From 2b9630d8dcc9e0a5875b7ebdcae67d3b838bf048 Mon Sep 17 00:00:00 2001
From: sirius0xdev <sirius0xdev@proton.me>
Date: Sun, 24 May 2026 15:52:52 -0400
Subject: [PATCH] Update rtx6000-vllm.yaml

---
 infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml b/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml
index 6135e98..d9ff93a 100644
--- a/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml
+++ b/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml
@@ -38,11 +38,11 @@ spec:
             - --tensor-parallel-size=1
             - --tokenizer-mode=hf 
             - --gpu-memory-utilization=0.90
-            - --max-model-len=131072
+            - --max-model-len=262144
             - --enable-auto-tool-choice 
             - --kv-cache-dtype=bfloat16
             - --max-num-batched-tokens=16384
-            - --max-num-seqs=64
+            - --max-num-seqs=4
             - --block-size=32 
             - --language-model-only
             - --enable-chunked-prefill