Skip to content

[None][fix] kv_cache_manager_v2: count only newly-added scratch blocks in resize #14531

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
lancelly merged 2 commits into from
May 26, 2026
Merged

[None][fix] kv_cache_manager_v2: count only newly-added scratch blocks in resize #14531

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c3ec577
[None][fix] kv_cache_manager_v2: count only newly-added scratch block…
lancelly May 25, 2026
2adfa0a
[None][refactor] kv_cache_manager_v2: clarify scratch slot accounting…
lancelly May 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion tensorrt_llm/runtime/kv_cache_manager_v2/_core/_kv_cache.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,6 +71,7 @@
div_up,
expect_type,
filled_list,
intersect,
make_typed,
map_optional,
stream_wait_events,
Expand Down Expand Up @@ -706,7 +707,10 @@ def resize(self, capacity: int | None, history_length: int | None = None) -> boo
continue
stale_beg, stale_end = stale_ranges[lc]
if enable_scratch:
num_scratch_blocks = len(scratch_ranges[lc])
# Only newly added blocks consume slots below; scratch range may
# extend before old_num_blocks when history_length < old_capacity.
new_block_range = HalfOpenRange(old_num_blocks, new_num_blocks)
num_scratch_blocks = len(intersect(scratch_ranges[lc], new_block_range))
Comment thread
lowsfer marked this conversation as resolved.
Outdated
num_new_normal_blocks = (new_num_blocks - old_num_blocks) - num_scratch_blocks
Comment thread
lowsfer marked this conversation as resolved.
Outdated
num_new_slots[lc] = num_new_normal_blocks * beam_width
else:
Expand Down
Loading