feat: multiple optimization profiles for disjoint input shape regimes#4325
Draft
cehongwang wants to merge 5 commits into
Draft
feat: multiple optimization profiles for disjoint input shape regimes#4325cehongwang wants to merge 5 commits into
cehongwang wants to merge 5 commits into
Conversation
github-actions
Bot
added
component: tests
cehongwang
force-pushed
the
cehongw/multi-optimization-profile
branch
from
f32fed3 to
427643d
Compare
github-actions
Bot
added
the
documentation
cehongwang
force-pushed
the
cehongw/multi-optimization-profile
branch
from
427643d to
2cd4797
Compare
apbose
force-pushed
the
abose/dynamic-shapes-passthrough
branch
from
f907b64 to
9f9055a
Compare
Add support for defining N optimization profiles at compile time via the list-based ``Input.profiles`` API and selecting the active profile at runtime (manual pin by index, or opt-in shape-based auto-selection). - AOT (torch.export) compile path builds one TRT optimization profile per declared profile index; submodules inherit the profile count via propagation across graph breaks. - Python and C++ runtimes expose a matching primitive engine API (set_active_profile / num_optimization_profiles / _active_profile_index / _auto_select_profiles) so the two runtimes remain interchangeable. - Profile selection is exposed through the optimization_profile context manager; auto-selection uses lazy (first-fitting) profile selection. - Backward compatible: engines without declared profiles keep the historical single-profile (dynamic) / no-profile (static) behavior. Includes an example and runtime tests covering dynamic submodule inputs.
cehongwang
force-pushed
the
cehongw/multi-optimization-profile
branch
from
2cd4797 to
a0eeae7
Compare
narendasan
reviewed
Jun 16, 2026
| if (profile_index == active_profile_index) { | ||
| return; | ||
| } | ||
| auto stream = c10::cuda::getCurrentCUDAStream(device_info.id); |
Collaborator
There was a problem hiding this comment.
Does this work with the green context pr?
| } | ||
| const auto& dims = ranges_it->second; | ||
| auto sizes = inputs[i].sizes(); | ||
| for (size_t d = 0; d < sizes.size(); ++d) { |
Collaborator
There was a problem hiding this comment.
Can we cache only what the dynamic dimension is for each profile and its ranges? Then we dont need to search mostly static dims
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
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.
Add support for defining N optimization profiles at compile time via the list-based
Input.profilesAPI and selecting the active profile at runtime (manual pin by index, or opt-in shape-based auto-selection).Includes an example and runtime tests covering dynamic submodule inputs.
Description
Please include a summary of the change and which issue is fixed. Please also include relevant motivation and context. List any dependencies that are required for this change.
Fixes # (issue)
Type of change
Please delete options that are not relevant and/or add your own.
Checklist: