[#5048][enhance] AutoDeploy: Optimize prepare_inputs #6634
Conversation
📝 WalkthroughWalkthroughThe changes introduce device-aware tensor management, host-pinned memory usage, and profiling instrumentation in the Changes
Sequence Diagram(s)sequenceDiagram
participant Executor as ADEngine
participant SeqInfo as SequenceInfo
participant Device as CUDA/Host
Executor->>SeqInfo: build_from_config (pass device string)
Executor->>SeqInfo: _prepare_inputs(new_tokens)
alt If new_tokens provided
SeqInfo->>Device: update_input_ids_with_new_tokens (host-pinned copy)
end
Executor->>SeqInfo: update_pos, assign_cache_loc, nest_sequences
SeqInfo->>Device: _update_position_ids (host-pinned, NVTX profiling)
SeqInfo->>Device: Tensor updates (input_ids, position_ids)
Executor->>Executor: _compute_logits (NVTX profiling)
Estimated code review effort🎯 4 (Complex) | ⏱️ ~40 minutes Suggested reviewers
📜 Recent review detailsConfiguration used: .coderabbit.yaml 📒 Files selected for processing (4)
🚧 Files skipped from review as they are similar to previous changes (2)
🧰 Additional context used📓 Path-based instructions (2)**/*.py📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Files:
**/*.{cpp,h,hpp,cc,cxx,cu,py}📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Files:
⏰ 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)
🔇 Additional comments (14)
✨ 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 comments)
Other keywords and placeholders
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
🧹 Nitpick comments (1)
tensorrt_llm/_torch/auto_deploy/custom_ops/attention_interface.py (1)
152-154: Consider lazy allocation for previous_batch_indices_cuda.The
previous_batch_indices_cudatensor is allocated during initialization but may not be used in all workflows. Consider lazy allocation whenupdate_input_ids_with_new_tokensis first called to save memory.
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (5)
tensorrt_llm/_torch/auto_deploy/compile/backends/torch_cudagraph.py(1 hunks)tensorrt_llm/_torch/auto_deploy/custom_ops/attention_interface.py(11 hunks)tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py(5 hunks)tensorrt_llm/_torch/auto_deploy/shim/demollm.py(1 hunks)tests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_kv_cache.py(1 hunks)
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py: The code developed for TensorRT-LLM should conform to Python 3.8+.
Indent Python code with 4 spaces. Do not use tabs.
Always maintain the namespace when importing in Python, even if only one class or function from a module is used.
Python filenames should use snake_case (e.g., some_file.py).
Python classes should use PascalCase (e.g., class SomeClass).
Python functions and methods should use snake_case (e.g., def my_awesome_function():).
Python local variables should use snake_case. Prefix k for variable names that start with a number (e.g., k_99th_percentile).
Python global variables should use upper snake_case and prefix G (e.g., G_MY_GLOBAL).
Python constants should use upper snake_case (e.g., MY_CONSTANT).
Avoid shadowing variables declared in an outer scope in Python.
Initialize all externally visible members of a Python class in the constructor.
For interfaces that may be used outside a file, prefer docstrings over comments in Python.
Comments in Python should be reserved for code within a function, or interfaces that are local to a file.
Use Google style docstrings for Python classes and functions, which can be parsed by Sphinx.
Attributes and variables in Python can be documented inline; attribute docstrings will be rendered under the docstring for the class.
Avoid using reflection in Python when functionality can be easily achieved without reflection.
When using try-except blocks in Python, limit the except to the smallest set of errors possible.
When using try-except blocks to handle multiple possible variable types in Python, keep the body of the try as small as possible, using the else block to implement the logic.
Files:
tensorrt_llm/_torch/auto_deploy/shim/ad_executor.pytensorrt_llm/_torch/auto_deploy/shim/demollm.pytests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_kv_cache.pytensorrt_llm/_torch/auto_deploy/compile/backends/torch_cudagraph.pytensorrt_llm/_torch/auto_deploy/custom_ops/attention_interface.py
**/*.{cpp,h,cu,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
All TensorRT-LLM Open Source Software code should contain an NVIDIA copyright header that includes the current year. This includes .cpp, .h, .cu, .py, and any other source files which are compiled or interpreted.
Files:
tensorrt_llm/_torch/auto_deploy/shim/ad_executor.pytensorrt_llm/_torch/auto_deploy/shim/demollm.pytests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_kv_cache.pytensorrt_llm/_torch/auto_deploy/compile/backends/torch_cudagraph.pytensorrt_llm/_torch/auto_deploy/custom_ops/attention_interface.py
🧠 Learnings (1)
📚 Learning: in tensorrt-llm's multimodal processing pipeline, shared tensor recovery using `from_shared_tensor()...
Learnt from: yechank-nvidia
PR: NVIDIA/TensorRT-LLM#6254
File: tensorrt_llm/_torch/pyexecutor/model_engine.py:1201-1204
Timestamp: 2025-07-22T09:22:14.726Z
Learning: In TensorRT-LLM's multimodal processing pipeline, shared tensor recovery using `from_shared_tensor()` is only needed during the context phase. Generation requests reuse the already-recovered tensor data and only need to call `strip_for_generation()` to remove unnecessary multimodal data while preserving the recovered tensors. This avoids redundant tensor recovery operations during generation.
Applied to files:
tensorrt_llm/_torch/auto_deploy/shim/demollm.py
🧬 Code Graph Analysis (2)
tensorrt_llm/_torch/auto_deploy/shim/demollm.py (1)
tensorrt_llm/_torch/auto_deploy/custom_ops/attention_interface.py (3)
update_input_pos(526-544)sequence_lengths(228-229)nest_sequences(487-518)
tests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_kv_cache.py (2)
tensorrt_llm/_torch/auto_deploy/custom_ops/attention_interface.py (3)
update_input_pos(526-544)update_position_ids(418-448)reset(351-369)tests/unittest/_torch/auto_deploy/_utils_test/_torch_test_utils.py (1)
all_close(6-21)
🔇 Additional comments (10)
tensorrt_llm/_torch/auto_deploy/shim/demollm.py (1)
118-119: LGTM! API update is correct.The change from
update_postoupdate_input_poscorrectly follows the refactored API inSequenceInfo. The added comment appropriately documents thatnest_sequenceshandles position ID updates.tensorrt_llm/_torch/auto_deploy/compile/backends/torch_cudagraph.py (1)
165-165: Good optimization for asynchronous data transfer.Using
copy_withnon_blocking=Trueenables asynchronous copying when possible, which aligns with the broader performance optimizations in this PR for efficient CPU-GPU data transfers.tests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_kv_cache.py (1)
210-211: Test updates correctly reflect the new API design.The split of position updates into
update_input_pos()andupdate_position_ids()is properly implemented. The comment explaining why explicitupdate_position_ids()calls are needed (whennest_sequencesis not called) provides good documentation.Also applies to: 215-216, 219-220, 224-225, 228-229
tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py (2)
97-102: Proper device handling implementation.The enhanced device handling correctly ensures that CUDA devices have an explicit index, defaulting to the current device when not specified. This prevents potential issues with device mismatches.
190-197: Efficient new token handling without CPU transfers.The refactored logic efficiently handles new tokens by:
- Using placeholder tokens (-1) to mark positions for new tokens
- Tracking batch indices for proper token placement
- Utilizing the new
update_input_ids_with_new_tokensmethod for in-place updatesThis avoids unnecessary CPU transfers and aligns with the performance optimization goals.
Also applies to: 212-218
tensorrt_llm/_torch/auto_deploy/custom_ops/attention_interface.py (5)
93-94: Excellent device-aware tensor management implementation.The changes introduce proper device-aware initialization with:
- Explicit device specification for all tensors
- Host-pinned memory for
seq_len_hostandinput_pos_hostto optimize CPU-GPU transfers- Maintenance of full-size tensor variants to avoid reallocation overhead
- Consistent 1D tensor shapes for flattened sequence handling
These optimizations align perfectly with the PR's performance goals.
Also applies to: 97-98, 130-154
417-448: Highly optimized position ID update implementation.The refactored
update_position_idsmethod includes excellent optimizations:
- NVTX profiling for performance analysis
- Fast path for generation mode (all seq_len == 1)
- Host-side computation to avoid device synchronization
- Non-blocking copies for better CPU-GPU overlap
- Conditional reallocation support
This is a well-thought-out performance improvement.
458-485: Efficient in-place token update implementation.The
update_input_ids_with_new_tokensmethod efficiently updates tokens using:
- Masked scatter for in-place updates where tokens are marked with -1
- Sorted indices to ensure correct alignment
- Host-pinned memory for index transfers
- Minimal memory allocations
This approach avoids unnecessary copies and aligns with the performance optimization goals.
487-518: Well-structured sequence nesting with flexible memory management.The refactored
nest_sequencesmethod provides:
- Flexible memory management with
allow_reallocparameter- Centralized sequence length updates via
_update_sequence_lengths- Efficient non-blocking copies with pinned memory
- Automatic position ID updates maintaining consistency
Good separation of concerns and performance optimization.
525-544: Clean refactoring with performance optimizations.The renamed
update_input_posmethod includes:
- More descriptive naming
- NVTX profiling instrumentation
- Host-pinned memory for efficient CPU-GPU transfers
- Non-blocking copy operations
These changes align with the overall performance optimization strategy.
galagam
force-pushed
the
user/galagam/ad-prepare-inputs-opt
branch
from
96c77bd to
3899933
Compare
galagam
changed the title
galagam
changed the title
galagam
force-pushed
the
user/galagam/ad-prepare-inputs-opt
branch
from
3899933 to
413cd3f
Compare
galagam
changed the title
|
/bot run |
galagam
force-pushed
the
user/galagam/ad-prepare-inputs-opt
branch
from
413cd3f to
ac4ae69
Compare
|
/bot run |
1 similar comment
|
/bot run |
|
PR_Github #14323 [ run ] triggered by Bot |
|
PR_Github #14323 [ run ] completed with state |
galagam
force-pushed
the
user/galagam/ad-prepare-inputs-opt
branch
from
ac4ae69 to
f8aa994
Compare
galagam
changed the title
|
/bot run |
|
PR_Github #14442 [ run ] triggered by Bot |
|
PR_Github #14442 [ run ] completed with state |
galagam
force-pushed
the
user/galagam/ad-prepare-inputs-opt
branch
from
f8aa994 to
f84a16f
Compare
|
/bot run |
GitHub Bot Help
Provide a user friendly way for developers to interact with a Jenkins server. Run See details below for each supported subcommand.
Launch build/test pipelines. All previously running jobs will be killed.
kill
Kill all running builds associated with pull request. skip
Skip testing for latest commit on pull request. reuse-pipeline
Reuse 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. |
|
/bot run |
|
PR_Github #14503 [ run ] triggered by Bot |
|
PR_Github #14503 [ run ] completed with state |
avoid gpu->cpu transfer when using overlap scheduler Signed-off-by: Suyog Gupta <[email protected]> Signed-off-by: Suyog Gupta <[email protected]> prealloc Signed-off-by: Suyog Gupta <[email protected]> optimize prepare input Signed-off-by: Suyog Gupta <[email protected]> more changes Signed-off-by: Suyog Gupta <[email protected]> remove more syncs Signed-off-by: Suyog Gupta <[email protected]> clean up Signed-off-by: Suyog Gupta <[email protected]> clean up position_id handling Signed-off-by: Suyog Gupta <[email protected]> revert spurious changes Signed-off-by: Suyog Gupta <[email protected]> revert spurious print Signed-off-by: Suyog Gupta <[email protected]> Update export_to_gm.py Signed-off-by: Suyog Gupta <[email protected]> revert spurious print Signed-off-by: Suyog Gupta <[email protected]> revert nvtx debug Signed-off-by: Suyog Gupta <[email protected]> revert nvtx debug Signed-off-by: Suyog Gupta <[email protected]>
update_position_ids is the bottleneck - add nvtx markers and small list comprehension perf improvement Signed-off-by: Gal Hubara Agam <[email protected]> optimize prepare_list for generation mode Signed-off-by: Gal Hubara Agam <[email protected]> bugfix in _update_position_ids optimized code Signed-off-by: Gal Hubara Agam <[email protected]>
Signed-off-by: Gal Hubara Agam <[email protected]>
|
/bot run |
|
PR_Github #14697 [ run ] triggered by Bot |
galagam
force-pushed
the
user/galagam/ad-prepare-inputs-opt
branch
from
f84a16f to
a0bcb2c
Compare
|
PR_Github #14697 [ run ] completed with state |
|
/bot run |
|
PR_Github #14702 [ run ] triggered by Bot |
|
PR_Github #14702 [ run ] completed with state |
Summary by CodeRabbit
New Features
Performance Improvements
Bug Fixes
Description
Optimize prepare_inputs routine in AutoDeploy, as part of the effort to reduce the performance gap compared to the default backend.
This PR includes two major fixes, and some other minor tweaks:
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.