[None][chore] Fix MAX_UTILIZATION reuse token budget on main#15066
Draft
brb-nv wants to merge 2 commits into
Draft
[None][chore] Fix MAX_UTILIZATION reuse token budget on main#15066brb-nv wants to merge 2 commits into
brb-nv wants to merge 2 commits into
Conversation
getNeededBlocksOneStep (MAX_UTILIZATION) derived estimatedReusableTokens from reusableBlocksAllocated (active-ref blocks only), the same conservative count correctly used for the block/capacity budget since NVIDIA#11731. In steady state, prefixes from completed requests are free-but-cached (no refs), so reusableBlocksAllocated drops to ~0 while reusableBlocksAll stays large. The micro batch scheduler then under-credits reuse, FCFS charges full context chunks, one request saturates max_num_tokens, and context requests serialize to ~1/iter -- inflating TTFT at high concurrency. Decouple the two budgets in getNeededBlocksOneStep: - Block/capacity budget keeps reusableBlocksAllocated (avoids the NVIDIA#11731 double-count / over-admission). - Token/compute budget now uses reusableBlocksAll, since cached KV is recovered via prepopulatedPromptLen regardless of ref state. This mirrors the accounting getRemainingBlocksToCompletion already uses for GUARANTEED_NO_EVICT, removing the discrepancy between the two policies. Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com> (cherry picked from commit 8d1df0d)
Add a regression test for getNeededBlocksOneStep showing that a free-but-cached shared prefix (reusableBlocksAllocated == 0, reusableBlocksAll > 0) now credits estimatedReusableTokens from reusableBlocksAll, while the block/capacity budget still credits only allocated reuse. The test models the FCFS micro-batch scheduler's compute-aware admission for a fixed token budget: with the conservative (allocated-only) estimate only one context request fits; with the all-cached estimate four fit. It fails before the fix (estimatedReusableTokens == 0) and passes after. Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>
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
1 participant
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.
Description
getNeededBlocksOneStep(MAX_UTILIZATION) derivedestimatedReusableTokensfromreusableBlocksAllocated(active-ref blocks only), the same conservative count correctly used for the block/capacity budget since #11731. In steady state, prefixes from completed requests are free-but-cached (no refs), soreusableBlocksAllocateddrops to ~0 whilereusableBlocksAllstays large. The micro batch scheduler then under-credits reuse, FCFS charges full context chunks, one request saturates max_num_tokens, and context requests serialize to ~1/iter -- inflating TTFT at high concurrency.Decouple the two budgets in
getNeededBlocksOneStep:reusableBlocksAllocated(avoids the [None][fix] Fix overly aggressive capacity scheduler #11731 double-count / over-admission).reusableBlocksAll, since cached KV is recovered viaprepopulatedPromptLenregardless of ref state.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)
If PR introduces API changes, an appropriate PR label is added - either
api-compatibleorapi-breaking. Forapi-breaking, includeBREAKINGin the PR title.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.