[NV TRT RTX EP] Reconfigure memory arena to grow with power of 2 #25800
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline,Windows x64 QNN CI Pipeline |
|
Azure Pipelines successfully started running 5 pipeline(s). |
jywu-msft
approved these changes
Aug 22, 2025
adrianlizarraga
pushed a commit
that referenced
this pull request
Aug 22, 2025
) This reconfiguration is done to NOT allocate tensors with an exact matching size. If that strategy is used a tensor will always trigger an allocation in the arena and not reuse memory since the memory size has to exactly match. This became a big problem with ORT GenAI since the arena grew constantly when prompting with different prompt lengths. No arena shrinkage was triggered to return older tensors. @skottmckay I am happy to be educated of a better usage of the allocators. Issues with this: Since the arena is not used for workspace allocations anymore (using reserve) it will likely not be possible in the future to allocate on a stream and immediately free memory after an enqueue call. That could have enabled workspace sharing in a multi model pipeline very nicely. @chilo-ms can you help merge this.
adrianlizarraga
added a commit
that referenced
this pull request
Aug 25, 2025
### Description Cherry-pick the following PRs into the `rel-1.23.0` branch: - #25592 - #25622 - #25688 - #25729 - #25743 - #25769 - #25745 - #25761 - #25751 - #25716 - #25228 - #25768 - #25788 - #25747 - #25800 - #25818 - #25762 - #25749 - #25831 ### Motivation and Context <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> --------- Co-authored-by: quic-tirupath <[email protected]> Co-authored-by: quic-calvnguy <[email protected]> Co-authored-by: qti-kromero <[email protected]> Co-authored-by: Jeff Kilpatrick <[email protected]> Co-authored-by: Scott McKay <[email protected]> Co-authored-by: David Fan <[email protected]> Co-authored-by: kuanyul-qti <[email protected]> Co-authored-by: Dmitri Smirnov <[email protected]> Co-authored-by: Chi Lo <[email protected]> Co-authored-by: Edward Chen <[email protected]> Co-authored-by: Chunye Wang@AMD <[email protected]> Co-authored-by: minfhong-qti <[email protected]> Co-authored-by: Vishal Agarwal <[email protected]> Co-authored-by: Maximilian Müller <[email protected]> Co-authored-by: Maximilian Müller <[email protected]> Co-authored-by: Changming Sun <[email protected]> Co-authored-by: adrastogi <[email protected]> Co-authored-by: Aditya Rastogi <[email protected]> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
gedoensmax
added a commit
to gedoensmax/onnxruntime
that referenced
this pull request
Sep 2, 2025
…rosoft#25800) This reconfiguration is done to NOT allocate tensors with an exact matching size. If that strategy is used a tensor will always trigger an allocation in the arena and not reuse memory since the memory size has to exactly match. This became a big problem with ORT GenAI since the arena grew constantly when prompting with different prompt lengths. No arena shrinkage was triggered to return older tensors. @skottmckay I am happy to be educated of a better usage of the allocators. Issues with this: Since the arena is not used for workspace allocations anymore (using reserve) it will likely not be possible in the future to allocate on a stream and immediately free memory after an enqueue call. That could have enabled workspace sharing in a multi model pipeline very nicely. @chilo-ms can you help merge this.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This reconfiguration is done to NOT allocate tensors with an exact matching size. If that strategy is used a tensor will always trigger an allocation in the arena and not reuse memory since the memory size has to exactly match.
This became a big problem with ORT GenAI since the arena grew constantly when prompting with different prompt lengths. No arena shrinkage was triggered to return older tensors. @skottmckay I am happy to be educated of a better usage of the allocators.
Issues with this:
Since the arena is not used for workspace allocations anymore (using reserve) it will likely not be possible in the future to allocate on a stream and immediately free memory after an enqueue call. That could have enabled workspace sharing in a multi model pipeline very nicely.
@chilo-ms can you help merge this.