kv-router: add block-weighted overlap counters - #34

Draft

Pernekhan wants to merge 12 commits into

mainfrom

claude/kv-router-block-weighted-metrics

Commits on Aug 25, 2026

kv_router: conditional disaggregation via DYN_MIN_REMOTE_PREFILL_TOKENS

Pernekhan
and
claude
committed
f29ed69
kv_router: fall back to local prefill on transient remote dispatch failure

Pernekhan
and
claude
committed
61a5d38
kv_router: derive session affinity from prompt prefix tokens

Pernekhan
and
claude
committed
6ecb6bc
kv_router: gate conditional disagg on uncached semantics via seen-session set

Pernekhan
and
claude
committed
494a0da
kv_router: KV-migration candidate telemetry (DYN_KV_MIGRATION_SAMPLE)

Pernekhan
and
claude
committed
86a1083
protocols: add kv_migration_source to PreprocessedRequest

Pernekhan
and
claude
committed
9e8e8c7
vllm: let a decode worker act as a KV migration source

Pernekhan
and
claude
committed
998cb05
kv_router: port overlap-driven KV migration + token gate to v1.4.1
Pernekhan
committed
b36c347
vllm: drop embedding_sequence_length passthrough dragged in by the cherry-pick
Pernekhan
committed
cc3f1d0
prefill_router: restore the stable prefix session id fix on 1.4.1
Pernekhan
committed
2819046

Commits on Aug 26, 2026

benchmarks: add the DSv4 harness used for the dynamo/engine evaluation
Pernekhan
committed
3f6f08d

Commits on Aug 27, 2026

kv-router: add block-weighted overlap counters
claude
committed
fb8d245