优化前缀缓存机制和UI，并引入一些新工具.preview

llm_server/start_llm.sh

@@ -5,6 +5,9 @@ SCRIPT_DIR="$(cd -- "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 VENV_DIR="${SCRIPT_DIR}/myenv"
 HOST_PORT="${HOST_PORT:-8900}"
 MODEL_DIR="${SCRIPT_DIR}/models/Qwen/Qwen3___6-35B-A3B-FP8"
+# 鉴权与模型名校验：SERVER_API_KEY 为空时 vLLM 不校验 API Key（任意值放行）
+SERVER_API_KEY="${SERVER_API_KEY:-}"
+SERVER_MODELNAME="${SERVER_MODELNAME:-Qwen/Qwen3.6-35B-A3B-FP8}"
 
 # 激活虚拟环境
 source "${VENV_DIR}/bin/activate"
@@ -17,6 +20,10 @@ fi
 
 echo "Starting vLLM server on port ${HOST_PORT} ..."
 echo "Model path: ${MODEL_DIR}"
+echo "Served model name: ${SERVER_MODELNAME}"
+if [ -n "${SERVER_API_KEY}" ]; then
+  echo "API Key 校验已启用"
+fi
 
 # 设置环境变量
 export VLLM_USE_MODELSCOPE=true
@@ -28,19 +35,25 @@ export TRANSFORMERS_OFFLINE=1
 export CUDA_VISIBLE_DEVICES=3
 
 # 启动 vLLM 服务
-vllm serve "${MODEL_DIR}" \
-  --port "${HOST_PORT}" \
-  --host 0.0.0.0 \
-  --tensor-parallel-size 1 \
-  --max-model-len 262144 \
-  --reasoning-parser qwen3 \
-  --enable-auto-tool-choice \
-  --tool-call-parser qwen3_coder \
-  --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}' \
-  --enable-prefix-caching \
-  --max_num_batched_tokens 4096 \
-  --max_num_seqs 8 \
-  --kv-cache-dtype fp8_e4m3 \
-  --chat-template "${SCRIPT_DIR}/chat_template.jinja" \
-  --safetensors-load-strategy eager \
-  --served-model-name Qwen/Qwen3.6-35B-A3B-FP8
+VLLM_ARGS=(
+  --port "${HOST_PORT}"
+  --host 0.0.0.0
+  --tensor-parallel-size 1
+  --max-model-len 262144
+  --reasoning-parser qwen3
+  --enable-auto-tool-choice
+  --tool-call-parser qwen3_coder
+  --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
+  --enable-prefix-caching
+  --max_num_batched_tokens 4096
+  --max_num_seqs 8
+  --kv-cache-dtype fp8_e4m3
+  --chat-template "${SCRIPT_DIR}/chat_template.jinja"
+  --safetensors-load-strategy eager
+  --served-model-name "${SERVER_MODELNAME}"
+)
+# API Key 为空时不传 --api-key（vLLM 默认放行任意值）
+if [ -n "${SERVER_API_KEY}" ]; then
+  VLLM_ARGS+=(--api-key "${SERVER_API_KEY}")
+fi
+vllm serve "${MODEL_DIR}" "${VLLM_ARGS[@]}"