From e4fba35e890635e1c56b77bdb51353972de29fa3 Mon Sep 17 00:00:00 2001
From: Frank <frank@dankware>
Date: Fri, 7 Aug 2026 08:49:59 -0700
Subject: [PATCH] Baseline: ThinkingCap 27B FP8 on dankclaw

- vLLM 0.26.0+rocm723
- --kv-cache-dtype fp8 --gpu-memory-utilization 0.80
- --max-model-len 32768 --max-num-batched-tokens 2048
- --max-num-seqs 32 --reasoning-parser qwen3 --language-model-only
- Baseline response time: 219s for 50 tokens
---
 scripts/serve-thinkingcap.sh | 29 +++++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)
 create mode 100644 scripts/serve-thinkingcap.sh

diff --git a/scripts/serve-thinkingcap.sh b/scripts/serve-thinkingcap.sh
new file mode 100644
index 0000000..a5f7d95
--- /dev/null
+++ b/scripts/serve-thinkingcap.sh
@@ -0,0 +1,29 @@
+#!/bin/bash
+# vLLM serve script for ThinkingCap-Qwen3.6-27B-FP8 on dankclaw (ROCm)
+# Run via systemd service: vllm-thinkingcap.service
+
+set -euo pipefail
+
+# Environment
+export HSA_OVERRIDE_GFX_VERSION=11.5.1
+export HF_HOME="/home/dank/.cache/huggingface"
+
+# vLLM virtual environment — use direct paths, no activate script
+VENV="/home/dank/vllm-env"
+export PATH="${VENV}/bin:${PATH}"
+
+# Model
+MODEL="bottlecapai/ThinkingCap-Qwen3.6-27B-FP8"
+
+# Serve
+exec "${VENV}/bin/vllm" serve "${MODEL}" \
+    --port 8000 \
+    --host 0.0.0.0 \
+    --kv-cache-dtype fp8 \
+    --gpu-memory-utilization 0.80 \
+    --max-model-len 32768 \
+    --max-num-batched-tokens 2048 \
+    --max-num-seqs 32 \
+    --reasoning-parser qwen3 \
+    --language-model-only \
+    --dtype auto