants over #define)
Constructor parameter that conflicts with a public member name gets trailing underscore (foo_)
Literal suffixes should be uppercase (e.g., 1234L not 1234l)
C++: use spaces only; indent 4 spaces
Run clang-format (LLVM style) before submitting; wrap lines at 120 characters
If formatting must be bypassed, use // clang-format off/on around the section
Prefer smart pointers; use unique_ptr for sole ownership, shared_ptr for shared; weak_ptr only in exceptional cases
Do not use deprecated pre-C++11 smart pointers
Use C++ style comments; avoid C comments except special inline cases; prefer // single-line
Capitalize and punctuate full-sentence comments
Follow Doxygen rules: use //! for comments and //!< for members in C++
Disable code with #if/#endif and mnemonic conditions; avoid commented-out code; avoid dead code
Do not throw exceptions across library boundaries
Use least-forceful casts; avoid removing const/volatile; avoid C-style and functional casts (except constructors); p...
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{cpp,cxx,cc,cu}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.{cpp,cxx,cc,cu}: Avoid literal values except for 0, nullptr, true, false; use named constexpr for other literals
Place semicolon of empty for/while loop on a new line
Always use brace-delimited bodies for switch/while/do-for/if/else
Use inline C comments in argument lists when parameter meaning is unclear (e.g., /* checkForErrors = */ false)
Do not use assignment in subexpressions (e.g., if (x = y) ... is forbidden)
Switch on enums should enumerate all values and omit default to catch new values at compile time
Structure switch statements; prohibit fallthrough except between empty cases; each case ends with break or throw; return at end of case not allowed; put break inside braces for compound case
Prefer anonymous namespaces over static for internal linkage of functions
Every defined function must be called at least once (no unused methods)
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{h,hpp,hxx,hh,cuh,cpp,cxx,cc,cu}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Parameter names must be consistent between declarations and definitions
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend NVIDIA copyright header (current year) to all source files
Files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
🧠 Learnings (1)
📚 Learning: 2025-08-06T08:18:28.669Z
Learnt from: zhengd-nv
PR: NVIDIA/TensorRT-LLM#6633
File: cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp:145-155
Timestamp: 2025-08-06T08:18:28.669Z
Learning: In cpp/tensorrt_llm/batch_manager/dataTransceiverImpl.cpp, the existing `mMtxForMap` mutex in DataSenderImpl is sufficient to synchronize measurement file operations in the `release` method, as all file operations occur within the same critical section that protects the `mRequestToSession` map access.
Applied to files:
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp
⏰ 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 (2)
cpp/tensorrt_llm/batch_manager/dataTransceiver.cpp (2)
227-237: Good refactoring - per-request state managementThe migration from single current request tracking to per-request mapping in
mRequestInfoMapproperly addresses the concurrency issues. The logic correctly stores theRequestInfoand initializes reference counting for each request.
159-159: Excellent architectural improvementThe immediate call to
sendResponse(llmRequest.mRequestId)after enqueueing the ready response creates a proper event-driven dispatch mechanism. This replaces the previous blocking approach with a more responsive design.
Tabrizian
force-pushed
the
user/imant/fixDisaggBug
branch
from
4fd2f93 to
d536bd2
Compare
|
/bot run --add-multi-gpu-test --disable-fail-fast |
|
PR_Github #16335 [ run ] triggered by Bot |
|
Inside pp_loop, I am somewhat concerned that the changes in this PR might potentially cause a deadlock issue, although I have not yet found a case that would lead to a deadlock. |
Tabrizian
force-pushed
the
user/imant/fixDisaggBug
branch
from
c480836 to
e90e18a
Compare
|
/bot run |
Signed-off-by: Iman Tabrizian <[email protected]>
Signed-off-by: Iman Tabrizian <[email protected]>
Signed-off-by: Iman Tabrizian <[email protected]>
Tabrizian
force-pushed
the
user/imant/fixDisaggBug
branch
from
e90e18a to
169a658
Compare
|
PR_Github #16581 [ run ] triggered by Bot |
|
PR_Github #16577 [ run ] completed with state |
|
PR_Github #16581 [ run ] completed with state |
|
/bot run --only-multi-gpu-test |
|
PR_Github #16652 [ run ] triggered by Bot |
|
PR_Github #16652 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #16688 [ run ] triggered by Bot |
|
PR_Github #16688 [ run ] completed with state |
Signed-off-by: Iman Tabrizian <[email protected]>
Signed-off-by: Iman Tabrizian <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
Signed-off-by: Iman Tabrizian <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
Signed-off-by: Iman Tabrizian <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
Signed-off-by: Iman Tabrizian <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
Signed-off-by: Iman Tabrizian <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
Signed-off-by: Iman Tabrizian <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
Signed-off-by: Iman Tabrizian <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
Signed-off-by: Iman Tabrizian <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
6 participants
Summary by CodeRabbit
Refactor
Performance
Bug Fixes
Tests
Chores
Description
Test Coverage
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.