[https://nvbugs/5537878][fix] Reserve an extra slot for padded batch #7998
Conversation
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5537878
branch
from
fdf4d6d to
7743a7f
Compare
|
/bot run |
📝 WalkthroughWalkthroughUpdates the CUDA graph padding path: padding uses runtime_draft_len; slot management now reuses a fixed dummy request ID without allocating a new slot; Eagle3 reserves one extra slot for the dummy; executor removes the SeqSlotManager usage; a redundant constant is removed; adds an integration test validating CUDA graph padding. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
participant T as Test
participant L as LLM (Eagle3)
participant E as PyExecutor
participant R as Eagle3ResourceManager
participant S as SlotManager
participant C as CudaGraphRunner
participant G as Engine
T->>L: generate(prompts, max_batch_size=4, cuda_graph.padding=true)
L->>E: submit requests
E->>R: prepare resources
R->>S: init with max_num_requests + 1
E->>C: build batch
C->>C: if batch_size < max_batch_size<br/>add dummy request (fixed ID)
C->>S: add_slot(request_id=dummy_id)
Note over S: New: reuse existing mapping<br/>for dummy_id (no new slot)
S-->>C: slot for dummy_id
C->>G: launch with padded batch<br/>(dummy uses runtime_draft_len)
G-->>C: results
C-->>E: trimmed results (exclude dummy)
E-->>L: responses
L-->>T: outputs
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes Pre-merge checks and finishing touches❌ Failed checks (2 warnings)
✅ Passed checks (1 passed)
✨ 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. 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: 0
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
tensorrt_llm/_torch/speculative/eagle3.py (1)
37-51: Reserve +1 slot: also extend seq_lens/start_indices to cover the dummy slotReserving an extra SlotManager slot is correct, but seq_lens/start_indices are still sized to max_num_requests. If the dummy maps to the extra slot ID, lookups will KeyError. Initialize these dicts with max_num_requests + 1 as well.
Apply this change:
- self.seq_lens = {i: 0 for i in range(max_num_requests)} - # start indices of each slot - self.start_indices = {i: 0 for i in range(max_num_requests)} + num_slots = max_num_requests + 1 + self.seq_lens = {i: 0 for i in range(num_slots)} + # start indices of each slot + self.start_indices = {i: 0 for i in range(num_slots)}
🧹 Nitpick comments (2)
tensorrt_llm/_torch/pyexecutor/resource_manager.py (1)
1045-1055: Reorder checks to allow dummy-slot reuse even when no free slots; avoid assert for production safetyCurrent order raises “No free slots” before reusing an existing dummy slot mapping. Also, relying on assert can be elided with -O. Prefer explicit guard and return for the dummy, and error for duplicate non-dummy IDs.
Apply this diff:
def add_slot(self, request_id: int): - if len(self.free_slots) == 0: - raise ValueError("No free slots") - if request_id in self.slot_mapping: - # CUDA graph dummy request could be added for different batches, - # but we only need to reserve slot for it once. - from .cuda_graph_runner import CUDA_GRAPH_DUMMY_REQUEST_ID - assert request_id == CUDA_GRAPH_DUMMY_REQUEST_ID - return self.slot_mapping[request_id] + if request_id in self.slot_mapping: + # CUDA graph dummy request can be added for different batches, + # but we only need to reserve a slot for it once. + from .cuda_graph_runner import CUDA_GRAPH_DUMMY_REQUEST_ID + if request_id != CUDA_GRAPH_DUMMY_REQUEST_ID: + # Duplicate non-dummy IDs indicate a logic bug upstream. + raise ValueError(f"Request {request_id} already has a slot") + return self.slot_mapping[request_id] + if len(self.free_slots) == 0: + raise ValueError("No free slots") slot = self.free_slots.pop() self.slot_mapping[request_id] = slot return slottests/unittest/_torch/speculative/test_eagle3.py (1)
372-429: Mark as high CUDA memory to align with other heavy testsThis test loads large models and already gates on device memory; add the high-memory mark to integrate with CI filters like the neighboring tests.
Apply this diff:
-@pytest.mark.parametrize("disable_overlap_scheduler", [True, False]) +@pytest.mark.high_cuda_memory +@pytest.mark.parametrize("disable_overlap_scheduler", [True, False]) def test_eagle3_cuda_graph_padding(disable_overlap_scheduler: bool):
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (7)
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py(1 hunks)tensorrt_llm/_torch/pyexecutor/model_engine.py(0 hunks)tensorrt_llm/_torch/pyexecutor/py_executor.py(0 hunks)tensorrt_llm/_torch/pyexecutor/resource_manager.py(1 hunks)tensorrt_llm/_torch/speculative/eagle3.py(1 hunks)tensorrt_llm/_torch/speculative/model_drafter.py(0 hunks)tests/unittest/_torch/speculative/test_eagle3.py(1 hunks)
💤 Files with no reviewable changes (3)
- tensorrt_llm/_torch/speculative/model_drafter.py
- tensorrt_llm/_torch/pyexecutor/model_engine.py
- tensorrt_llm/_torch/pyexecutor/py_executor.py
🧰 Additional context used
📓 Path-based instructions (3)
**/*.{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:
tensorrt_llm/_torch/speculative/eagle3.pytensorrt_llm/_torch/pyexecutor/cuda_graph_runner.pytests/unittest/_torch/speculative/test_eagle3.pytensorrt_llm/_torch/pyexecutor/resource_manager.py
**/*.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/speculative/eagle3.pytensorrt_llm/_torch/pyexecutor/cuda_graph_runner.pytests/unittest/_torch/speculative/test_eagle3.pytensorrt_llm/_torch/pyexecutor/resource_manager.py
**/*.{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:
tensorrt_llm/_torch/speculative/eagle3.pytensorrt_llm/_torch/pyexecutor/cuda_graph_runner.pytests/unittest/_torch/speculative/test_eagle3.pytensorrt_llm/_torch/pyexecutor/resource_manager.py
🧠 Learnings (1)
📚 Learning: 2025-08-08T04:10:19.038Z
Learnt from: djns99
PR: NVIDIA/TensorRT-LLM#6728
File: cpp/tensorrt_llm/plugins/mixtureOfExperts/mixtureOfExpertsPlugin.cpp:966-966
Timestamp: 2025-08-08T04:10:19.038Z
Learning: TensorRT plugins currently don't support padding functionality, and TensorRT is not getting new features (in maintenance mode). This means that duplicating parameters like mExpertHiddenSize in function calls, even with TODO comments, can be acceptable as pragmatic solutions within these constraints.
Applied to files:
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py
🧬 Code graph analysis (4)
tensorrt_llm/_torch/speculative/eagle3.py (1)
tensorrt_llm/_torch/pyexecutor/resource_manager.py (1)
SlotManager(1025-1067)
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (1)
tensorrt_llm/_torch/pyexecutor/model_engine.py (1)
runtime_draft_len(474-475)
tests/unittest/_torch/speculative/test_eagle3.py (2)
tensorrt_llm/llmapi/llm_args.py (4)
KvCacheConfig(976-1110)CudaGraphConfig(108-165)EagleDecodingConfig(442-480)speculative_model_dir(1438-1439)tensorrt_llm/llmapi/llm.py (3)
LLM(1052-1068)generate(241-319)shutdown(737-744)
tensorrt_llm/_torch/pyexecutor/resource_manager.py (1)
tensorrt_llm/executor/result.py (1)
request_id(556-557)
⏰ 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 (1)
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (1)
306-317: Use runtime_draft_len for dummy padding requestSwitching to engine.runtime_draft_len ensures correct draft length (0 when spec decode disabled). LGTM.
Please confirm runtime_draft_len cannot change mid-run; if it can, padding_dummy_request should be (re)created per change to avoid stale KV sizing.
|
PR_Github #19950 [ run ] triggered by Bot |
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5537878
branch
from
7743a7f to
4755dcd
Compare
ziyixiong-nv
changed the title
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5537878
branch
from
4755dcd to
d897b5b
Compare
|
/bot run |
|
PR_Github #19953 [ run ] triggered by Bot |
|
PR_Github #19950 [ run ] completed with state |
|
PR_Github #19953 [ run ] completed with state |
|
/bot run |
|
PR_Github #20005 [ run ] triggered by Bot |
|
PR_Github #20005 [ run ] completed with state |
|
/bot run |
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5537878
branch
from
d897b5b to
e0d32a2
Compare
|
/bot run |
|
PR_Github #20070 [ run ] triggered by Bot |
|
PR_Github #20071 [ run ] triggered by Bot |
|
PR_Github #20070 [ run ] completed with state |
|
PR_Github #20071 [ run ] completed with state |
|
PR_Github #20336 [ run ] triggered by Bot |
|
PR_Github #20336 [ run ] completed with state |
|
/bot run |
|
PR_Github #20365 [ run ] triggered by Bot |
|
PR_Github #20365 [ run ] completed with state |
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5537878
branch
from
71abfea to
b222c13
Compare
|
/bot run |
|
PR_Github #20436 [ run ] triggered by Bot |
|
PR_Github #20436 [ run ] completed with state |
|
/bot run |
|
PR_Github #20504 [ run ] triggered by Bot |
|
PR_Github #20504 [ run ] completed with state |
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5537878
branch
from
b222c13 to
8ec7d9b
Compare
|
/bot run |
|
PR_Github #20515 [ run ] triggered by Bot |
|
PR_Github #20515 [ run ] completed with state |
Signed-off-by: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com>
Signed-off-by: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com>
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5537878
branch
from
8ec7d9b to
21a4958
Compare
|
/bot run |
|
PR_Github #20561 [ run ] triggered by Bot |
|
PR_Github #20561 [ run ] completed with state |
|
/bot run |
|
PR_Github #20586 [ run ] triggered by Bot |
|
PR_Github #20586 [ run ] completed with state |
…VIDIA#7998) Signed-off-by: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com>
…VIDIA#7998) Signed-off-by: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com>
…VIDIA#7998) Signed-off-by: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com>
3 participants
Summary by CodeRabbit
Description
There could be dummy request for padding batch when using CUDA graph, so we also need to reserve the slot for the dummy request.
When some requests don't have draft requests, we'll need to pad the draft batches and reserve slots for them in the spec_resource_manager. However, the slots are not released until https://github.com/NVIDIA/TensorRT-LLM/blob/main/tensorrt_llm/_torch/pyexecutor/py_executor.py#L1811 . To resolve this issue, we can reserve one extra slot for the dummy request at the initialization phase.
There's also another alternative solution. We can release the slots in spec_resource_manager earlier in https://github.com/NVIDIA/TensorRT-LLM/blob/main/tensorrt_llm/_torch/speculative/model_drafter.py#L222 ,L226 and L236 for the requests that don't need draft requests, but it might be inefficient enough as we need to operate the slots more frequently.
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.