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

[lancelly]

Summary

Fixes IndexError: pop from empty list in KVCache.resize() when SWA scratch reuse is enabled and history_length < old_capacity (rewind window allowed since #14403). Reproduces with DSv4 + MTP + block_reuse.

Root cause

In tensorrt_llm/runtime/kv_cache_manager_v2/_core/_kv_cache.py, the enable_scratch branch of resize() was computing new-slot demand from the full scratch_ranges[lc]:

num_scratch_blocks = len(scratch_ranges[lc])
num_new_normal_blocks = (new_num_blocks - old_num_blocks) - num_scratch_blocks
num_new_slots[lc] = num_new_normal_blocks * beam_width

But the block construction loop below only iterates [old_num_blocks, new_num_blocks) and only those blocks pop slots:

for ordinal in typed_range(old_num_blocks, new_num_blocks):
    ...
    if enable_scratch and ordinal in scratch_ranges[lc]:
        continue
    slot = slots[lc].pop()

compute_scratch_range() derives its lower bound from history_length, not from old_capacity. After #14403 (Fix MTP by scratch reuse rewind), history_length is allowed to be smaller than old_capacity by up to max_rewind_len, so scratch_ranges[lc] can start before old_num_blocks. When that happens, scratch blocks that sit in the old range are still subtracted from the allocation count but are never re-encountered in the construction loop, so the allocation under-requests by exactly that overlap and the loop's slots[lc].pop() underflows.

Concrete example

tokens_per_block = 128
old_capacity     = 1024  -> old_num_blocks = 8
new_capacity     = 1152  -> new_num_blocks = 9
history_length   = 896   -> input_range starts at block 7
scratch_ranges[lc] = [7, 8)        # block 7 is OLD, not newly added
newly added range  = [8, 9)        # construction loop iterates only block 8

old code:    num_new_slots = (9-8) - 1 = 0
loop pops:   block 8 needs 1 slot   -> IndexError

Fix

Intersect scratch_ranges[lc] with [old_num_blocks, new_num_blocks) before counting, so the request matches what the loop will actually consume.

_take_excess_scratch_slots is intentionally unchanged — it sizes the scratch slot pool to serve the full scratch range, not just the newly-added portion.

[lancelly]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #50206 [ run ] triggered by Bot. Commit: c3ec577 Link to invocation

[tensorrt-cicd]

PR_Github #50206 [ run ] completed with state SUCCESS. Commit: c3ec577
/LLM/main/L0_MergeRequest_PR pipeline #39744 completed with status: 'SUCCESS'

CI Report

Link to invocation

[lancelly]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #50217 [ run ] triggered by Bot. Commit: 2adfa0a Link to invocation

[tensorrt-cicd]

PR_Github #50217 [ run ] completed with state SUCCESS. Commit: 2adfa0a
/LLM/main/L0_MergeRequest_PR pipeline #39754 completed with status: 'SUCCESS'

CI Report

Link to invocation