From 54267b081808a546ae4c59e89028754fc6399336 Mon Sep 17 00:00:00 2001
From: sirius0xdev <sirius0xdev@proton.me>
Date: Fri, 22 May 2026 23:27:18 -0400
Subject: [PATCH] Update rtx6000-vllm.yaml

---
 infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml b/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml
index 186d1d6..5cd1928 100644
--- a/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml
+++ b/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml
@@ -37,14 +37,14 @@ spec:
             - --port=8000
             - --tensor-parallel-size=1
             - --tokenizer-mode=hf 
-            - --gpu-memory-utilization=0.90  
-            #- --enforce-eager 
+            - --gpu-memory-utilization=0.90
             - --max-model-len=131072
             - --enable-auto-tool-choice 
             - --kv-cache-dtype=bfloat16
-            - --max-num-batched-tokens=32768 
-            - --max-num-seqs=32
+            - --max-num-batched-tokens=65536
+            - --max-num-seqs=64
             - --block-size=32 
+            - --language-model-only
             - --enable-chunked-prefill
             - --trust-remote-code
             - --dtype=auto