[None][feat] KV reuse probe

[lancelly]

Summary

Stacked on top of #14353. Adds the KV cache reuse probe API from #14333 by cherry-picking f9968c6bd3 onto the feat/dsv4-refactor-and-probe stack.

• Picks up the full content of [None][chore] ReuseScope refactor (cp #14140) #14353 (ReuseScope refactor + DSv4 follow-ups).
• Cherry-picks Add KV cache reuse probe ([None][feat] add KV cache reuse probe #14333, f9968c6bd3) — adds KVCacheManager.probe_reuse() plus the BlockRadixTree.match → ReuseMatch refactor that splits the prune logic out of _KVCache._setup_for_reuse.

Conflict resolution notes (tensorrt_llm/runtime/kv_cache_manager_v2/_core/_kv_cache.py)

#14333 was based on main, which doesn't have DSv4's reuse-stats tracking. The cherry-pick conflicted in two regions; resolved as follows:

1. _KVCache.__init__ — keep _pending_stats = _PendingStats() and the post-setup _refresh_generation_alloc_ready() call, while adopting the new reuse_match: ReuseMatch | None parameter from [None][feat] add KV cache reuse probe #14333.
2. _setup_for_reuse — adopt [None][feat] add KV cache reuse probe #14333's lean shape (use match.blocks / match.num_tokens from a pre-pruned ReuseMatch), then re-insert the DSv4 stats tracking inside the per-life-cycle loop (full_reused_blocks, partial_reused_blocks, _pending_stats.record_reuse, mark_stats_dirty) using num_tokens instead of the removed get_num_matched_tokens(matched) helper. Indexing was migrated from matched[ordinal][0] to matched[ordinal] since matched is now list[Block].
3. _KVCache.__init__ previously defaulted expected_prompt_length from len(input_tokens). After [None][feat] add KV cache reuse probe #14333 removes the input_tokens parameter, that fallback was moved up one level into KVCacheManager.create_kv_cache so the public-API behavior is preserved.

[lancelly]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #49639 [ run ] triggered by Bot. Commit: 12396e6 Link to invocation

[tensorrt-cicd]

PR_Github #49639 [ run ] completed with state SUCCESS. Commit: 12396e6
/LLM/main/L0_MergeRequest_PR pipeline #39258 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[lancelly]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #49704 [ run ] triggered by Bot. Commit: 12396e6 Link to invocation

[tensorrt-cicd]

PR_Github #49704 [ run ] completed with state SUCCESS. Commit: 12396e6
/LLM/main/L0_MergeRequest_PR pipeline #39310 completed with status: 'SUCCESS'

CI Report

Link to invocation

[lancelly]

/bot run --disable-fail-fast

[lancelly]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #49721 [ run ] triggered by Bot. Commit: 04a52f1 Link to invocation

[tensorrt-cicd]

PR_Github #49721 [ run ] completed with state SUCCESS. Commit: 04a52f1
/LLM/main/L0_MergeRequest_PR pipeline #39327 completed with status: 'SUCCESS'

CI Report

Link to invocation

[pcastonguay]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #49753 [ run ] triggered by Bot. Commit: fb14a8f Link to invocation

[peihu-nv]

LGTM.

One follow up needed for full E2E is wiring the probe reuse to the existing probe_prefix_match_length(...) I think.

[tensorrt-cicd]

PR_Github #49753 [ run ] completed with state SUCCESS. Commit: fb14a8f
/LLM/main/L0_MergeRequest_PR pipeline #39354 completed with status: 'SUCCESS'

CI Report

Link to invocation