[#13909][fix] Reuse hidden_states buffer across CUDA graph captures in Eagle3#13930
Draft
achartier wants to merge 6 commits intoNVIDIA:mainfrom
Draft
[#13909][fix] Reuse hidden_states buffer across CUDA graph captures in Eagle3#13930achartier wants to merge 6 commits intoNVIDIA:mainfrom
achartier wants to merge 6 commits intoNVIDIA:mainfrom
Conversation
…Eagle3 Previously, Eagle3OneModelSpecMetadata allocated a new hidden_states buffer (max_num_tokens × hidden_size × num_capture_layers) per CUDA graph capture. This caused memory to grow linearly with the number of graphs captured, wasting ~2.75 GiB for 16 captures on Qwen3-235B. Fix: Always create Eagle3ResourceManager in the one-model flow and reuse its pre-allocated hidden_states buffer instead of allocating a new one per capture. Fixes: NVIDIA#13909 Signed-off-by: Spurthi Sandiri <spurthi@amazon.com>
Remove Memory-Debug logger.debug call and its logger import added for investigation — not needed in the final fix. Signed-off-by: Aurelien Chartier <2567591+achartier@users.noreply.github.com>
Eagle3OneModelDynamicTreeResourceManager does not have a hidden_states attribute, so the unconditional access added in the previous commit would raise AttributeError in dynamic-tree one-model mode. Fall back to allocating a fresh tensor when the resource manager doesn't expose the buffer. Signed-off-by: Aurelien Chartier <2567591+achartier@users.noreply.github.com>
When reusing the resource manager's hidden_states buffer, verify the column dimension matches hidden_size * num_capture_layers. Fails fast with a clear message if the buffer was allocated with different capture layer settings. Signed-off-by: Aurelien Chartier <2567591+achartier@users.noreply.github.com>
Include resource manager type, hidden_size, and capture layers in the assertion to make mismatches easier to diagnose in multi-model setups. Signed-off-by: Aurelien Chartier <2567591+achartier@users.noreply.github.com>
Add hidden_states = None to Eagle3OneModelDynamicTreeResourceManager so the attribute is always present, then replace the hasattr duck-typing with a direct None check. Signed-off-by: Aurelien Chartier <2567591+achartier@users.noreply.github.com>
Collaborator
Author
|
/bot run |
Collaborator
|
PR_Github #47474 [ run ] triggered by Bot. Commit: |
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
2 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.
@coderabbitai summary
Description
Copy of !13920 addressing review
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
Update tava architecture diagram if there is a significant design change in PR.
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
To see a list of available CI bot commands, please comment
/bot help.