A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 2, 03:30 UTC
Nit Changes in Tiered KV Offload (#44293)
Signed-off-by: Rotem Shavitt <rshavitt@gmail.com>
Rotem Shavitt committed
Jun 3, 2026 at 07:53 UTC
3f0a91bb96f8d72e0498b95c166e817deae14d62
3 files changed
+8
tests/v1/kv_offload/tiering/test_fs_tier.py
renamed
tests/v1/kv_offload/tiering/test_tiering_offloading.py
renamed
vllm/v1/kv_offload/tiering/fs/manager.py
+8
@@ -52,6 +52,14 @@ class FileSystemTierManager(SecondaryTierManager):
52
submit_store / submit_load are non-blocking: they enqueue tasks and return.
53
get_finished_jobs() polls job completion and returns completed JobResults.
54
55
+ Cross-process sharing:
56
+ In order to enable KV cache sharing between multiple vLLM instances
57
+ using the same ``root_dir`` (e.g., via a shared PVC) the environment
58
+ variable ``PYTHONHASHSEED`` must be set to the same fixed value
59
+ (e.g., "0") on all instances. Without this, each process initializes
60
+ ``NONE_HASH`` (the chain-hash seed for block content hashes) with
61
+ random bytes, producing different block filenames for identical token
62
+ content.
63
"""
64
65
def __init__(