[Bugfix][Prefix caching] Fix inconsistent hybrid Mamba prefix cache boundaries (#60533)

Signed-off-by: Wei Zhao <weizha@aws-cmh-slurm-1-login-01.cm.cluster>
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: Wei Zhao <weizha@aws-cmh-slurm-1-login-01.cm.cluster>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Wei Zhao
2026-10-08 18:33:48 +00:00
committed by GitHub
co-authored by Wei Zhao Codex zjy0516 Claude Opus 5.5 mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 1664c4f444
commit d8fa463e2b
17 changed files with 546 additions and 200 deletions
+11
View File
@@ -230,6 +230,15 @@ class CacheConfig:
for hybrid models where requests occupy multiple KV cache groups."""
kv_cache_max_concurrency: float | None = field(default=None, init=False)
"""Per-DP-engine maximum concurrency at max_model_len tokens."""
# The two differ when hashes are finer than the scheduler block but partial
# hash hits are off (e.g. a sliding-window group cannot look up sub-block
# hits), so hits only land on the scheduler block.
# TODO(zjy0516): Merge into one value
hash_block_size: int | None = field(default=None, init=False)
"""Tokens per prefix-cache block hash."""
cache_hit_alignment_tokens: int | None = field(default=None, init=False)
"""Token granularity at which prefix-cache hits land: `hash_block_size`
with partial hash hits, otherwise the scheduler block size."""
kv_sharing_fast_prefill: bool = False
"""In some KV sharing setups, e.g. YOCO (https://arxiv.org/abs/2405.05254),
@@ -298,6 +307,8 @@ class CacheConfig:
"effective_attention_block_size",
"kv_cache_size_tokens",
"kv_cache_max_concurrency",
"hash_block_size",
"cache_hit_alignment_tokens",
# Feature toggles not impacting compiled graph shape
"kv_sharing_fast_prefill",
"swa_bounded_replay",