Repository navigation
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What & why
Small top-logprob requests allocated a token-index vector for the entire vocabulary on every call, even for count=0.
Keep at most the requested candidates in a heap and allocate none for zero. Preserve descending-logit/ascending-token ties, log-sum-exp and the selected token logprob.
Area
Engine, CLI or server (
edge0-infer/).How was it tested?
The target, sampling oracle and RNG checks pass. A 262144-token regression compares candidates with a full ordering. Largest measured allocations for counts 0/1/20 are 0/48/960 bytes; main allocates 1048576 bytes even for zero and fails the bound. This measures allocation, not end-to-end generation throughput.
All 14 L0 checks passed in a local integration with the other focused fixes.
git diff --checkpassed.Local C++ environment: Linux x86_64, GCC 14.2, CPU preset. Targeted builds downgrade pre-existing ignored-attributes, dangling-reference, range-loop-construct and misleading-indentation warnings (plus maybe-uninitialized in the existing diagnostics source); these accommodations are not source changes. The all-target build is blocked by existing assets_test.cpp use of std::count_if without . A clean full L0/L1 result, GPU tests and model-asset execution are not claimed.
Checklist
ctest --preset cpu -L "L0|L1"passes locally