From 3133a3b32fb6aa491c74b43c66169dd7c587b529 Mon Sep 17 00:00:00 2001 From: Sirius Claw Date: Wed, 22 Apr 2026 06:32:35 +0000 Subject: [PATCH] fix(vllm): update to recommended qwen3_coder tool parser, auto max-model-len, and latest image Addresses tokenizer backend / parser mismatch on heretic-style Qwen3.5 models per model card recommendations. --- apps/base/customer1/openclaw/a100-vllm.yaml | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/apps/base/customer1/openclaw/a100-vllm.yaml b/apps/base/customer1/openclaw/a100-vllm.yaml index 7fe4336..e6fbcf4 100644 --- a/apps/base/customer1/openclaw/a100-vllm.yaml +++ b/apps/base/customer1/openclaw/a100-vllm.yaml @@ -24,7 +24,7 @@ spec: effect: "NoSchedule" containers: - name: vllm-brain - image: vllm/vllm-openai:v0.19.1-ubuntu2404 + image: vllm/vllm-openai:latest command: ["python3", "-m", "vllm.entrypoints.openai.api_server"] env: - name: HF_TOKEN @@ -36,12 +36,12 @@ spec: - --tensor-parallel-size=1 - --gpu-memory-utilization=0.95 - --kv-cache-dtype=fp8 - - --max-model-len=49152 + - --max-model-len=auto - --enable-auto-tool-choice - --trust-remote-code - --dtype=auto - --enable-prefix-caching - - --tool-call-parser=hermes + - --tool-call-parser=qwen3_coder - --reasoning-parser=qwen3 - --disable-custom-all-reduce