From ee3cf7c7eea75b04bb20f68bbb70b1bd3f3d5416 Mon Sep 17 00:00:00 2001 From: sirius0xdev Date: Sun, 19 Apr 2026 23:41:15 +0000 Subject: [PATCH] dial in vllm settings --- apps/base/customer1/openclaw/vllm-gemma.yaml | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/apps/base/customer1/openclaw/vllm-gemma.yaml b/apps/base/customer1/openclaw/vllm-gemma.yaml index 7ec0d25..1a0ef62 100644 --- a/apps/base/customer1/openclaw/vllm-gemma.yaml +++ b/apps/base/customer1/openclaw/vllm-gemma.yaml @@ -36,16 +36,16 @@ spec: - --host=0.0.0.0 - --port=8000 - --tensor-parallel-size=1 - - --gpu-memory-utilization=0.78 # ← Critical: lower than default 0.90 - - --max-model-len=16384 # Start conservative (you can raise to 32768 later) + - --gpu-memory-utilization=0.98 # ← Critical: lower than default 0.90 + - --max-model-len=24576 # Start conservative (you can raise to 32768 later) - --max-num-batched-tokens=8192 - - --max-num-seqs=4 # Lower for single-user / low-concurrency + - --max-num-seqs=16 # Lower for single-user / low-concurrency - --enforce-eager # Disables CUDA graphs → big memory saver at startup - --disable-custom-all-reduce + - --enable-auto-tool-choice - --trust-remote-code - --dtype=auto - --tool-call-parser qwen3_xml - - --reasoning-parser qwen3 - --enable-prefix-caching ##- --limit-mm-per-prompt='{"image"=1, "video"=0}' # Optional: limit vision if not heavily using images yet ports: