[TRTLLM-14903][fix] Free partially-allocated warmup dummy KV blocks and count spec extra tokens in warmup block estimates - #17162

Open

brnguyen2 wants to merge 6 commits into

NVIDIA:mainfrom

brnguyen2:fix/TRTLLM-14903-estimation-warmup-leak

Commits on Aug 5, 2026

[TRTLLM-14903][fix] Free partially-allocated warmup dummy KV blocks and count spec extra tokens in warmup block estimates
brnguyen2
committed
dfe10b9
[None][fix] Launch LMBenchmark client without PYTHONSAFEPATH
brnguyen2
committed
70b1717
[TRTLLM-14903][fix] Clean up warmup block estimate expression
brnguyen2
committed
8e53f2b
[TRTLLM-14903][fix] Document one-engine warmup token accounting
brnguyen2
committed
252d83f
[TRTLLM-14903][fix] Attempt all warmup cleanup and re-raise cleanup failures
brnguyen2
committed
32da14b
[TRTLLM-14903][test] Cover partial add_dummy_requests failure cleanup
brnguyen2
committed
01196c0