feat: v1.2.0 - 三项优化

1. 模型选择改为两步式：先搜索选择基模型，再选择量化版本
   - 自动选中默认量化版本(Q4_K_M)，后台可配置
   - 模型新增 base_model 字段，按基模型分组
   - 后台模型管理增加基模型字段

2. 参数搜索改为全局搜索
   - 搜索时跨所有分类显示匹配参数
   - 隐藏分类标签页，显示匹配数量

3. 自然语言支持 LLM 接口
   - 后台可配置 LLM API (URL/Key/Model/System Prompt)
   - 启用后优先调用 LLM 解析，失败自动回退正则解析
   - 兼容 OpenAI API 格式

README.md

@@ -0,0 +1,106 @@
+# 🦙 llama.cpp 命令生成器
+
+> 一个用于生成 llama.cpp 执行命令的 Web 工具，支持参数可视化配置、自然语言生成、实时显存估算。
+
+## ✨ 功能特性
+
+### 命令生成
+- **多版本支持**：内置 b6310 和 b10068 两个版本的参数定义，可按版本精准匹配参数
+- **GPU 模式 / GPU+CPU 模式**：默认 GPU 模式，可切换混合模式查看系统内存占用
+- **多 GPU**：最多支持 4 张显卡，每张可选不同型号，自动计算总显存
+- **智能输出**：仅输出非默认值的参数，生成干净可用的命令
+
+### 自然语言生成
+输入自然语言描述，自动解析为参数配置：
+
+```
+用RTX 4090跑70B模型, 上下文8192, 温度0.7, flash attention
+```
+
+支持中文和英文关键词，涵盖 GPU 型号、数量、上下文、温度、Top-K/P、线程数、端口等。
+
+### 实时显存估算
+- 选择模型和 GPU 后实时计算显存占用
+- 分项显示：模型权重 / KV缓存 / 计算开销 / CUDA开销
+- 进度条颜色预警：绿(正常) → 黄(警告) → 红(超限)
+- GPU+CPU 模式额外显示 CPU 侧权重和 KV 缓存占用
+
+### 参数管理
+- 参数按分类标签页组织：通用 / 采样 / 服务器 / 模型来源 / LoRA / 日志 / 高级
+- 重要参数直接显示，次要参数点击展开
+- 影响显存的参数标记 ⚡ 标识
+- 支持参数搜索（按键名、标志、描述）
+
+### 模型选择
+- 内置 26 个常见模型预设（Llama 3/3.1、Qwen2.5、DeepSeek V2/R1、Mistral、Gemma2、Phi-3、GLM-4 等）
+- 支持搜索过滤
+- 后台可增删改
+
+### 后台管理
+- **GPU 管理**：增删改查显卡型号、显存、计算能力
+- **版本管理**：增删改查 llama.cpp 版本
+- **参数管理**：按版本增删改查参数定义
+- **模型管理**：增删改查模型预设
+- **系统设置**：修改管理密码等配置
+
+## 🚀 部署
+
+### 环境要求
+- Python 3.8+
+- Flask
+- flask-cors
+
+### 安装
+```bash
+cd llama-cmd-gen
+pip install flask flask-cors
+```
+
+### 启动
+```bash
+python3 app.py
+# 或
+./start.sh
+```
+
+服务默认监听 `0.0.0.0:16052`。
+
+### 访问
+- **主界面**：`http://<host>:16052/`
+- **后台管理**：`http://<host>:16052/admin`（需密码登录，默认 `admin123`）
+
+## 📁 项目结构
+
+```
+llama-cmd-gen/
+├── app.py              # Flask 主应用 + API
+├── db.py               # 数据库初始化 + 默认数据
+├── data.db             # SQLite 数据库（运行时生成）
+├── start.sh            # 启动脚本
+├── requirements.txt    # Python 依赖
+├── logs/
+│   └── app.log         # 运行日志
+└── static/
+    ├── index.html       # 主界面
+    ├── admin.html       # 后台管理界面
+    ├── css/
+    │   └── style.css    # 样式
+    └── js/
+        ├── main.js      # 主界面逻辑
+        └── admin.js     # 后台逻辑
+```
+
+## 🛠️ 技术栈
+
+- **后端**：Python / Flask / SQLite
+- **前端**：原生 HTML / CSS / JavaScript（无框架依赖）
+- **数据**：SQLite 本地存储，包含 GPU、版本、参数、模型等表
+
+## 📝 版本历史
+
+- **v1.1.0** — 修复输入焦点丢失、加参数搜索、模型选择器、后台密码登录、复制兼容性
+- **v1.0.0** — 初始版本：命令生成、自然语言解析、显存估算、多 GPU、后台管理
+
+## 📄 License
+
+MIT

app.py

@@ -7,6 +7,7 @@ import json
 import re
 import math
 import functools
+import urllib.request
 from flask import Flask, request, jsonify, send_from_directory, session, redirect
 from flask_cors import CORS
 from db import get_db, init_db, DB_PATH
@@ -600,16 +601,86 @@ def get_models():
     db.close()
     return jsonify(result)
 
+@app.route('/api/models/grouped')
+def get_models_grouped():
+    db = get_db()
+    models = db.execute('SELECT * FROM models ORDER BY base_model, sort_order, name').fetchall()
+    db.close()
+    grouped = {}
+    for m in models:
+        d = dict(m)
+        base = d['base_model']
+        if base not in grouped:
+            grouped[base] = []
+        grouped[base].append(d)
+    # Get default quant from settings
+    db2 = get_db()
+    dq = db2.execute("SELECT value FROM settings WHERE key = 'default_quant'").fetchone()
+    db2.close()
+    default_quant = dq['value'] if dq else 'Q4_K_M'
+    return jsonify({'models': grouped, 'default_quant': default_quant})
+
 
 # ----- Parse Natural Language -----
 @app.route('/api/parse-nl', methods=['POST'])
 def parse_nl():
     data = request.json
     text = data.get('text', '')
+    # Try LLM API first if enabled
+    db = get_db()
+    llm_enabled = db.execute("SELECT value FROM settings WHERE key = 'llm_enabled'").fetchone()
+    if llm_enabled and llm_enabled['value'] == 'true':
+        llm_url = db.execute("SELECT value FROM settings WHERE key = 'llm_api_url'").fetchone()
+        llm_key = db.execute("SELECT value FROM settings WHERE key = 'llm_api_key'").fetchone()
+        llm_model = db.execute("SELECT value FROM settings WHERE key = 'llm_api_model'").fetchone()
+        llm_prompt = db.execute("SELECT value FROM settings WHERE key = 'llm_system_prompt'").fetchone()
+        db.close()
+        url = llm_url['value'] if llm_url else ''
+        key = llm_key['value'] if llm_key else ''
+        model = llm_model['value'] if llm_model else ''
+        system_prompt = llm_prompt['value'] if llm_prompt else ''
+        if url:
+            try:
+                result = call_llm_for_parsing(url, key, model, system_prompt, text)
+                if result:
+                    return jsonify(result)
+            except Exception as e:
+                print(f'LLM parse failed: {e}', file=sys.stderr)
+    else:
+        db.close()
+    # Fallback to regex parsing
     result = parse_natural_language(text)
     return jsonify(result)
 
 
+def call_llm_for_parsing(url, key, model, system_prompt, user_text):
+    """Call LLM API to parse natural language into params."""
+    headers = {'Content-Type': 'application/json'}
+    if key:
+        headers['Authorization'] = f'Bearer {key}'
+    body = {
+        'model': model,
+        'messages': [
+            {'role': 'system', 'content': system_prompt},
+            {'role': 'user', 'content': user_text}
+        ],
+        'temperature': 0.1,
+        'max_tokens': 2000,
+    }
+    req = urllib.request.Request(url, data=json.dumps(body).encode('utf-8'), headers=headers, method='POST')
+    with urllib.request.urlopen(req, timeout=30) as resp:
+        data = json.loads(resp.read().decode('utf-8'))
+    # OpenAI-compatible response
+    content = data['choices'][0]['message']['content']
+    # Try to extract JSON from the response
+    content = content.strip()
+    if content.startswith('```'):
+        content = re.sub(r'^```\w*\n?', '', content)
+        content = re.sub(r'\n?```$', '', content)
+    result = json.loads(content)
+    return result
+
+
 # ==================== Admin API ====================
 
     # All admin routes below require login
@@ -800,9 +871,9 @@ def admin_add_model():
     data = request.json
     db = get_db()
     db.execute(
-        '''INSERT INTO models (name, size_gb, layers, embd, kv_heads, head_dim, attention_heads, quant, description, sort_order)
-        VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)''',
-        (data['name'], data['size_gb'], data['layers'], data['embd'],
+        '''INSERT INTO models (base_model, name, size_gb, layers, embd, kv_heads, head_dim, attention_heads, quant, description, sort_order)
+        VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)''',
+        (data['base_model'], data['name'], data['size_gb'], data['layers'], data['embd'],
          data['kv_heads'], data['head_dim'], data['attention_heads'],
          data.get('quant', ''), data.get('description', ''), data.get('sort_order', 0))
     )
@@ -818,9 +889,9 @@ def admin_model_edit(mid):
     if request.method == 'PUT':
         data = request.json
         db.execute(
-            '''UPDATE models SET name=?, size_gb=?, layers=?, embd=?, kv_heads=?,
+            '''UPDATE models SET base_model=?, name=?, size_gb=?, layers=?, embd=?, kv_heads=?,
             head_dim=?, attention_heads=?, quant=?, description=?, sort_order=? WHERE id=?''',
-            (data['name'], data['size_gb'], data['layers'], data['embd'],
+            (data['base_model'], data['name'], data['size_gb'], data['layers'], data['embd'],
              data['kv_heads'], data['head_dim'], data['attention_heads'],
              data.get('quant', ''), data.get('description', ''), data.get('sort_order', 0), mid)
         )

db.py

@@ -80,6 +80,7 @@ def init_db():
     c.execute('''
         CREATE TABLE IF NOT EXISTS models (
             id INTEGER PRIMARY KEY AUTOINCREMENT,
+            base_model TEXT NOT NULL,
             name TEXT NOT NULL,
             size_gb REAL NOT NULL,
             layers INTEGER NOT NULL,
@@ -312,44 +313,79 @@ def insert_default_data(conn):
             (v2_id,) + p)
 
     # ===== Default models =====
+    # (base_model, name, size_gb, layers, embd, kv_heads, head_dim, attention_heads, quant, description, sort_order)
     default_models = [
-        ("Llama-3-8B-Instruct (Q4_K_M)", 4.9, 32, 4096, 8, 128, 32, "Q4_K_M", "Meta Llama 3 8B Instruct, Q4_K_M 量化", 1),
-        ("Llama-3-8B-Instruct (Q8_0)", 8.5, 32, 4096, 8, 128, 32, "Q8_0", "Meta Llama 3 8B Instruct, Q8_0 量化", 2),
-        ("Llama-3-8B-Instruct (FP16)", 15.5, 32, 4096, 8, 128, 32, "FP16", "Meta Llama 3 8B Instruct, FP16", 3),
-        ("Llama-3-70B-Instruct (Q4_K_M)", 38.5, 80, 8192, 8, 128, 64, "Q4_K_M", "Meta Llama 3 70B Instruct, Q4_K_M 量化", 4),
-        ("Llama-3-70B-Instruct (Q8_0)", 74.0, 80, 8192, 8, 128, 64, "Q8_0", "Meta Llama 3 70B Instruct, Q8_0 量化", 5),
-        ("Llama-3-70B-Instruct (FP16)", 138.0, 80, 8192, 8, 128, 64, "FP16", "Meta Llama 3 70B Instruct, FP16", 6),
-        ("Llama-3.1-8B-Instruct (Q4_K_M)", 4.9, 32, 4096, 8, 128, 32, "Q4_K_M", "Meta Llama 3.1 8B Instruct, Q4_K_M 量化", 7),
-        ("Llama-3.1-70B-Instruct (Q4_K_M)", 38.5, 80, 8192, 8, 128, 64, "Q4_K_M", "Meta Llama 3.1 70B Instruct, Q4_K_M 量化", 8),
-        ("Llama-3.1-70B-Instruct (Q8_0)", 74.0, 80, 8192, 8, 128, 64, "Q8_0", "Meta Llama 3.1 70B Instruct, Q8_0 量化", 9),
-        ("Llama-3.1-405B-Instruct (Q4_K_M)", 226.0, 126, 16384, 8, 128, 128, "Q4_K_M", "Meta Llama 3.1 405B Instruct, Q4_K_M 量化", 10),
-        ("Qwen2.5-7B-Instruct (Q4_K_M)", 4.7, 28, 3584, 4, 128, 28, "Q4_K_M", "Qwen2.5 7B Instruct, Q4_K_M 量化", 11),
-        ("Qwen2.5-14B-Instruct (Q4_K_M)", 8.7, 40, 5120, 8, 128, 40, "Q4_K_M", "Qwen2.5 14B Instruct, Q4_K_M 量化", 12),
-        ("Qwen2.5-32B-Instruct (Q4_K_M)", 19.5, 64, 5120, 8, 128, 64, "Q4_K_M", "Qwen2.5 32B Instruct, Q4_K_M 量化", 13),
-        ("Qwen2.5-72B-Instruct (Q4_K_M)", 42.0, 80, 8192, 8, 128, 64, "Q4_K_M", "Qwen2.5 72B Instruct, Q4_K_M 量化", 14),
-        ("Qwen2.5-72B-Instruct (Q8_0)", 75.0, 80, 8192, 8, 128, 64, "Q8_0", "Qwen2.5 72B Instruct, Q8_0 量化", 15),
-        ("DeepSeek-V2-Chat (Q4_K_M)", 23.0, 60, 5120, 8, 128, 60, "Q4_K_M", "DeepSeek V2 Chat, Q4_K_M 量化", 16),
-        ("DeepSeek-V2.5-Chat (Q4_K_M)", 23.0, 60, 5120, 8, 128, 60, "Q4_K_M", "DeepSeek V2.5 Chat, Q4_K_M 量化", 17),
-        ("DeepSeek-R1-Distill-Qwen-32B (Q4_K_M)", 19.5, 64, 5120, 8, 128, 64, "Q4_K_M", "DeepSeek R1 Distill Qwen 32B, Q4_K_M 量化", 18),
-        ("DeepSeek-R1-Distill-Llama-70B (Q4_K_M)", 42.0, 80, 8192, 8, 128, 64, "Q4_K_M", "DeepSeek R1 Distill Llama 70B, Q4_K_M 量化", 19),
-        ("Mistral-7B-Instruct-v0.3 (Q4_K_M)", 4.4, 32, 4096, 8, 128, 32, "Q4_K_M", "Mistral 7B Instruct v0.3, Q4_K_M 量化", 20),
-        ("Mixtral-8x7B-Instruct (Q4_K_M)", 26.0, 32, 4096, 8, 128, 32, "Q4_K_M", "Mixtral 8x7B Instruct, Q4_K_M 量化", 21),
-        ("Gemma-2-9B-It (Q4_K_M)", 5.4, 42, 3584, 4, 256, 14, "Q4_K_M", "Google Gemma 2 9B It, Q4_K_M 量化", 22),
-        ("Gemma-2-27B-It (Q4_K_M)", 16.5, 46, 4608, 4, 128, 36, "Q4_K_M", "Google Gemma 2 27B It, Q4_K_M 量化", 23),
-        ("Phi-3-Mini-4K-Instruct (Q4_K_M)", 2.5, 32, 3072, 32, 96, 32, "Q4_K_M", "Microsoft Phi-3 Mini 4K Instruct, Q4_K_M 量化", 24),
-        ("Phi-3-Medium-14B-Instruct (Q4_K_M)", 8.4, 40, 5120, 10, 128, 40, "Q4_K_M", "Microsoft Phi-3 Medium 14B Instruct, Q4_K_M 量化", 25),
-        ("GLM-4-9B-Chat (Q4_K_M)", 5.5, 40, 4096, 4, 128, 40, "Q4_K_M", "Zhipu GLM-4 9B Chat, Q4_K_M 量化", 26),
+        # Llama-3-8B-Instruct
+        ("Llama-3-8B-Instruct", "Llama-3-8B-Instruct (Q4_K_M)", 4.9, 32, 4096, 8, 128, 32, "Q4_K_M", "Meta Llama 3 8B Instruct, Q4_K_M 量化", 1),
+        ("Llama-3-8B-Instruct", "Llama-3-8B-Instruct (Q8_0)", 8.5, 32, 4096, 8, 128, 32, "Q8_0", "Meta Llama 3 8B Instruct, Q8_0 量化", 2),
+        ("Llama-3-8B-Instruct", "Llama-3-8B-Instruct (FP16)", 15.5, 32, 4096, 8, 128, 32, "FP16", "Meta Llama 3 8B Instruct, FP16", 3),
+        # Llama-3-70B-Instruct
+        ("Llama-3-70B-Instruct", "Llama-3-70B-Instruct (Q4_K_M)", 38.5, 80, 8192, 8, 128, 64, "Q4_K_M", "Meta Llama 3 70B Instruct, Q4_K_M 量化", 4),
+        ("Llama-3-70B-Instruct", "Llama-3-70B-Instruct (Q8_0)", 74.0, 80, 8192, 8, 128, 64, "Q8_0", "Meta Llama 3 70B Instruct, Q8_0 量化", 5),
+        ("Llama-3-70B-Instruct", "Llama-3-70B-Instruct (FP16)", 138.0, 80, 8192, 8, 128, 64, "FP16", "Meta Llama 3 70B Instruct, FP16", 6),
+        # Llama-3.1-8B-Instruct
+        ("Llama-3.1-8B-Instruct", "Llama-3.1-8B-Instruct (Q4_K_M)", 4.9, 32, 4096, 8, 128, 32, "Q4_K_M", "Meta Llama 3.1 8B Instruct, Q4_K_M 量化", 7),
+        ("Llama-3.1-8B-Instruct", "Llama-3.1-8B-Instruct (Q8_0)", 8.5, 32, 4096, 8, 128, 32, "Q8_0", "Meta Llama 3.1 8B Instruct, Q8_0 量化", 8),
+        # Llama-3.1-70B-Instruct
+        ("Llama-3.1-70B-Instruct", "Llama-3.1-70B-Instruct (Q4_K_M)", 38.5, 80, 8192, 8, 128, 64, "Q4_K_M", "Meta Llama 3.1 70B Instruct, Q4_K_M 量化", 9),
+        ("Llama-3.1-70B-Instruct", "Llama-3.1-70B-Instruct (Q8_0)", 74.0, 80, 8192, 8, 128, 64, "Q8_0", "Meta Llama 3.1 70B Instruct, Q8_0 量化", 10),
+        # Llama-3.1-405B-Instruct
+        ("Llama-3.1-405B-Instruct", "Llama-3.1-405B-Instruct (Q4_K_M)", 226.0, 126, 16384, 8, 128, 128, "Q4_K_M", "Meta Llama 3.1 405B Instruct, Q4_K_M 量化", 11),
+        # Qwen2.5-7B-Instruct
+        ("Qwen2.5-7B-Instruct", "Qwen2.5-7B-Instruct (Q4_K_M)", 4.7, 28, 3584, 4, 128, 28, "Q4_K_M", "Qwen2.5 7B Instruct, Q4_K_M 量化", 12),
+        ("Qwen2.5-7B-Instruct", "Qwen2.5-7B-Instruct (Q8_0)", 7.6, 28, 3584, 4, 128, 28, "Q8_0", "Qwen2.5 7B Instruct, Q8_0 量化", 13),
+        # Qwen2.5-14B-Instruct
+        ("Qwen2.5-14B-Instruct", "Qwen2.5-14B-Instruct (Q4_K_M)", 8.7, 40, 5120, 8, 128, 40, "Q4_K_M", "Qwen2.5 14B Instruct, Q4_K_M 量化", 14),
+        # Qwen2.5-32B-Instruct
+        ("Qwen2.5-32B-Instruct", "Qwen2.5-32B-Instruct (Q4_K_M)", 19.5, 64, 5120, 8, 128, 64, "Q4_K_M", "Qwen2.5 32B Instruct, Q4_K_M 量化", 15),
+        ("Qwen2.5-32B-Instruct", "Qwen2.5-32B-Instruct (Q8_0)", 32.0, 64, 5120, 8, 128, 64, "Q8_0", "Qwen2.5 32B Instruct, Q8_0 量化", 16),
+        # Qwen2.5-72B-Instruct
+        ("Qwen2.5-72B-Instruct", "Qwen2.5-72B-Instruct (Q4_K_M)", 42.0, 80, 8192, 8, 128, 64, "Q4_K_M", "Qwen2.5 72B Instruct, Q4_K_M 量化", 17),
+        ("Qwen2.5-72B-Instruct", "Qwen2.5-72B-Instruct (Q8_0)", 75.0, 80, 8192, 8, 128, 64, "Q8_0", "Qwen2.5 72B Instruct, Q8_0 量化", 18),
+        # DeepSeek-V2-Chat
+        ("DeepSeek-V2-Chat", "DeepSeek-V2-Chat (Q4_K_M)", 23.0, 60, 5120, 8, 128, 60, "Q4_K_M", "DeepSeek V2 Chat, Q4_K_M 量化", 19),
+        # DeepSeek-V2.5-Chat
+        ("DeepSeek-V2.5-Chat", "DeepSeek-V2.5-Chat (Q4_K_M)", 23.0, 60, 5120, 8, 128, 60, "Q4_K_M", "DeepSeek V2.5 Chat, Q4_K_M 量化", 20),
+        # DeepSeek-R1-Distill-Qwen-32B
+        ("DeepSeek-R1-Distill-Qwen-32B", "DeepSeek-R1-Distill-Qwen-32B (Q4_K_M)", 19.5, 64, 5120, 8, 128, 64, "Q4_K_M", "DeepSeek R1 Distill Qwen 32B, Q4_K_M 量化", 21),
+        ("DeepSeek-R1-Distill-Qwen-32B", "DeepSeek-R1-Distill-Qwen-32B (Q8_0)", 32.0, 64, 5120, 8, 128, 64, "Q8_0", "DeepSeek R1 Distill Qwen 32B, Q8_0 量化", 22),
+        # DeepSeek-R1-Distill-Llama-70B
+        ("DeepSeek-R1-Distill-Llama-70B", "DeepSeek-R1-Distill-Llama-70B (Q4_K_M)", 42.0, 80, 8192, 8, 128, 64, "Q4_K_M", "DeepSeek R1 Distill Llama 70B, Q4_K_M 量化", 23),
+        ("DeepSeek-R1-Distill-Llama-70B", "DeepSeek-R1-Distill-Llama-70B (Q8_0)", 75.0, 80, 8192, 8, 128, 64, "Q8_0", "DeepSeek R1 Distill Llama 70B, Q8_0 量化", 24),
+        # Mistral-7B-Instruct-v0.3
+        ("Mistral-7B-Instruct-v0.3", "Mistral-7B-Instruct-v0.3 (Q4_K_M)", 4.4, 32, 4096, 8, 128, 32, "Q4_K_M", "Mistral 7B Instruct v0.3, Q4_K_M 量化", 25),
+        ("Mistral-7B-Instruct-v0.3", "Mistral-7B-Instruct-v0.3 (Q8_0)", 7.5, 32, 4096, 8, 128, 32, "Q8_0", "Mistral 7B Instruct v0.3, Q8_0 量化", 26),
+        # Mixtral-8x7B-Instruct
+        ("Mixtral-8x7B-Instruct", "Mixtral-8x7B-Instruct (Q4_K_M)", 26.0, 32, 4096, 8, 128, 32, "Q4_K_M", "Mixtral 8x7B Instruct, Q4_K_M 量化", 27),
+        # Gemma-2-9B-It
+        ("Gemma-2-9B-It", "Gemma-2-9B-It (Q4_K_M)", 5.4, 42, 3584, 4, 256, 14, "Q4_K_M", "Google Gemma 2 9B It, Q4_K_M 量化", 28),
+        # Gemma-2-27B-It
+        ("Gemma-2-27B-It", "Gemma-2-27B-It (Q4_K_M)", 16.5, 46, 4608, 4, 128, 36, "Q4_K_M", "Google Gemma 2 27B It, Q4_K_M 量化", 29),
+        # Phi-3-Mini-4K-Instruct
+        ("Phi-3-Mini-4K-Instruct", "Phi-3-Mini-4K-Instruct (Q4_K_M)", 2.5, 32, 3072, 32, 96, 32, "Q4_K_M", "Microsoft Phi-3 Mini 4K Instruct, Q4_K_M 量化", 30),
+        # Phi-3-Medium-14B-Instruct
+        ("Phi-3-Medium-14B-Instruct", "Phi-3-Medium-14B-Instruct (Q4_K_M)", 8.4, 40, 5120, 10, 128, 40, "Q4_K_M", "Microsoft Phi-3 Medium 14B Instruct, Q4_K_M 量化", 31),
+        # GLM-4-9B-Chat
+        ("GLM-4-9B-Chat", "GLM-4-9B-Chat (Q4_K_M)", 5.5, 40, 4096, 4, 128, 40, "Q4_K_M", "Zhipu GLM-4 9B Chat, Q4_K_M 量化", 32),
+        ("GLM-4-9B-Chat", "GLM-4-9B-Chat (Q8_0)", 9.0, 40, 4096, 4, 128, 40, "Q8_0", "Zhipu GLM-4 9B Chat, Q8_0 量化", 33),
     ]
 
     for m in default_models:
-        c.execute('''INSERT INTO models (name, size_gb, layers, embd, kv_heads, head_dim, attention_heads, quant, description, sort_order)
-                     VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)''', m)
+        c.execute('''INSERT INTO models (base_model, name, size_gb, layers, embd, kv_heads, head_dim, attention_heads, quant, description, sort_order)
+                     VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)''', m)
 
     # ===== Default settings =====
     c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("admin_password", "admin123"))
     c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("default_gpu", "RTX 3090"))
     c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("default_version", "b10068"))
     c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("default_mode", "gpu"))
+    c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("default_quant", "Q4_K_M"))
+    # LLM API settings for natural language parsing
+    c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("llm_enabled", "false"))
+    c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("llm_api_url", ""))
+    c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("llm_api_key", ""))
+    c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("llm_api_model", ""))
+    c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("llm_system_prompt", "你是一个llama.cpp命令行参数解析助手。用户会用自然语言描述他想运行的模型和参数配置，你需要将其解析为JSON格式的参数。\n\n可用的参数键包括: model, ctx_size, n_gpu_layers, threads, batch_size, temperature, top_k, top_p, flash_attn, port, host, parallel, split_mode, mlock, numa, repeat_penalty, presence_penalty, frequency_penalty, seed, min_p, typical, mirostat, mirostat_lr, mirostat_ent。\n\nGPU型号会通过 _gpu_name 字段返回，GPU数量通过 _gpu_count 返回。\n\n只返回JSON，不要其他文本。"))
 
     conn.commit()
 

static/admin.html

@@ -123,7 +123,8 @@
             <div class="admin-add-form">
                 <h3>添加新模型</h3>
                 <div class="form-grid">
-                    <input type="text" id="model-name" placeholder="模型名称">
+                    <input type="text" id="model-basemodel" placeholder="基模型 (如: Llama-3-8B-Instruct)">
+                    <input type="text" id="model-name" placeholder="完整名称 (含量化)">
                     <input type="number" id="model-size" placeholder="大小(GB)" step="0.1">
                     <input type="number" id="model-layers" placeholder="层数">
                     <input type="number" id="model-embd" placeholder="嵌入维度">
@@ -137,7 +138,7 @@
                 </div>
             </div>
             <table class="admin-table">
-                <thead><tr><th>ID</th><th>名称</th><th>大小(GB)</th><th>层数</th><th>EMBD</th><th>KV</th><th>HD</th><th>Heads</th><th>量化</th><th>排序</th><th>操作</th></tr></thead>
+                <thead><tr><th>ID</th><th>基模型</th><th>名称</th><th>大小(GB)</th><th>层数</th><th>EMBD</th><th>KV</th><th>HD</th><th>Heads</th><th>量化</th><th>排序</th><th>操作</th></tr></thead>
                 <tbody id="model-table-body"></tbody>
             </table>
         </section>

static/css/style.css

@@ -177,7 +177,10 @@ header h1 {
 .vram-breakdown { margin-top: 10px; font-size: 0.85em; color: var(--text-dim); display: grid; grid-template-columns: repeat(auto-fill, minmax(200px, 1fr)); gap: 6px; }
 .vram-breakdown .breakdown-item { display: flex; justify-content: space-between; padding: 4px 8px; background: var(--bg); border-radius: 4px; }
 
-/* ===== Model Selection ===== */
+/* ===== Model Selection (two-step) ===== */
+.model-step { margin-bottom: 12px; }
+.model-step.hidden { display: none; }
+.model-step-label { display: block; font-size: 0.85em; color: var(--text-dim); margin-bottom: 6px; }
 .model-select-container { position: relative; }
 .model-select-container input {
     width: 100%;
@@ -215,8 +218,18 @@ header h1 {
 .model-name { font-size: 0.9em; }
 .model-meta { font-size: 0.8em; color: var(--text-dim); }
 
+.quant-btn {
+    display: inline-flex; flex-direction: column; align-items: center; gap: 2px;
+    padding: 8px 16px; background: var(--bg-input); border: 1px solid var(--border);
+    color: var(--text); border-radius: var(--radius); cursor: pointer; font-size: 0.85em; margin-right: 8px; margin-bottom: 8px; transition: all 0.15s;
+}
+.quant-btn:hover { border-color: var(--accent2); }
+.quant-btn.active { background: var(--accent); color: white; border-color: var(--accent); }
+.quant-size { font-size: 0.75em; color: var(--text-dim); }
+.quant-btn.active .quant-size { color: rgba(255,255,255,0.8); }
+
 .model-selected-info { margin-top: 12px; }
-.model-detail { display: flex; gap: 16px; flex-wrap: wrap; font-size: 0.85em; }
+.model-detail { display: flex; gap: 12px; flex-wrap: wrap; font-size: 0.85em; }
 .detail-item { background: var(--bg-input); padding: 4px 10px; border-radius: 4px; }
 .detail-item b { color: var(--accent2); }
 
@@ -235,7 +248,7 @@ header h1 {
 .tab-btn.active { background: var(--accent); color: white; border-color: var(--accent); }
 
 /* ===== Param Search ===== */
-.param-search-container { margin-bottom: 12px; }
+.param-search-container { margin-bottom: 12px; position: relative; }
 .param-search-container input {
     width: 100%;
     background: var(--bg-input);
@@ -246,6 +259,14 @@ header h1 {
     font-size: 0.9em;
 }
 .param-search-container input:focus { outline: none; border-color: var(--accent); }
+.search-hint { font-size: 0.8em; color: var(--text-dim); margin-left: 8px; }
+
+/* ===== Settings (wide textarea) ===== */
+.settings-form .setting-item-wide { grid-column: 1 / -1; }
+.settings-form textarea {
+    background: var(--bg-input); border: 1px solid var(--border); color: var(--text);
+    padding: 8px 12px; border-radius: var(--radius); font-size: 0.85em; resize: vertical; width: 100%; font-family: inherit;
+}
 
 /* ===== Parameter Items (one per line) ===== */
 #param-container { display: flex; flex-direction: column; gap: 6px; }

static/index.html

@@ -39,8 +39,6 @@
         <h2>🖥️ GPU 配置</h2>
         <div id="gpu-slots"></div>
         <button class="btn-add" id="btn-add-gpu" onclick="addGpuSlot()">+ 添加显卡</button>
-
-        <!-- VRAM Display -->
         <div class="vram-display" id="vram-display">
             <div class="vram-bar-container">
                 <div class="vram-bar" id="vram-bar"></div>
@@ -71,13 +69,23 @@
         </div>
     </section>
 
-    <!-- Model Selection -->
+    <!-- Model Selection (two-step) -->
     <section class="panel">
         <h2>📦 选择模型</h2>
-        <div class="model-select-container">
-            <input type="text" id="model-search" placeholder="搜索模型名称..." oninput="filterModels()" autocomplete="off">
-            <div class="model-dropdown" id="model-dropdown"></div>
+        <!-- Step 1: Select base model -->
+        <div class="model-step">
+            <label class="model-step-label">第1步：选择模型</label>
+            <div class="model-select-container">
+                <input type="text" id="model-search" placeholder="搜索模型名称..." oninput="filterBaseModels()" autocomplete="off" onfocus="this.select()">
+                <div class="model-dropdown" id="model-dropdown"></div>
+            </div>
         </div>
+        <!-- Step 2: Select quant -->
+        <div class="model-step hidden" id="quant-step">
+            <label class="model-step-label">第2步：选择量化版本</label>
+            <div id="quant-options"></div>
+        </div>
+        <!-- Selected model info -->
         <div id="model-selected-info" class="model-selected-info"></div>
     </section>
 
@@ -94,7 +102,11 @@
     <!-- Parameters -->
     <section class="panel">
         <h2>⚙️ 参数配置</h2>
-        <div class="param-tabs">
+        <div class="param-search-container">
+            <input type="text" id="param-search" placeholder="搜索全部参数..." oninput="onParamSearch()" autocomplete="off">
+            <span id="search-hint" class="search-hint"></span>
+        </div>
+        <div class="param-tabs" id="param-tabs">
             <button class="tab-btn active" data-cat="common" onclick="switchTab('common')">通用参数</button>
             <button class="tab-btn" data-cat="sampling" onclick="switchTab('sampling')">采样参数</button>
             <button class="tab-btn" data-cat="server" onclick="switchTab('server')">服务器参数</button>
@@ -103,9 +115,6 @@
             <button class="tab-btn" data-cat="logging" onclick="switchTab('logging')">日志</button>
             <button class="tab-btn" data-cat="advanced" onclick="switchTab('advanced')">高级</button>
         </div>
-        <div class="param-search-container">
-            <input type="text" id="param-search" placeholder="搜索参数名称或标志..." oninput="onParamSearch()" autocomplete="off">
-        </div>
         <div id="param-container"></div>
         <button class="btn-toggle-advanced" onclick="toggleHiddenParams()" id="toggle-advanced-btn">▼ 显示更多参数</button>
     </section>
@@ -120,7 +129,6 @@
         </div>
     </section>
 </div>
-
 <script src="/js/main.js"></script>
 </body>
 </html>

static/js/admin.js

@@ -239,7 +239,8 @@ function renderModelTable() {
     document.getElementById('model-table-body').innerHTML = adminState.models.map(m => `
         <tr>
             <td>${m.id}</td>
-            <td><input type="text" value="${m.name}" onchange="updateModel(${m.id}, 'name', this.value)"></td>
+            <td><input type="text" value="${m.base_model}" onchange="updateModel(${m.id}, 'base_model', this.value)" style="width:140px"></td>
+            <td><input type="text" value="${m.name}" onchange="updateModel(${m.id}, 'name', this.value)" style="width:200px"></td>
             <td><input type="number" value="${m.size_gb}" step="0.1" onchange="updateModel(${m.id}, 'size_gb', this.value)" style="width:70px"></td>
             <td><input type="number" value="${m.layers}" onchange="updateModel(${m.id}, 'layers', this.value)" style="width:60px"></td>
             <td><input type="number" value="${m.embd}" onchange="updateModel(${m.id}, 'embd', this.value)" style="width:70px"></td>
@@ -254,6 +255,7 @@ function renderModelTable() {
 
 async function addModel() {
     const data = {
+        base_model: document.getElementById('model-basemodel').value,
         name: document.getElementById('model-name').value,
         size_gb: parseFloat(document.getElementById('model-size').value) || 0,
         layers: parseInt(document.getElementById('model-layers').value) || 0,
@@ -265,16 +267,16 @@ async function addModel() {
         description: document.getElementById('model-desc').value,
         sort_order: parseInt(document.getElementById('model-order').value) || 0,
     };
-    if (!data.name || !data.size_gb || !data.layers) { alert('请填写模型名称、大小和层数'); return; }
+    if (!data.base_model || !data.name || !data.size_gb || !data.layers) { alert('请填写基模型、名称、大小和层数'); return; }
     await fetch('/api/admin/models', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify(data) });
-    ['model-name','model-size','model-layers','model-embd','model-kv','model-hdim','model-heads','model-quant','model-desc','model-order'].forEach(id => document.getElementById(id).value = id === 'model-order' ? '0' : '');
+    ['model-basemodel','model-name','model-size','model-layers','model-embd','model-kv','model-hdim','model-heads','model-quant','model-desc','model-order'].forEach(id => document.getElementById(id).value = id === 'model-order' ? '0' : '');
     await loadAdminModels();
 }
 
 async function updateModel(id, field, value) {
     const m = adminState.models.find(m => m.id === id);
     if (!m) return;
-    const data = { name: m.name, size_gb: m.size_gb, layers: m.layers, embd: m.embd, kv_heads: m.kv_heads, head_dim: m.head_dim, attention_heads: m.attention_heads, quant: m.quant, description: m.description, sort_order: m.sort_order };
+    const data = { base_model: m.base_model, name: m.name, size_gb: m.size_gb, layers: m.layers, embd: m.embd, kv_heads: m.kv_heads, head_dim: m.head_dim, attention_heads: m.attention_heads, quant: m.quant, description: m.description, sort_order: m.sort_order };
     if (['size_gb'].includes(field)) data[field] = parseFloat(value) || 0;
     else if (['layers','embd','kv_heads','head_dim','attention_heads','sort_order'].includes(field)) data[field] = parseInt(value) || 0;
     else data[field] = value;
@@ -297,11 +299,27 @@ async function loadAdminSettings() {
 }
 
 function renderSettings() {
-    document.getElementById('settings-form').innerHTML = Object.entries(adminState.settings).map(([key, value]) => `
-        <div class="setting-item">
-            <label>${key}</label>
-            <input type="text" value="${value}" onchange="adminState.settings['${key}'] = this.value">
-        </div>`).join('');
+    const labels = {
+        'admin_password': '管理密码',
+        'default_gpu': '默认GPU',
+        'default_version': '默认版本',
+        'default_mode': '默认模式',
+        'default_quant': '默认量化版本',
+        'llm_enabled': '启用LLM解析 (true/false)',
+        'llm_api_url': 'LLM API URL',
+        'llm_api_key': 'LLM API Key',
+        'llm_api_model': 'LLM 模型名',
+        'llm_system_prompt': 'LLM 系统提示词',
+    };
+    document.getElementById('settings-form').innerHTML = Object.entries(adminState.settings).map(([key, value]) => {
+        const label = labels[key] || key;
+        const isLong = key === 'llm_system_prompt';
+        if (isLong) {
+            return `<div class="setting-item setting-item-wide"><label>${label}</label><textarea rows="4" onchange="adminState.settings['${key}'] = this.value">${value}</textarea></div>`;
+        }
+        const type = key === 'llm_api_key' ? 'password' : 'text';
+        return `<div class="setting-item"><label>${label}</label><input type="${type}" value="${value}" onchange="adminState.settings['${key}'] = this.value"></div>`;
+    }).join('');
 }
 
 async function saveSettings() {

static/js/main.js

@@ -1,59 +1,42 @@
 // ===== State =====
 let state = {
-    versions: [],
-    gpus: [],
-    models: [],
-    filteredModels: [],
-    selectedModel: null,
-    currentVersionId: null,
-    params: [],
-    paramValues: {},
-    mode: 'gpu',
-    gpuSlots: [],
-    showHidden: false,
-    currentTab: 'common',
-    paramSearchText: '',
+    versions: [], gpus: [], models: [], modelsGrouped: {},
+    baseModelList: [], filteredBaseModels: [],
+    currentVersionId: null, params: [], paramValues: {},
+    mode: 'gpu', gpuSlots: [], showHidden: false,
+    currentTab: 'common', paramSearchText: '',
+    selectedModel: null, defaultQuant: 'Q4_K_M',
 };
 
 // ===== Init =====
 async function init() {
     await loadVersions();
     await loadGpus();
-    await loadModels();
-    const defaultVer = state.versions.find(v => v.version_tag === 'b10068') || state.versions[0];
-    if (defaultVer) {
-        state.currentVersionId = defaultVer.id;
-        document.getElementById('version-select').value = defaultVer.id;
-        await loadParams(defaultVer.id);
-    }
-    const defaultGpu = state.gpus.find(g => g.name === 'RTX 3090') || state.gpus[0];
-    if (defaultGpu) {
-        state.gpuSlots = [{ ...defaultGpu }];
-        renderGpuSlots();
-    }
-    renderParams();
-    generateCommand();
-    updateEstimate();
+    await loadModelsGrouped();
+    const dv = state.versions.find(v => v.version_tag === 'b10068') || state.versions[0];
+    if (dv) { state.currentVersionId = dv.id; document.getElementById('version-select').value = dv.id; await loadParams(dv.id); }
+    const dg = state.gpus.find(g => g.name === 'RTX 3090') || state.gpus[0];
+    if (dg) { state.gpuSlots = [{ ...dg }]; renderGpuSlots(); }
+    renderParams(); generateCommand(); updateEstimate();
 }
 
 // ===== Load Data =====
 async function loadVersions() {
-    const res = await fetch('/api/versions');
-    state.versions = await res.json();
-    const sel = document.getElementById('version-select');
-    sel.innerHTML = state.versions.map(v => `<option value="${v.id}">${v.version_tag} - ${v.description}</option>`).join('');
+    const res = await fetch('/api/versions'); state.versions = await res.json();
+    document.getElementById('version-select').innerHTML = state.versions.map(v => `<option value="${v.id}">${v.version_tag} - ${v.description}</option>`).join('');
 }
+async function loadGpus() { const res = await fetch('/api/gpus'); state.gpus = await res.json(); }
 
-async function loadGpus() {
-    const res = await fetch('/api/gpus');
-    state.gpus = await res.json();
-}
-
-async function loadModels() {
-    const res = await fetch('/api/models');
-    state.models = await res.json();
-    state.filteredModels = state.models;
-    renderModelDropdown();
+async function loadModelsGrouped() {
+    const res = await fetch('/api/models/grouped');
+    const data = await res.json();
+    state.modelsGrouped = data.models || {};
+    state.defaultQuant = data.default_quant || 'Q4_K_M';
+    state.baseModelList = Object.keys(state.modelsGrouped).sort();
+    state.filteredBaseModels = state.baseModelList;
+    // Also flat list for backward compat
+    state.models = Object.values(state.modelsGrouped).flat();
+    renderBaseModelDropdown();
 }
 
 async function loadParams(versionId) {
@@ -61,98 +44,67 @@ async function loadParams(versionId) {
     state.params = await res.json();
     state.paramValues = {};
     state.params.forEach(p => {
-        if (p.param_type === 'boolean') {
-            state.paramValues[p.param_key] = p.default_value === 'true';
-        } else {
-            state.paramValues[p.param_key] = p.default_value || '';
-        }
+        state.paramValues[p.param_key] = p.param_type === 'boolean' ? (p.default_value === 'true') : (p.default_value || '');
     });
-    renderParams();
-    generateCommand();
-    updateEstimate();
+    renderParams(); generateCommand(); updateEstimate();
 }
 
 // ===== Version Change =====
-async function onVersionChange() {
-    const sel = document.getElementById('version-select');
-    state.currentVersionId = parseInt(sel.value);
-    await loadParams(state.currentVersionId);
-}
+async function onVersionChange() { state.currentVersionId = parseInt(document.getElementById('version-select').value); await loadParams(state.currentVersionId); }
 
 // ===== Mode Switch =====
 function switchMode(mode) {
     state.mode = mode;
-    document.querySelectorAll('.mode-btn').forEach(btn => {
-        btn.classList.toggle('active', btn.dataset.mode === mode);
-    });
+    document.querySelectorAll('.mode-btn').forEach(btn => btn.classList.toggle('active', btn.dataset.mode === mode));
     document.getElementById('memory-panel').classList.toggle('hidden', mode !== 'gpu_cpu');
-    generateCommand();
-    updateEstimate();
+    generateCommand(); updateEstimate();
 }
 
 // ===== GPU Slots =====
 function renderGpuSlots() {
-    const container = document.getElementById('gpu-slots');
-    container.innerHTML = state.gpuSlots.map((gpu, i) => {
-        const options = state.gpus.map(g =>
-            `<option value="${g.name}" ${g.name === gpu.name ? 'selected' : ''}>${g.name} (${(g.vram_mb / 1024).toFixed(0)}GB)</option>`
-        ).join('');
-        return `
-            <div class="gpu-slot">
-                <span class="gpu-index">GPU ${i + 1}</span>
-                <select onchange="updateGpuSlot(${i}, this.value)">
-                    <option value="">-- 选择GPU --</option>
-                    ${options}
-                </select>
-                <span class="vram-info">显存: ${(gpu.vram_mb / 1024).toFixed(1)} GB</span>
-                ${state.gpuSlots.length > 1 ? `<button class="btn-remove" onclick="removeGpuSlot(${i})">×</button>` : ''}
-            </div>
-        `;
+    document.getElementById('gpu-slots').innerHTML = state.gpuSlots.map((gpu, i) => {
+        const opts = state.gpus.map(g => `<option value="${g.name}" ${g.name === gpu.name ? 'selected' : ''}>${g.name} (${(g.vram_mb/1024).toFixed(0)}GB)</option>`).join('');
+        return `<div class="gpu-slot"><span class="gpu-index">GPU ${i+1}</span><select onchange="updateGpuSlot(${i}, this.value)"><option value="">-- 选择GPU --</option>${opts}</select><span class="vram-info">显存: ${(gpu.vram_mb/1024).toFixed(1)} GB</span>${state.gpuSlots.length > 1 ? `<button class="btn-remove" onclick="removeGpuSlot(${i})">×</button>` : ''}</div>`;
     }).join('');
     document.getElementById('btn-add-gpu').style.display = state.gpuSlots.length >= 4 ? 'none' : 'block';
     updateEstimate();
 }
+function addGpuSlot() { if (state.gpuSlots.length >= 4) return; const dg = state.gpus.find(g => g.name === 'RTX 3090') || state.gpus[0]; state.gpuSlots.push({ ...dg }); renderGpuSlots(); generateCommand(); }
+function removeGpuSlot(i) { state.gpuSlots.splice(i, 1); renderGpuSlots(); generateCommand(); }
+function updateGpuSlot(i, name) { const g = state.gpus.find(g => g.name === name); if (g) state.gpuSlots[i] = { ...g }; renderGpuSlots(); generateCommand(); }
 
-function addGpuSlot() {
-    if (state.gpuSlots.length >= 4) return;
-    const defaultGpu = state.gpus.find(g => g.name === 'RTX 3090') || state.gpus[0];
-    state.gpuSlots.push({ ...defaultGpu });
-    renderGpuSlots();
-    generateCommand();
-}
-
-function removeGpuSlot(index) {
-    state.gpuSlots.splice(index, 1);
-    renderGpuSlots();
-    generateCommand();
-}
-
-function updateGpuSlot(index, name) {
-    const gpu = state.gpus.find(g => g.name === name);
-    if (gpu) state.gpuSlots[index] = { ...gpu };
-    renderGpuSlots();
-    generateCommand();
-}
-
-// ===== Model Selection =====
-function filterModels() {
+// ===== Model Selection (two-step) =====
+function filterBaseModels() {
     const text = document.getElementById('model-search').value.toLowerCase();
-    state.filteredModels = state.models.filter(m =>
-        m.name.toLowerCase().includes(text) ||
-        (m.quant || '').toLowerCase().includes(text) ||
-        (m.description || '').toLowerCase().includes(text)
-    );
-    renderModelDropdown();
-    document.getElementById('model-dropdown').style.display = state.filteredModels.length > 0 ? 'block' : 'none';
+    state.filteredBaseModels = state.baseModelList.filter(n => n.toLowerCase().includes(text));
+    renderBaseModelDropdown();
+    document.getElementById('model-dropdown').style.display = state.filteredBaseModels.length > 0 ? 'block' : 'none';
 }
 
-function renderModelDropdown() {
-    const dd = document.getElementById('model-dropdown');
-    dd.innerHTML = state.filteredModels.map(m => `
-        <div class="model-option" onclick="selectModel(${m.id})">
-            <span class="model-name">${m.name}</span>
-            <span class="model-meta">${m.size_gb}GB | ${m.layers}L | ${m.quant || ''}</span>
-        </div>
+function renderBaseModelDropdown() {
+    document.getElementById('model-dropdown').innerHTML = state.filteredBaseModels.map(n => {
+        const quants = state.modelsGrouped[n] || [];
+        const quantStr = quants.map(q => q.quant).filter(Boolean).join(', ');
+        return `<div class="model-option" onclick="selectBaseModel('${n}')"><span class="model-name">${n}</span><span class="model-meta">${quantStr}</span></div>`;
+    }).join('');
+}
+
+function selectBaseModel(baseName) {
+    document.getElementById('model-search').value = baseName;
+    document.getElementById('model-dropdown').style.display = 'none';
+    const variants = state.modelsGrouped[baseName] || [];
+    if (variants.length === 0) return;
+    // Show quant step
+    document.getElementById('quant-step').classList.remove('hidden');
+    // Auto-select default quant
+    const defaultVariant = variants.find(v => v.quant === state.defaultQuant) || variants[0];
+    renderQuantOptions(baseName, variants, defaultVariant.id);
+    selectModel(defaultVariant.id);
+}
+
+function renderQuantOptions(baseName, variants, selectedId) {
+    document.getElementById('quant-options').innerHTML = variants.map(v => `
+        <button class="quant-btn ${v.id === selectedId ? 'active' : ''}" onclick="selectModel(${v.id})">${v.quant || 'FP16'}<span class="quant-size">${v.size_gb}GB</span></button>
     `).join('');
 }
 
@@ -160,19 +112,20 @@ function selectModel(id) {
     const m = state.models.find(m => m.id === id);
     if (!m) return;
     state.selectedModel = m;
-    document.getElementById('model-search').value = m.name;
-    document.getElementById('model-dropdown').style.display = 'none';
+    // Update quant buttons active state
+    const variants = state.modelsGrouped[m.base_model] || [];
+    renderQuantOptions(m.base_model, variants, id);
+    // Show model detail
     document.getElementById('model-selected-info').innerHTML = `
         <div class="model-detail">
+            <span class="detail-item"><b>模型:</b> ${m.name}</span>
             <span class="detail-item"><b>大小:</b> ${m.size_gb} GB</span>
             <span class="detail-item"><b>层数:</b> ${m.layers}</span>
-            <span class="detail-item"><b>嵌入维度:</b> ${m.embd}</span>
+            <span class="detail-item"><b>嵌入:</b> ${m.embd}</span>
             <span class="detail-item"><b>KV Heads:</b> ${m.kv_heads}</span>
             <span class="detail-item"><b>Head Dim:</b> ${m.head_dim}</span>
-            <span class="detail-item"><b>Attention Heads:</b> ${m.attention_heads}</span>
             <span class="detail-item"><b>量化:</b> ${m.quant || 'N/A'}</span>
-        </div>
-    `;
+        </div>`;
     updateEstimate();
 }
 
@@ -181,25 +134,39 @@ function switchTab(cat) {
     state.currentTab = cat;
     state.paramSearchText = '';
     document.getElementById('param-search').value = '';
-    document.querySelectorAll('.tab-btn').forEach(btn => {
-        btn.classList.toggle('active', btn.dataset.cat === cat);
-    });
+    document.getElementById('search-hint').textContent = '';
+    document.querySelectorAll('.tab-btn').forEach(btn => btn.classList.toggle('active', btn.dataset.cat === cat));
+    document.getElementById('param-tabs').style.display = '';
     renderParams();
 }
 
 function onParamSearch() {
     state.paramSearchText = document.getElementById('param-search').value.toLowerCase();
+    const hint = document.getElementById('search-hint');
+    if (state.paramSearchText) {
+        const count = state.params.filter(p =>
+            p.param_key.toLowerCase().includes(state.paramSearchText) ||
+            (p.long_flag||'').toLowerCase().includes(state.paramSearchText) ||
+            (p.short_flag||'').toLowerCase().includes(state.paramSearchText) ||
+            (p.description||'').toLowerCase().includes(state.paramSearchText)
+        ).length;
+        hint.textContent = `找到 ${count} 个匹配参数`;
+        document.getElementById('param-tabs').style.display = 'none';
+    } else {
+        hint.textContent = '';
+        document.getElementById('param-tabs').style.display = '';
+    }
     renderParams();
 }
 
 function getFilteredParams() {
     let params = state.params.filter(p => p.category === state.currentTab);
     if (state.paramSearchText) {
-        params = params.filter(p =>
+        params = state.params.filter(p =>
             p.param_key.toLowerCase().includes(state.paramSearchText) ||
-            (p.long_flag || '').toLowerCase().includes(state.paramSearchText) ||
-            (p.short_flag || '').toLowerCase().includes(state.paramSearchText) ||
-            (p.description || '').toLowerCase().includes(state.paramSearchText)
+            (p.long_flag||'').toLowerCase().includes(state.paramSearchText) ||
+            (p.short_flag||'').toLowerCase().includes(state.paramSearchText) ||
+            (p.description||'').toLowerCase().includes(state.paramSearchText)
         );
     }
     return params;
@@ -207,121 +174,72 @@ function getFilteredParams() {
 
 function renderParams() {
     const container = document.getElementById('param-container');
-    const allParams = getFilteredParams();
-
-    // When searching, show all matching regardless of important/hidden
+    const all = getFilteredParams();
     if (state.paramSearchText) {
-        container.innerHTML = allParams.map(p => renderParamItem(p)).join('');
+        container.innerHTML = all.map(p => renderParamItem(p)).join('');
         document.getElementById('toggle-advanced-btn').style.display = 'none';
         return;
     }
-
-    const importantParams = allParams.filter(p => p.is_important === 1);
-    const otherParams = allParams.filter(p => p.is_important === 0);
-
-    let html = importantParams.map(p => renderParamItem(p)).join('');
-    if (state.showHidden) {
-        html += otherParams.map(p => renderParamItem(p)).join('');
-    }
-
+    const imp = all.filter(p => p.is_important === 1);
+    const other = all.filter(p => p.is_important === 0);
+    let html = imp.map(p => renderParamItem(p)).join('');
+    if (state.showHidden) html += other.map(p => renderParamItem(p)).join('');
     container.innerHTML = html;
-
-    const toggleBtn = document.getElementById('toggle-advanced-btn');
-    if (otherParams.length > 0 && !state.paramSearchText) {
-        toggleBtn.style.display = 'block';
-        toggleBtn.textContent = state.showHidden ? '▲ 收起更多参数' : '▼ 显示更多参数';
-    } else {
-        toggleBtn.style.display = 'none';
-    }
+    const btn = document.getElementById('toggle-advanced-btn');
+    if (other.length > 0 && !state.paramSearchText) { btn.style.display = 'block'; btn.textContent = state.showHidden ? '▲ 收起更多参数' : '▼ 显示更多参数'; }
+    else btn.style.display = 'none';
 }
 
 function renderParamItem(p) {
     const val = state.paramValues[p.param_key];
-    const defaultVal = p.default_value;
-    const isModified = isParamModified(p, val);
-    const modifiedClass = isModified ? 'modified' : '';
-    const vramBadge = p.affects_vram ? '<span class="vram-badge">⚡显存</span>' : '';
+    const isMod = isParamModified(p, val);
+    const mc = isMod ? 'modified' : '';
+    const vb = p.affects_vram ? '<span class="vram-badge">⚡显存</span>' : '';
     const flag = p.short_flag || p.long_flag;
-
-    let inputHtml = '';
+    let inp = '';
     if (p.param_type === 'boolean') {
-        inputHtml = `<input type="checkbox" ${val ? 'checked' : ''} onchange="setParam('${p.param_key}', this.checked)">`;
+        inp = `<input type="checkbox" ${val ? 'checked' : ''} onchange="setParam('${p.param_key}', this.checked)">`;
     } else if (p.param_type === 'select') {
-        const options = Array.isArray(p.options) ? p.options : [];
-        inputHtml = `<select onchange="setParam('${p.param_key}', this.value)">
-            ${options.map(o => `<option value="${o}" ${String(val) === String(o) ? 'selected' : ''}>${o}</option>`).join('')}
-        </select>`;
+        const opts = Array.isArray(p.options) ? p.options : [];
+        inp = `<select onchange="setParam('${p.param_key}', this.value)">${opts.map(o => `<option value="${o}" ${String(val)===String(o) ? 'selected':''}>${o}</option>`).join('')}</select>`;
     } else if (p.param_type === 'number') {
         const step = p.step || 'any';
         const min = p.min_value !== null ? `min="${p.min_value}"` : '';
         const max = p.max_value !== null ? `max="${p.max_value}"` : '';
-        inputHtml = `<input type="number" value="${val}" step="${step}" ${min} ${max}
-            onchange="setParam('${p.param_key}', this.value)" data-key="${p.param_key}">`;
+        inp = `<input type="number" value="${val}" step="${step}" ${min} ${max} onchange="setParam('${p.param_key}', this.value)" data-key="${p.param_key}">`;
     } else {
-        inputHtml = `<input type="text" value="${val}" onchange="setParam('${p.param_key}', this.value)" data-key="${p.param_key}">`;
+        inp = `<input type="text" value="${val}" onchange="setParam('${p.param_key}', this.value)" data-key="${p.param_key}">`;
     }
-
-    const unitHtml = p.unit ? `<span class="param-unit">${p.unit}</span>` : '';
-
-    return `
-        <div class="param-item ${modifiedClass}" data-key="${p.param_key}">
-            <span class="param-flag">${flag}</span>
-            ${vramBadge}
-            <span class="param-desc-text">${p.description}</span>
-            ${inputHtml}
-            ${unitHtml}
-        </div>
-    `;
+    const uh = p.unit ? `<span class="param-unit">${p.unit}</span>` : '';
+    return `<div class="param-item ${mc}" data-key="${p.param_key}"><span class="param-flag">${flag}</span>${vb}<span class="param-desc-text">${p.description}</span>${inp}${uh}</div>`;
 }
 
 function isParamModified(p, val) {
-    const defaultVal = p.default_value;
-    if (p.param_type === 'boolean') {
-        return (val === true || val === 'true') !== (defaultVal === 'true');
-    }
-    return String(val) !== String(defaultVal);
+    const dv = p.default_value;
+    if (p.param_type === 'boolean') return (val===true||val==='true') !== (dv==='true');
+    return String(val) !== String(dv);
 }
 
 function setParam(key, value) {
     state.paramValues[key] = value;
-    // Only update modified visual state, command, and estimate — do NOT re-render
     const item = document.querySelector(`.param-item[data-key="${key}"]`);
-    if (item) {
-        const p = state.params.find(p => p.param_key === key);
-        if (p && isParamModified(p, value)) {
-            item.classList.add('modified');
-        } else {
-            item.classList.remove('modified');
-        }
-    }
-    generateCommand();
-    updateEstimate();
+    if (item) { const p = state.params.find(p => p.param_key === key); if (p && isParamModified(p, value)) item.classList.add('modified'); else item.classList.remove('modified'); }
+    generateCommand(); updateEstimate();
 }
 
-function toggleHiddenParams() {
-    state.showHidden = !state.showHidden;
-    renderParams();
-}
+function toggleHiddenParams() { state.showHidden = !state.showHidden; renderParams(); }
 
 // ===== VRAM Estimation =====
 async function updateEstimate() {
     const params = collectParamsForEstimate();
-    const gpuSelections = state.gpuSlots.map(s => ({ name: s.name, vram_mb: s.vram_mb }));
-
-    const sysMemory = parseFloat(document.getElementById('sys-memory').value) || 0;
-    const sysMemUnit = document.getElementById('sys-memory-unit').value;
-    let sysMemoryGb = 0;
-    if (sysMemUnit === '1') sysMemoryGb = sysMemory;
-    else if (sysMemUnit === '2') sysMemoryGb = sysMemory / 1024;
-
-    const res = await fetch('/api/estimate', {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ params, gpu_selections: gpuSelections, mode: state.mode, system_memory_gb: sysMemoryGb })
-    });
+    const gpuSel = state.gpuSlots.map(s => ({ name: s.name, vram_mb: s.vram_mb }));
+    const sm = parseFloat(document.getElementById('sys-memory').value) || 0;
+    const smu = document.getElementById('sys-memory-unit').value;
+    let sysGb = 0; if (smu === '1') sysGb = sm; else if (smu === '2') sysGb = sm / 1024;
+    const res = await fetch('/api/estimate', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ params, gpu_selections: gpuSel, mode: state.mode, system_memory_gb: sysGb }) });
     const data = await res.json();
     renderVramDisplay(data);
-    if (state.mode === 'gpu_cpu') renderRamDisplay(data, sysMemoryGb);
+    if (state.mode === 'gpu_cpu') renderRamDisplay(data, sysGb);
 }
 
 function collectParamsForEstimate() {
@@ -333,124 +251,49 @@ function collectParamsForEstimate() {
         params._model_kv_heads = state.selectedModel.kv_heads;
         params._model_head_dim = state.selectedModel.head_dim;
         params._model_heads = state.selectedModel.attention_heads;
-    } else {
-        params._model_size_gb = 0;
-        params._model_layers = 0;
-        params._model_embd = 0;
-        params._model_kv_heads = 0;
-        params._model_head_dim = 0;
-        params._model_heads = 0;
-    }
+    } else { params._model_size_gb = 0; params._model_layers = 0; params._model_embd = 0; params._model_kv_heads = 0; params._model_head_dim = 0; params._model_heads = 0; }
     return params;
 }
 
 function renderVramDisplay(data) {
-    const bar = document.getElementById('vram-bar');
-    const label = document.getElementById('vram-label');
-    const breakdown = document.getElementById('vram-breakdown');
-
-    if (!data.total_vram_available_mb || data.total_vram_available_mb === 0) {
-        bar.style.width = '0%';
-        label.textContent = '请选择GPU';
-        breakdown.innerHTML = '';
-        return;
-    }
-
-    const percent = data.usage_percent || 0;
-    bar.style.width = Math.min(percent, 100) + '%';
-    bar.className = 'vram-bar' + (percent > 90 ? ' danger' : percent > 75 ? ' warning' : '');
-    label.textContent = `VRAM: ${data.total_gb}GB / ${data.total_vram_available_gb}GB (${percent}%)`;
-
-    breakdown.innerHTML = `
-        <div class="breakdown-item"><span class="label">模型权重</span><span class="value">${data.weights_gb}GB</span></div>
-        <div class="breakdown-item"><span class="label">KV缓存</span><span class="value">${data.kv_cache_gb}GB</span></div>
-        <div class="breakdown-item"><span class="label">计算/开销</span><span class="value">${(data.compute_mb / 1024).toFixed(2)}GB</span></div>
-        <div class="breakdown-item"><span class="label">CUDA开销</span><span class="value">${(data.cuda_overhead_mb / 1024).toFixed(2)}GB</span></div>
-    `;
+    const bar = document.getElementById('vram-bar'), label = document.getElementById('vram-label'), bd = document.getElementById('vram-breakdown');
+    if (!data.total_vram_available_mb) { bar.style.width = '0%'; label.textContent = '请选择GPU'; bd.innerHTML = ''; return; }
+    const pct = data.usage_percent || 0;
+    bar.style.width = Math.min(pct, 100) + '%';
+    bar.className = 'vram-bar' + (pct > 90 ? ' danger' : pct > 75 ? ' warning' : '');
+    label.textContent = `VRAM: ${data.total_gb}GB / ${data.total_vram_available_gb}GB (${pct}%)`;
+    bd.innerHTML = `<div class="breakdown-item"><span class="label">模型权重</span><span class="value">${data.weights_gb}GB</span></div><div class="breakdown-item"><span class="label">KV缓存</span><span class="value">${data.kv_cache_gb}GB</span></div><div class="breakdown-item"><span class="label">计算/开销</span><span class="value">${(data.compute_mb/1024).toFixed(2)}GB</span></div><div class="breakdown-item"><span class="label">CUDA开销</span><span class="value">${(data.cuda_overhead_mb/1024).toFixed(2)}GB</span></div>`;
 }
 
-function renderRamDisplay(data, sysMemoryGb) {
-    const bar = document.getElementById('ram-bar');
-    const label = document.getElementById('ram-label');
-    const breakdown = document.getElementById('ram-breakdown');
-
-    if (sysMemoryGb === 0) {
-        bar.style.width = '0%';
-        label.textContent = `内存: ${data.cpu_total_gb}GB (无上限)`;
-    } else {
-        const percent = data.cpu_usage_percent || 0;
-        bar.style.width = Math.min(percent, 100) + '%';
-        bar.className = 'vram-bar ram-bar' + (percent > 90 ? ' danger' : percent > 75 ? ' warning' : '');
-        label.textContent = `内存: ${data.cpu_total_gb}GB / ${sysMemoryGb}GB (${percent}%)`;
-    }
-
-    breakdown.innerHTML = `
-        <div class="breakdown-item"><span class="label">CPU模型权重</span><span class="value">${data.cpu_weights_gb}GB</span></div>
-        <div class="breakdown-item"><span class="label">CPU KV缓存</span><span class="value">${(data.cpu_kv_cache_mb / 1024).toFixed(2)}GB</span></div>
-    `;
+function renderRamDisplay(data, sysGb) {
+    const bar = document.getElementById('ram-bar'), label = document.getElementById('ram-label'), bd = document.getElementById('ram-breakdown');
+    if (sysGb === 0) { bar.style.width = '0%'; label.textContent = `内存: ${data.cpu_total_gb}GB (无上限)`; }
+    else { const pct = data.cpu_usage_percent || 0; bar.style.width = Math.min(pct, 100) + '%'; bar.className = 'vram-bar ram-bar' + (pct > 90 ? ' danger' : pct > 75 ? ' warning' : ''); label.textContent = `内存: ${data.cpu_total_gb}GB / ${sysGb}GB (${pct}%)`; }
+    bd.innerHTML = `<div class="breakdown-item"><span class="label">CPU模型权重</span><span class="value">${data.cpu_weights_gb}GB</span></div><div class="breakdown-item"><span class="label">CPU KV缓存</span><span class="value">${(data.cpu_kv_cache_mb/1024).toFixed(2)}GB</span></div>`;
 }
 
 // ===== Natural Language =====
 async function parseNaturalLanguage() {
     const text = document.getElementById('nl-input').value;
     if (!text.trim()) return;
-
-    const res = await fetch('/api/parse-nl', {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ text })
-    });
+    const res = await fetch('/api/parse-nl', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ text }) });
     const data = await res.json();
-
-    let resultHtml = '<div>解析结果:</div>';
-
+    let html = '<div>解析结果:</div>';
     if (data._gpu_name) {
-        const gpu = state.gpus.find(g => g.name === data._gpu_name);
-        if (gpu) {
-            if (data._gpu_count && data._gpu_count > 1) {
-                state.gpuSlots = [];
-                for (let i = 0; i < Math.min(data._gpu_count, 4); i++) state.gpuSlots.push({ ...gpu });
-            } else {
-                state.gpuSlots = [{ ...gpu }];
-            }
-            renderGpuSlots();
-            resultHtml += `<span class="parsed-param"><span class="key">GPU:</span> <span class="value">${data._gpu_name}${data._gpu_count > 1 ? ' x' + data._gpu_count : ''}</span></span>`;
-        }
+        const g = state.gpus.find(g => g.name === data._gpu_name);
+        if (g) { const cnt = data._gpu_count > 1 ? Math.min(data._gpu_count, 4) : 1; state.gpuSlots = []; for (let i = 0; i < cnt; i++) state.gpuSlots.push({ ...g }); renderGpuSlots(); html += `<span class="parsed-param"><span class="key">GPU:</span> <span class="value">${data._gpu_name}${cnt > 1 ? ' x'+cnt : ''}</span></span>`; }
     }
-
-    if (data._mode) {
-        switchMode(data._mode);
-        resultHtml += `<span class="parsed-param"><span class="key">模式:</span> <span class="value">${data._mode}</span></span>`;
-    }
-
-    for (const [key, value] of Object.entries(data)) {
-        if (key.startsWith('_')) continue;
-        state.paramValues[key] = value;
-        resultHtml += `<span class="parsed-param"><span class="key">${key}:</span> <span class="value">${value}</span></span>`;
-    }
-
-    document.getElementById('nl-result').innerHTML = resultHtml;
-    renderParams();
-    generateCommand();
-    updateEstimate();
+    if (data._mode) { switchMode(data._mode); html += `<span class="parsed-param"><span class="key">模式:</span> <span class="value">${data._mode}</span></span>`; }
+    for (const [k, v] of Object.entries(data)) { if (k.startsWith('_')) continue; state.paramValues[k] = v; html += `<span class="parsed-param"><span class="key">${k}:</span> <span class="value">${v}</span></span>`; }
+    document.getElementById('nl-result').innerHTML = html;
+    renderParams(); generateCommand(); updateEstimate();
 }
 
 // ===== Generate Command =====
 async function generateCommand() {
     const params = { ...state.paramValues };
     Object.keys(params).forEach(k => { if (k.startsWith('_')) delete params[k]; });
-
-    const res = await fetch('/api/generate', {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({
-            version_id: state.currentVersionId,
-            params,
-            mode: state.mode,
-            gpu_selections: state.gpuSlots,
-            binary: document.getElementById('binary-select').value,
-        })
-    });
+    const res = await fetch('/api/generate', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ version_id: state.currentVersionId, params, mode: state.mode, gpu_selections: state.gpuSlots, binary: document.getElementById('binary-select').value }) });
     const data = await res.json();
     document.getElementById('command-output').textContent = data.command;
 }
@@ -458,24 +301,11 @@ async function generateCommand() {
 // ===== Copy =====
 function copyCommand() {
     const text = document.getElementById('command-output').textContent;
-    // Use textarea fallback for non-HTTPS environments
     const ta = document.createElement('textarea');
-    ta.value = text;
-    ta.style.position = 'fixed';
-    ta.style.left = '-9999px';
-    document.body.appendChild(ta);
-    ta.select();
-    try {
-        document.execCommand('copy');
-        const btn = document.querySelector('.btn-copy');
-        const orig = btn.textContent;
-        btn.textContent = '✅ 已复制!';
-        setTimeout(() => btn.textContent = orig, 2000);
-    } catch(e) {
-        alert('复制失败，请手动选择文本复制');
-    }
+    ta.value = text; ta.style.position = 'fixed'; ta.style.left = '-9999px';
+    document.body.appendChild(ta); ta.select();
+    try { document.execCommand('copy'); const btn = document.querySelector('.btn-copy'); const o = btn.textContent; btn.textContent = '✅ 已复制!'; setTimeout(() => btn.textContent = o, 2000); } catch(e) { alert('复制失败，请手动选择文本复制'); }
     document.body.removeChild(ta);
 }
 
-// ===== Init =====
 window.addEventListener('DOMContentLoaded', init);