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 14, 05:30 UTC
[LMCache][Example] Align the PYTHONHASHSEED for prefillers and decoders for KV chunks hashing (#21161)
Signed-off-by: zejunchen-zejun <zejun.chen@amd.com>
zejunchen-zejun committed
Aug 12, 2025 at 17:05 UTC
50f2aae1b4afb8799bc6a38254639e031997e61c
1 file changed
+8
examples/others/lmcache/disagg_prefill_lmcache_v1/disagg_vllm_launcher.sh
+8
@@ -15,6 +15,14 @@ else
15
MODEL=$2
16
fi
17
18
+# The prefillers and decoders in LMCache use the same hash seed for all chunk keys.
19
+# This seed must be aligned so that decoders can identify and retrieve KV cache
20
+# entries stored by prefillers.
21
+#
22
+# WARNING: Using a fixed hash seed is insecure and makes the application vulnerable to
23
+# denial-of-service attacks. In a production environment, this should be set to a
24
+# secure random value. This is set to a fixed value for demonstration purposes only.
25
+export PYTHONHASHSEED=${VLLM_PYTHON_HASH_SEED:-123}
26
27
if [[ $1 == "prefiller" ]]; then
28
# Prefiller listens on port 8100