[https://nvbugs/5451342][fix] Use runtime max_batch_size when cuda_graph_config.max_batch_size is not provided in trtllm-bench #7031
Conversation
… provided Signed-off-by: Jiagan Cheng <[email protected]>
📝 WalkthroughWalkthroughAdds PyTorch-specific post-processing in get_llm_args: after applying extra LLM API options, it adjusts cuda_graph_config for the PyTorch backend, defaulting cuda_graph_config.max_batch_size from settings_config.max_batch_size when not otherwise specified, and returns the updated args. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
participant Caller
participant Configuration as Configuration.get_llm_args
participant Utils as update_llm_args_with_extra_options
Caller->>Configuration: get_llm_args()
Configuration->>Utils: merge llm_args with extra_llm_api_options
Utils-->>Configuration: updated_llm_args
alt backend == "pytorch"
Note over Configuration: Extract cuda_graph_config
Configuration->>Configuration: if no batch_sizes/max_batch_size<br/>set cuda_graph_config.max_batch_size = settings_config.max_batch_size
Configuration->>Configuration: write back cuda_graph_config
else backend != "pytorch"
Note over Configuration: No additional processing
end
Configuration-->>Caller: return updated_llm_args
Estimated code review effort🎯 2 (Simple) | ⏱️ ~10 minutes Possibly related PRs
Suggested reviewers
Tip 🔌 Remote MCP (Model Context Protocol) integration is now available!Pro plan users can now connect to remote MCP servers from the Integrations page. Connect with popular remote MCPs such as Notion and Linear to add more context to your reviews and chats. ✨ 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: 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/llmapi/llm_args.py (1)
150-153: Off-by-one omits power-of-two 256 (and others) when max_batch_size isn’t a power-of-twoThe intent says “Add powers of 2 up to max_batch_size,” but the current range excludes the endpoint. For example, with max_batch_size=300, 256 should be included but isn’t. Fix by making the upper bound inclusive.
Apply this diff:
- # Add powers of 2 up to max_batch_size - batch_sizes += [ - 2**i for i in range(8, math.floor(math.log(max_batch_size, 2))) - ] + # Add powers of 2 up to max_batch_size (inclusive) + batch_sizes += [ + 2**i for i in range(8, math.floor(math.log2(max_batch_size)) + 1) + ]
🧹 Nitpick comments (2)
tensorrt_llm/llmapi/llm_args.py (2)
2233-2237: Nit: clarify comment wording and avoid magic number
- “base_llm_args.max_batch_size” can confuse readers; the field here is self.max_batch_size (runtime).
- Consider extracting 128 into a named constant to avoid magic numbers and centralize defaults.
Apply this diff to tighten the comment:
- # Use the max batch size from: - # 1. cuda_graph_config.max_batch_size, if provided, - # 2. base_llm_args.max_batch_size, if provided, - # 3. default value 128. + # Resolve max batch size from (in order of precedence): + # 1. cuda_graph_config.max_batch_size (if provided), + # 2. self.max_batch_size (runtime, if provided), + # 3. default value 128.Optionally, define a constant once and use it here:
# Near the top-level of this module (e.g., after imports) DEFAULT_CUDA_GRAPH_MAX_BATCH_SIZE = 128And then:
max_batch_size = ( config.max_batch_size or self.max_batch_size or DEFAULT_CUDA_GRAPH_MAX_BATCH_SIZE )
2236-2236: Fix Ruff E501 (line too long) by wrapping the expressionRuff flagged this line as exceeding 120 chars. Break it across multiple lines using parentheses or intermediate variables so it passes lint without altering behavior.
📜 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 (1)
tensorrt_llm/llmapi/llm_args.py(1 hunks)
🧰 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:
tensorrt_llm/llmapi/llm_args.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/llmapi/llm_args.py
🪛 Ruff (0.12.2)
tensorrt_llm/llmapi/llm_args.py
2236-2236: Line too long (124 > 120)
(E501)
⏰ 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/llmapi/llm_args.py (1)
2233-2242: Precedence fix restores expected behavior and propagates resolved valueGood call using the fallback chain (config.max_batch_size -> self.max_batch_size -> 128) and writing the resolved value back to config.max_batch_size. This mirrors pre-1.0 defaulting and ensures downstream consumers (e.g., get_pytorch_backend_config) see a consistent, non-zero max_batch_size and a generated batch size list.
|
/bot run |
|
PR_Github #15757 [ run ] triggered by Bot |
|
PR_Github #15757 [ run ] completed with state |
|
/bot run |
|
PR_Github #15863 [ run ] triggered by Bot |
|
PR_Github #15863 [ run ] completed with state |
|
/bot run |
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Nitpick comments (2)
tests/integration/defs/accuracy/test_llm_api_pytorch.py (2)
2018-2018: Confirm KV cache memory fraction (0.8) won’t cause warmup/CG OOMsBumping free_gpu_memory_fraction to 0.8 reduces headroom for non-KV allocations and larger CUDA graph captures (now defaulting to runtime max_batch_size). Most tests here use 0.6–0.75. Please sanity-check this on smallest target GPUs that run this suite.
If you want to align with existing tests, here’s a minimal change:
- kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.8) + kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
2024-2026: Add a regression assertion to lock in CUDA Graph fallback precedenceGiven the PR restores “use runtime max_batch_size when config.max_batch_size is None”, add a quick assertion to prevent regressions in future refactors.
Apply this diff inside the with-block before running tasks:
with LLM(f"{llm_models_root()}/Qwen3/Qwen3-8B", tensor_parallel_size=tp_size, pipeline_parallel_size=pp_size, moe_expert_parallel_size=ep_size, kv_cache_config=kv_cache_config, **pytorch_config, enable_attention_dp=attention_dp) as llm: + # Validate: when CudaGraphConfig.max_batch_size is unspecified, + # it should default to the runtime (llm.args) max_batch_size. + assert llm.args.cuda_graph_config is not None + assert llm.args.cuda_graph_config.max_batch_size == llm.args.max_batch_size task = CnnDailymail(self.MODEL_NAME) task.evaluate(llm) task = MMLU(self.MODEL_NAME) task.evaluate(llm)
📜 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 (1)
tests/integration/defs/accuracy/test_llm_api_pytorch.py(1 hunks)
🧰 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_llm_api_pytorch.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_llm_api_pytorch.py
🧠 Learnings (1)
📓 Common learnings
Learnt from: eopXD
PR: NVIDIA/TensorRT-LLM#6768
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:577-579
Timestamp: 2025-08-20T06:56:02.859Z
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.
⏰ 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
|
PR_Github #15911 [ run ] triggered by Bot |
|
PR_Github #15911 [ run ] completed with state |
|
/bot run |
|
PR_Github #15974 [ run ] triggered by Bot |
|
PR_Github #15974 [ run ] completed with state |
Signed-off-by: Jiagan Cheng <[email protected]>
jiaganc
force-pushed
the
dev-jiaganc-fix-cuda-graph-max-batchsize
branch
from
bd35a62 to
93d792d
Compare
|
/bot run |
|
PR_Github #15989 [ run ] triggered by Bot |
|
PR_Github #15989 [ run ] completed with state |
|
/bot run |
|
PR_Github #16028 [ run ] triggered by Bot |
|
PR_Github #16028 [ run ] completed with state |
|
The extra graphs cause lots of OOM in CI tests. Maybe we should use |
Signed-off-by: Jiagan Cheng <[email protected]>
|
/bot run |
jiaganc
changed the title
|
PR_Github #16051 [ run ] triggered by Bot |
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)
tensorrt_llm/bench/dataclasses/configuration.py (3)
83-86: Python 3.8 compatibility: replace dict union operator.The codebase targets Python 3.8+, but the dict union operator “|” requires Python 3.9+. Use dict unpacking to remain compatible.
Apply this diff:
- llm_args["kv_cache_config"] = backend_cache_config | kv_cache_config + llm_args["kv_cache_config"] = {**backend_cache_config, **kv_cache_config}
65-67: Potential AttributeError when decoding_config is None.RuntimeConfig.decoding_config is Optional, but get_llm_args unconditionally calls self.decoding_config.get_decoding_config(). If decoding_config is None, this will crash.
Apply this diff to avoid the crash:
- "decoding_config": - self.decoding_config.get_decoding_config(), + "decoding_config": + (self.decoding_config.get_decoding_config() + if self.decoding_config is not None else None),
1-1: Add NVIDIA copyright header.All source files must prepend the current-year NVIDIA copyright header.
Apply this diff:
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
🧹 Nitpick comments (2)
tensorrt_llm/bench/dataclasses/configuration.py (2)
120-125: Consider capping graph cache or exposing a bench-only cap to mitigate CI OOMs.Bench defaults set cuda_graph_cache_size to 1000, and combined with large inferred max_batch_size this can blow up memory in CI. A modest bench-only cap (e.g., 128–256) or an env-tunable ceiling could help.
If you want, I can wire an env var like TLLM_BENCH_CUDA_GRAPH_DEFAULT_MAX_CAP to clamp inferred max_batch_size=min(runtime_max, cap).
95-101: Clarify emptybatch_sizesas unset in bench configurationTo align with the LLM-API’s handling—where an empty list (
[]) is treated the same as “not provided”—update the check intensorrt_llm/bench/dataclasses/configuration.py(around lines 95–101) to use truthiness rather thanis not None. For example:# Use runtime max_batch_size if neither field is supplied. - batch_sizes_set = cuda_graph_config.get("batch_sizes", None) is not None - max_batch_size_set = cuda_graph_config.get("max_batch_size", None) is not None - if not batch_sizes_set and not max_batch_size_set: + batch_sizes = cuda_graph_config.get("batch_sizes") + max_batch_size = cuda_graph_config.get("max_batch_size") + if not batch_sizes and not max_batch_size: cuda_graph_config[ "max_batch_size"] = self.settings_config.max_batch_size• This treats both
Noneand[]as “unset,” ensuring that an explicit empty list falls through to the default-injection branch.
• No existing unit tests cover the empty-list case in the bench code—consider adding one to prevent regressions.
📜 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 (1)
tensorrt_llm/bench/dataclasses/configuration.py(1 hunks)
🧰 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:
tensorrt_llm/bench/dataclasses/configuration.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/bench/dataclasses/configuration.py
🔇 Additional comments (1)
tensorrt_llm/bench/dataclasses/configuration.py (1)
87-89: Return-path looks good after extra options merge.The flow of updating llm_args with extra_llm_api_options, post-processing, then returning updated_llm_args is correct and aligns with the PR objective.
|
PR_Github #16051 [ run ] completed with state |
|
/bot run |
|
PR_Github #16505 [ run ] triggered by Bot |
|
PR_Github #16505 [ run ] completed with state |
…aph_config.max_batch_size is not provided in trtllm-bench (#7031) Signed-off-by: Jiagan Cheng <[email protected]>
…aph_config.max_batch_size is not provided in trtllm-bench (NVIDIA#7031) Signed-off-by: Jiagan Cheng <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
…aph_config.max_batch_size is not provided in trtllm-bench (NVIDIA#7031) Signed-off-by: Jiagan Cheng <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
…aph_config.max_batch_size is not provided in trtllm-bench (NVIDIA#7031) Signed-off-by: Jiagan Cheng <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
…aph_config.max_batch_size is not provided in trtllm-bench (NVIDIA#7031) Signed-off-by: Jiagan Cheng <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
…aph_config.max_batch_size is not provided in trtllm-bench (NVIDIA#7031) Signed-off-by: Jiagan Cheng <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
…aph_config.max_batch_size is not provided in trtllm-bench (NVIDIA#7031) Signed-off-by: Jiagan Cheng <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
…aph_config.max_batch_size is not provided in trtllm-bench (NVIDIA#7031) Signed-off-by: Jiagan Cheng <[email protected]> Signed-off-by: Wangshanshan <[email protected]>
Summary by CodeRabbit
Description
In 0.21.0,
cuda_graph_max_batch_sizeis a field inTorchLlmArgs. When it's not specified by user,trtllm-benchwill use the value fromget_settings, which use runtimemax_batch_sizeas the default value.However, in 1.0.0, CUDA graph configs are move to
CudaGraphConfig, and all the configs fromget_settingswill be overwritten by configs fromextra_llm_api_options. In the bug's case,cuda_graph_config.max_batch_sizeis set toNone. Therefore,validate_cuda_graph_configcan only use the default value 128 to generate the batch size list. Lacking CUDA graphs larger than 128 causes the perfermance regression.This PR makes
trtllm-benchuse runtimemax_batch_sizewhencuda_graph_config.max_batch_sizeis not provided.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.