feat: v1.4.0 - 五项优化

1. 模型选择: 点击搜索框弹出top5模型，支持搜索过滤
2. 选模型后自动调整参数: ctx_size设为模型默认上下文，n_gpu_layers设为all
   - models表新增 default_ctx 字段，各模型预设上下文值
3. 参数配置新增「已改参数」标签页
   - 放在最右边，绿色样式区别于其他标签
   - 显示所有被修改过(非默认值)的参数
4. 复制按钮移到命令文本框左上角悬空
5. 自然语言默认提示文本可在后台管理中编辑
   - 新增 /api/settings/public 接口
   - 后台系统设置新增 nl_default_text 配置项

app.py

@@ -621,6 +621,15 @@ def get_models_grouped():
     return jsonify({'models': grouped, 'default_quant': default_quant})
 
 
+# ----- Public Settings (only nl_default_text) -----
+@app.route('/api/settings/public')
+def get_public_settings():
+    db = get_db()
+    nl = db.execute("SELECT value FROM settings WHERE key = 'nl_default_text'").fetchone()
+    db.close()
+    return jsonify({'nl_default_text': nl['value'] if nl else ''})
+
+
 # ----- Parse Natural Language -----
 @app.route('/api/parse-nl', methods=['POST'])
 def parse_nl():
@@ -871,11 +880,11 @@ def admin_add_model():
     data = request.json
     db = get_db()
     db.execute(
-        '''INSERT INTO models (base_model, name, size_gb, layers, embd, kv_heads, head_dim, attention_heads, quant, description, sort_order)
-        VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)''',
+        '''INSERT INTO models (base_model, name, size_gb, layers, embd, kv_heads, head_dim, attention_heads, default_ctx, 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))
+         data.get('default_ctx', 0), data.get('quant', ''), data.get('description', ''), data.get('sort_order', 0))
     )
     db.commit()
     db.close()
@@ -890,10 +899,10 @@ def admin_model_edit(mid):
         data = request.json
         db.execute(
             '''UPDATE models SET base_model=?, name=?, size_gb=?, layers=?, embd=?, kv_heads=?,
-            head_dim=?, attention_heads=?, quant=?, description=?, sort_order=? WHERE id=?''',
+            head_dim=?, attention_heads=?, default_ctx=?, quant=?, description=?, sort_order=? WHERE id=?''',
             (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)
+             data.get('default_ctx', 0), data.get('quant', ''), data.get('description', ''), data.get('sort_order', 0), mid)
         )
         db.commit()
         db.close()

db.py

@@ -88,6 +88,7 @@ def init_db():
             kv_heads INTEGER NOT NULL,
             head_dim INTEGER NOT NULL,
             attention_heads INTEGER NOT NULL,
+            default_ctx INTEGER DEFAULT 0,
             quant TEXT DEFAULT '',
             description TEXT DEFAULT '',
             sort_order INTEGER DEFAULT 0
@@ -313,66 +314,66 @@ 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)
+    # (base_model, name, size_gb, layers, embd, kv_heads, head_dim, attention_heads, default_ctx, quant, description, sort_order)
     default_models = [
-        # 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),
+        # Llama-3-8B-Instruct (ctx 8192)
+        ("Llama-3-8B-Instruct", "Llama-3-8B-Instruct (Q4_K_M)", 4.9, 32, 4096, 8, 128, 32, 8192, "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, 8192, "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, 8192, "FP16", "Meta Llama 3 8B Instruct, FP16", 3),
+        # Llama-3-70B-Instruct (ctx 8192)
+        ("Llama-3-70B-Instruct", "Llama-3-70B-Instruct (Q4_K_M)", 38.5, 80, 8192, 8, 128, 64, 8192, "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, 8192, "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, 8192, "FP16", "Meta Llama 3 70B Instruct, FP16", 6),
+        # Llama-3.1-8B-Instruct (ctx 131072)
+        ("Llama-3.1-8B-Instruct", "Llama-3.1-8B-Instruct (Q4_K_M)", 4.9, 32, 4096, 8, 128, 32, 131072, "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, 131072, "Q8_0", "Meta Llama 3.1 8B Instruct, Q8_0 量化", 8),
+        # Llama-3.1-70B-Instruct (ctx 131072)
+        ("Llama-3.1-70B-Instruct", "Llama-3.1-70B-Instruct (Q4_K_M)", 38.5, 80, 8192, 8, 128, 64, 131072, "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, 131072, "Q8_0", "Meta Llama 3.1 70B Instruct, Q8_0 量化", 10),
+        # Llama-3.1-405B-Instruct (ctx 131072)
+        ("Llama-3.1-405B-Instruct", "Llama-3.1-405B-Instruct (Q4_K_M)", 226.0, 126, 16384, 8, 128, 128, 131072, "Q4_K_M", "Meta Llama 3.1 405B Instruct, Q4_K_M 量化", 11),
+        # Qwen2.5-7B-Instruct (ctx 32768)
+        ("Qwen2.5-7B-Instruct", "Qwen2.5-7B-Instruct (Q4_K_M)", 4.7, 28, 3584, 4, 128, 28, 32768, "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, 32768, "Q8_0", "Qwen2.5 7B Instruct, Q8_0 量化", 13),
+        # Qwen2.5-14B-Instruct (ctx 32768)
+        ("Qwen2.5-14B-Instruct", "Qwen2.5-14B-Instruct (Q4_K_M)", 8.7, 40, 5120, 8, 128, 40, 32768, "Q4_K_M", "Qwen2.5 14B Instruct, Q4_K_M 量化", 14),
+        # Qwen2.5-32B-Instruct (ctx 32768)
+        ("Qwen2.5-32B-Instruct", "Qwen2.5-32B-Instruct (Q4_K_M)", 19.5, 64, 5120, 8, 128, 64, 32768, "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, 32768, "Q8_0", "Qwen2.5 32B Instruct, Q8_0 量化", 16),
+        # Qwen2.5-72B-Instruct (ctx 32768)
+        ("Qwen2.5-72B-Instruct", "Qwen2.5-72B-Instruct (Q4_K_M)", 42.0, 80, 8192, 8, 128, 64, 32768, "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, 32768, "Q8_0", "Qwen2.5 72B Instruct, Q8_0 量化", 18),
+        # DeepSeek-V2-Chat (ctx 4096)
+        ("DeepSeek-V2-Chat", "DeepSeek-V2-Chat (Q4_K_M)", 23.0, 60, 5120, 8, 128, 60, 4096, "Q4_K_M", "DeepSeek V2 Chat, Q4_K_M 量化", 19),
+        # DeepSeek-V2.5-Chat (ctx 4096)
+        ("DeepSeek-V2.5-Chat", "DeepSeek-V2.5-Chat (Q4_K_M)", 23.0, 60, 5120, 8, 128, 60, 4096, "Q4_K_M", "DeepSeek V2.5 Chat, Q4_K_M 量化", 20),
+        # DeepSeek-R1-Distill-Qwen-32B (ctx 131072)
+        ("DeepSeek-R1-Distill-Qwen-32B", "DeepSeek-R1-Distill-Qwen-32B (Q4_K_M)", 19.5, 64, 5120, 8, 128, 64, 131072, "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, 131072, "Q8_0", "DeepSeek R1 Distill Qwen 32B, Q8_0 量化", 22),
+        # DeepSeek-R1-Distill-Llama-70B (ctx 131072)
+        ("DeepSeek-R1-Distill-Llama-70B", "DeepSeek-R1-Distill-Llama-70B (Q4_K_M)", 42.0, 80, 8192, 8, 128, 64, 131072, "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, 131072, "Q8_0", "DeepSeek R1 Distill Llama 70B, Q8_0 量化", 24),
+        # Mistral-7B-Instruct-v0.3 (ctx 32768)
+        ("Mistral-7B-Instruct-v0.3", "Mistral-7B-Instruct-v0.3 (Q4_K_M)", 4.4, 32, 4096, 8, 128, 32, 32768, "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, 32768, "Q8_0", "Mistral 7B Instruct v0.3, Q8_0 量化", 26),
+        # Mixtral-8x7B-Instruct (ctx 32768)
+        ("Mixtral-8x7B-Instruct", "Mixtral-8x7B-Instruct (Q4_K_M)", 26.0, 32, 4096, 8, 128, 32, 32768, "Q4_K_M", "Mixtral 8x7B Instruct, Q4_K_M 量化", 27),
+        # Gemma-2-9B-It (ctx 8192)
+        ("Gemma-2-9B-It", "Gemma-2-9B-It (Q4_K_M)", 5.4, 42, 3584, 4, 256, 14, 8192, "Q4_K_M", "Google Gemma 2 9B It, Q4_K_M 量化", 28),
+        # Gemma-2-27B-It (ctx 8192)
+        ("Gemma-2-27B-It", "Gemma-2-27B-It (Q4_K_M)", 16.5, 46, 4608, 4, 128, 36, 8192, "Q4_K_M", "Google Gemma 2 27B It, Q4_K_M 量化", 29),
+        # Phi-3-Mini-4K-Instruct (ctx 4096)
+        ("Phi-3-Mini-4K-Instruct", "Phi-3-Mini-4K-Instruct (Q4_K_M)", 2.5, 32, 3072, 32, 96, 32, 4096, "Q4_K_M", "Microsoft Phi-3 Mini 4K Instruct, Q4_K_M 量化", 30),
+        # Phi-3-Medium-14B-Instruct (ctx 14336)
+        ("Phi-3-Medium-14B-Instruct", "Phi-3-Medium-14B-Instruct (Q4_K_M)", 8.4, 40, 5120, 10, 128, 40, 14336, "Q4_K_M", "Microsoft Phi-3 Medium 14B Instruct, Q4_K_M 量化", 31),
+        # GLM-4-9B-Chat (ctx 131072)
+        ("GLM-4-9B-Chat", "GLM-4-9B-Chat (Q4_K_M)", 5.5, 40, 4096, 4, 128, 40, 131072, "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, 131072, "Q8_0", "Zhipu GLM-4 9B Chat, Q8_0 量化", 33),
     ]
 
     for m in default_models:
-        c.execute('''INSERT INTO models (base_model, 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, default_ctx, quant, description, sort_order)
+                     VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)''', m)
 
     # ===== Default settings =====
     c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("admin_password", "admin123"))
@@ -385,7 +386,7 @@ def insert_default_data(conn):
     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。"))
+    c.execute("INSERT OR IGNORE INTO settings (key, value) VALUES (?, ?)", ("nl_default_text", "用自然语言描述你想要的配置，支持多行输入。\n例如：\n用RTX 4090跑Llama-3-70B，上下文8192\n温度0.7，开启flash attention\n端口设为8080"))
 
     conn.commit()
 

static/admin.html

@@ -131,6 +131,7 @@
                     <input type="number" id="model-kv" placeholder="KV Heads">
                     <input type="number" id="model-hdim" placeholder="Head Dim">
                     <input type="number" id="model-heads" placeholder="Attention Heads">
+                    <input type="number" id="model-ctx" placeholder="默认上下文" value="0">
                     <input type="text" id="model-quant" placeholder="量化 (如: Q4_K_M)">
                     <input type="text" id="model-desc" placeholder="描述">
                     <input type="number" id="model-order" placeholder="排序" value="0">
@@ -138,7 +139,7 @@
                 </div>
             </div>
             <table class="admin-table">
-                <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>
+                <thead><tr><th>ID</th><th>基模型</th><th>名称</th><th>大小</th><th>层数</th><th>EMBD</th><th>KV</th><th>HD</th><th>Heads</th><th>默认CTX</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

@@ -14,6 +14,102 @@
     --radius: 8px;
 }
 
+/* ===== Sticky Memory Bar ===== */
+.sticky-mem-bar {
+    position: sticky;
+    top: 0;
+    z-index: 200;
+    background: var(--bg-panel);
+    border: 1px solid var(--border);
+    border-radius: var(--radius);
+    padding: 6px 12px;
+    margin-bottom: 16px;
+    box-shadow: 0 2px 8px rgba(0,0,0,0.4);
+}
+.sticky-vram-row, .sticky-ram-row {
+    display: flex;
+    align-items: center;
+    gap: 8px;
+}
+.sticky-ram-row { margin-top: 4px; }
+.sticky-ram-row.hidden { display: none; }
+.sticky-label {
+    font-size: 0.8em;
+    font-weight: 600;
+    color: var(--text-dim);
+    min-width: 60px;
+}
+.sticky-bar-container {
+    position: relative;
+    flex: 1;
+    height: 22px;
+    background: var(--bg);
+    border-radius: 4px;
+    overflow: hidden;
+    border: 1px solid var(--border);
+}
+.sticky-bar {
+    height: 100%;
+    background: linear-gradient(90deg, #4ecca3, #e9c46a);
+    border-radius: 3px;
+    transition: width 0.3s;
+    width: 0%;
+}
+.sticky-bar.warning { background: linear-gradient(90deg, #e9c46a, #f4a261); }
+.sticky-bar.danger { background: linear-gradient(90deg, #f4a261, #e94560); }
+.ram-sticky-bar { background: linear-gradient(90deg, #4e9cca, #4ecca3); }
+.sticky-bar-text {
+    position: absolute;
+    top: 50%; left: 50%;
+    transform: translate(-50%, -50%);
+    font-size: 0.75em;
+    font-weight: 600;
+    color: white;
+    text-shadow: 1px 1px 2px rgba(0,0,0,0.8);
+    white-space: nowrap;
+}
+
+/* ===== Inline Memory Config ===== */
+.inline-memory-config { margin-top: 12px; padding-top: 12px; border-top: 1px dashed var(--border); }
+.inline-memory-config.hidden { display: none; }
+.memory-input-row { display: flex; align-items: center; gap: 8px; margin-bottom: 8px; }
+.memory-input-row label { font-size: 0.85em; color: var(--text-dim); }
+.memory-input-row input {
+    background: var(--bg-input); border: 1px solid var(--border); color: var(--text);
+    padding: 6px 10px; border-radius: 4px; width: 100px;
+}
+.memory-unit { font-size: 0.85em; color: var(--text-dim); }
+
+/* ===== Parameter Items (show long flag) ===== */
+.param-item .param-flag-long {
+    font-family: monospace;
+    font-size: 0.8em;
+    color: var(--text-dim);
+    background: var(--bg);
+    padding: 2px 6px;
+    border-radius: 3px;
+    min-width: 80px;
+    text-align: center;
+    flex-shrink: 0;
+}
+
+/* ===== Command Hints ===== */
+.command-hint { margin-top: 8px; font-size: 0.85em; }
+.hint-error { color: var(--accent); padding: 6px 10px; background: rgba(233, 69, 96, 0.1); border-radius: 4px; margin-bottom: 4px; border-left: 3px solid var(--accent); }
+.hint-warning { color: #e9c46a; padding: 6px 10px; background: rgba(233, 196, 106, 0.1); border-radius: 4px; margin-bottom: 4px; border-left: 3px solid #e9c46a; }
+.hint-info { color: var(--text-dim); padding: 6px 10px; background: rgba(136, 146, 176, 0.1); border-radius: 4px; margin-bottom: 4px; border-left: 3px solid var(--text-dim); }
+.hint-ok { color: var(--accent2); padding: 6px 10px; background: rgba(78, 204, 163, 0.1); border-radius: 4px; margin-bottom: 4px; border-left: 3px solid var(--accent2); }
+
+/* ===== Natural Language Textarea ===== */
+.nl-input-container { display: flex; gap: 8px; align-items: flex-end; }
+.nl-input-container textarea {
+    flex: 1; background: var(--bg-input); border: 1px solid var(--border); color: var(--text);
+    padding: 10px 14px; border-radius: var(--radius); font-size: 0.95em; resize: vertical; font-family: inherit;
+    line-height: 1.5;
+}
+.nl-input-container textarea:focus { outline: none; border-color: var(--accent); }
+.nl-input-container button { padding: 10px 24px; background: var(--accent2); color: var(--bg); border: none; border-radius: var(--radius); cursor: pointer; font-weight: 600; flex-shrink: 0; }
+
 * { margin: 0; padding: 0; box-sizing: border-box; }
 
 body {
@@ -152,7 +248,7 @@ header h1 {
     box-shadow: var(--shadow);
 }
 
-.panel h2 { font-size: 1.2em; margin-bottom: 16px; color: var(--accent2); border-bottom: 1px solid var(--border); padding-bottom: 8px; }
+.panel h2 { font-size: 1.2em; margin-top: 0; margin-bottom: 16px; color: var(--accent2); border-bottom: 1px solid var(--border); padding-bottom: 8px; }
 
 .hidden { display: none !important; }
 
@@ -166,8 +262,9 @@ header h1 {
 .btn-add { background: transparent; border: 1px dashed var(--border); color: var(--text-dim); padding: 8px 16px; border-radius: var(--radius); cursor: pointer; width: 100%; margin-top: 8px; }
 .btn-add:hover { border-color: var(--accent2); color: var(--accent2); }
 
-/* ===== VRAM Display ===== */
-.vram-display { margin-top: 16px; padding: 12px; background: var(--bg-input); border-radius: var(--radius); }
+/* ===== VRAM Display (hidden when sticky bar visible) ===== */
+.vram-display { margin-top: 16px; padding: 12px; background: var(--bg-input); border-radius: var(--radius); display: none; }
+.vram-display.visible { display: block; }
 .vram-bar-container { position: relative; height: 30px; background: var(--bg); border-radius: 4px; overflow: hidden; border: 1px solid var(--border); }
 .vram-bar { height: 100%; background: linear-gradient(90deg, #4ecca3, #e9c46a); border-radius: 3px; transition: width 0.3s; width: 0%; }
 .vram-bar.warning { background: linear-gradient(90deg, #e9c46a, #f4a261); }
@@ -234,9 +331,6 @@ header h1 {
 .detail-item b { color: var(--accent2); }
 
 /* ===== Natural Language ===== */
-.nl-input-container { display: flex; gap: 8px; }
-.nl-input-container input { flex: 1; background: var(--bg-input); border: 1px solid var(--border); color: var(--text); padding: 10px 14px; border-radius: var(--radius); }
-.nl-input-container button { padding: 10px 24px; background: var(--accent2); color: var(--bg); border: none; border-radius: var(--radius); cursor: pointer; font-weight: 600; }
 .nl-result { margin-top: 10px; font-size: 0.9em; color: var(--text-dim); }
 .parsed-param { display: inline-block; margin: 2px 4px; padding: 2px 8px; background: var(--bg-input); border-radius: 4px; border: 1px solid var(--border); }
 .parsed-param .key { color: var(--accent2); }
@@ -246,6 +340,8 @@ header h1 {
 .param-tabs { display: flex; gap: 4px; margin-bottom: 12px; flex-wrap: wrap; }
 .tab-btn { padding: 6px 14px; background: var(--bg-input); border: 1px solid var(--border); color: var(--text-dim); border-radius: var(--radius); cursor: pointer; font-size: 0.85em; }
 .tab-btn.active { background: var(--accent); color: white; border-color: var(--accent); }
+.tab-btn.tab-modified { border-color: var(--accent2); color: var(--accent2); margin-left: auto; }
+.tab-btn.tab-modified.active { background: var(--accent2); color: var(--bg); border-color: var(--accent2); }
 
 /* ===== Param Search ===== */
 .param-search-container { margin-bottom: 12px; position: relative; }
@@ -282,7 +378,9 @@ header h1 {
     transition: border-color 0.15s;
 }
 .param-item:hover { border-color: var(--accent2); }
-.param-item.modified { border-color: var(--accent); background: rgba(233, 69, 96, 0.05); }
+.param-item.modified { border-color: var(--accent2); }
+
+/* keep old .modified for backwards compat but remove red styling */
 
 .param-item .param-flag {
     font-family: monospace;
@@ -341,12 +439,25 @@ header h1 {
 
 /* ===== Command Output ===== */
 .command-panel { position: sticky; bottom: 20px; }
-.command-output { background: #0d1117; border: 1px solid var(--border); border-radius: var(--radius); padding: 16px; font-family: 'Cascadia Code', 'Fira Code', monospace; font-size: 0.9em; color: #4ecca3; word-break: break-all; min-height: 60px; white-space: pre-wrap; }
-.command-actions { margin-top: 12px; display: flex; gap: 8px; }
-.btn-copy, .btn-gen { padding: 8px 20px; border: none; border-radius: var(--radius); cursor: pointer; font-size: 0.9em; font-weight: 500; }
-.btn-copy { background: var(--accent2); color: var(--bg); }
-.btn-copy:hover { opacity: 0.85; }
-.btn-gen { background: var(--bg-input); color: var(--text); border: 1px solid var(--border); }
+.command-output-wrap { position: relative; }
+.command-output { position: relative; background: #0d1117; border: 1px solid var(--border); border-radius: var(--radius); padding: 16px 16px 16px 70px; font-family: 'Cascadia Code', 'Fira Code', monospace; font-size: 0.9em; color: #4ecca3; word-break: break-all; min-height: 60px; white-space: pre-wrap; }
+.command-copy-btn {
+    position: absolute;
+    top: 6px;
+    left: 6px;
+    background: var(--bg-input);
+    border: 1px solid var(--border);
+    color: var(--text-dim);
+    padding: 3px 10px;
+    border-radius: 4px;
+    cursor: pointer;
+    font-size: 0.75em;
+    transition: all 0.15s;
+    opacity: 0.7;
+    z-index: 5;
+}
+.command-copy-btn:hover { opacity: 1; border-color: var(--accent2); color: var(--accent2); }
+.command-copy-btn.copied { background: var(--accent2); color: var(--bg); border-color: var(--accent2); opacity: 1; }
 
 /* ===== Admin ===== */
 .admin-tabs { display: flex; gap: 4px; margin-bottom: 20px; flex-wrap: wrap; }

static/index.html

@@ -12,6 +12,24 @@
         <h1>🦙 llama.cpp 命令生成器</h1>
     </header>
 
+    <!-- Sticky memory bar (always visible on top) -->
+    <div id="sticky-mem-bar" class="sticky-mem-bar">
+        <div class="sticky-vram-row" id="sticky-vram-row">
+            <span class="sticky-label">🎮 VRAM</span>
+            <div class="sticky-bar-container">
+                <div class="sticky-bar" id="sticky-vram-bar"></div>
+                <span class="sticky-bar-text" id="sticky-vram-text">等待配置...</span>
+            </div>
+        </div>
+        <div class="sticky-ram-row hidden" id="sticky-ram-row">
+            <span class="sticky-label">💾 内存</span>
+            <div class="sticky-bar-container">
+                <div class="sticky-bar ram-sticky-bar" id="sticky-ram-bar"></div>
+                <span class="sticky-bar-text" id="sticky-ram-text">等待配置...</span>
+            </div>
+        </div>
+    </div>
+
     <!-- Version & Mode Selection -->
     <section class="top-controls">
         <div class="control-group">
@@ -34,11 +52,13 @@
         </div>
     </section>
 
-    <!-- GPU Configuration -->
+    <!-- GPU Configuration + System Memory (merged) -->
     <section class="panel" id="gpu-panel">
         <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>
@@ -46,46 +66,38 @@
             </div>
             <div class="vram-breakdown" id="vram-breakdown"></div>
         </div>
-    </section>
 
-    <!-- System Memory (for GPU+CPU mode) -->
-    <section class="panel hidden" id="memory-panel">
-        <h2>💾 系统内存配置</h2>
-        <div class="control-group">
-            <label>系统内存上限</label>
-            <input type="number" id="sys-memory" value="0" min="0" step="1" onchange="updateEstimate()">
-            <select id="sys-memory-unit" onchange="updateEstimate()">
-                <option value="0">无限制</option>
-                <option value="1">GB</option>
-                <option value="2">MB</option>
-            </select>
-        </div>
-        <div class="vram-display" id="ram-display">
-            <div class="vram-bar-container">
-                <div class="vram-bar ram-bar" id="ram-bar"></div>
-                <div class="vram-label" id="ram-label">内存: 等待配置...</div>
+        <!-- System Memory (inline, only for GPU+CPU mode) -->
+        <div class="inline-memory-config hidden" id="inline-memory-config">
+            <div class="memory-input-row">
+                <label>系统内存上限</label>
+                <input type="number" id="sys-memory" value="0" min="0" step="1" onchange="updateEstimate()" placeholder="0 = 无限制">
+                <span class="memory-unit">GB</span>
+            </div>
+            <div class="vram-display" id="ram-display">
+                <div class="vram-bar-container">
+                    <div class="vram-bar ram-bar" id="ram-bar"></div>
+                    <div class="vram-label" id="ram-label">内存: 等待配置...</div>
+                </div>
+                <div class="vram-breakdown" id="ram-breakdown"></div>
             </div>
-            <div class="vram-breakdown" id="ram-breakdown"></div>
         </div>
     </section>
 
     <!-- Model Selection (two-step) -->
     <section class="panel">
         <h2>📦 选择模型</h2>
-        <!-- 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()">
+                <input type="text" id="model-search" placeholder="搜索模型名称..." oninput="filterBaseModels()" onfocus="onModelSearchFocus()" onblur="onModelSearchBlur()" autocomplete="off">
                 <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>
 
@@ -93,7 +105,7 @@
     <section class="panel">
         <h2>💬 自然语言生成</h2>
         <div class="nl-input-container">
-            <input type="text" id="nl-input" placeholder="例如: 用RTX 4090跑70B模型, 上下文8192, 温度0.7, flash attention" />
+            <textarea id="nl-input" rows="3" placeholder="加载中..."></textarea>
             <button onclick="parseNaturalLanguage()">解析</button>
         </div>
         <div id="nl-result" class="nl-result"></div>
@@ -114,6 +126,7 @@
             <button class="tab-btn" data-cat="lora" onclick="switchTab('lora')">LoRA</button>
             <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>
+            <button class="tab-btn tab-modified" data-cat="modified" onclick="switchTab('modified')">已改参数</button>
         </div>
         <div id="param-container"></div>
         <button class="btn-toggle-advanced" onclick="toggleHiddenParams()" id="toggle-advanced-btn">▼ 显示更多参数</button>
@@ -122,11 +135,11 @@
     <!-- Generated Command -->
     <section class="panel command-panel">
         <h2>📋 生成命令</h2>
-        <div class="command-output" id="command-output">请在上方配置参数...</div>
-        <div class="command-actions">
-            <button onclick="copyCommand()" class="btn-copy">📋 复制命令</button>
-            <button onclick="generateCommand()" class="btn-gen">🔄 重新生成</button>
+        <div class="command-output-wrap">
+            <button class="command-copy-btn" onclick="copyCommand()" id="copy-btn">📋 复制</button>
+            <div class="command-output" id="command-output">请在上方配置参数...</div>
         </div>
+        <div class="command-hint" id="command-hint"></div>
     </section>
 </div>
 <script src="/js/main.js"></script>

static/js/admin.js

@@ -241,14 +241,15 @@ function renderModelTable() {
             <td>${m.id}</td>
             <td><input type="text" value="${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(${m.id}, 'embd', this.value)" style="width:70px"></td>
-            <td><input type="number" value="${m.kv_heads}" onchange="updateModel(${m.id}, 'kv_heads', this.value)" style="width:60px"></td>
-            <td><input type="number" value="${m.head_dim}" onchange="updateModel(${m.id}, 'head_dim', this.value)" style="width:60px"></td>
-            <td><input type="number" value="${m.attention_heads}" onchange="updateModel(${m.id}, 'attention_heads', this.value)" style="width:60px"></td>
-            <td><input type="text" value="${m.quant || ''}" onchange="updateModel(${m.id}, 'quant', this.value)" style="width:80px"></td>
-            <td><input type="number" value="${m.sort_order}" onchange="updateModel(${m.id}, 'sort_order', this.value)" style="width:50px"></td>
+            <td><input type="number" value="${m.size_gb}" step="0.1" onchange="updateModel(${m.id}, 'size_gb', this.value)" style="width:60px"></td>
+            <td><input type="number" value="${m.layers}" onchange="updateModel(${m.id}, 'layers', this.value)" style="width:50px"></td>
+            <td><input type="number" value="${m.embd}" onchange="updateModel(${m.id}, 'embd', this.value)" style="width:60px"></td>
+            <td><input type="number" value="${m.kv_heads}" onchange="updateModel(${m.id}, 'kv_heads', this.value)" style="width:50px"></td>
+            <td><input type="number" value="${m.head_dim}" onchange="updateModel(${m.id}, 'head_dim', this.value)" style="width:50px"></td>
+            <td><input type="number" value="${m.attention_heads}" onchange="updateModel(${m.id}, 'attention_heads', this.value)" style="width:50px"></td>
+            <td><input type="number" value="${m.default_ctx || 0}" onchange="updateModel(${m.id}, 'default_ctx', this.value)" style="width:70px"></td>
+            <td><input type="text" value="${m.quant || ''}" onchange="updateModel(${m.id}, 'quant', this.value)" style="width:70px"></td>
+            <td><input type="number" value="${m.sort_order}" onchange="updateModel(${m.id}, 'sort_order', this.value)" style="width:40px"></td>
             <td><button class="btn-action btn-delete" onclick="deleteModel(${m.id})">删除</button></td>
         </tr>`).join('');
 }
@@ -263,22 +264,23 @@ async function addModel() {
         kv_heads: parseInt(document.getElementById('model-kv').value) || 0,
         head_dim: parseInt(document.getElementById('model-hdim').value) || 0,
         attention_heads: parseInt(document.getElementById('model-heads').value) || 0,
+        default_ctx: parseInt(document.getElementById('model-ctx').value) || 0,
         quant: document.getElementById('model-quant').value,
         description: document.getElementById('model-desc').value,
         sort_order: parseInt(document.getElementById('model-order').value) || 0,
     };
     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-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' : '');
+    ['model-basemodel','model-name','model-size','model-layers','model-embd','model-kv','model-hdim','model-heads','model-ctx','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 = { 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 };
+    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, default_ctx: m.default_ctx || 0, 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 if (['layers','embd','kv_heads','head_dim','attention_heads','default_ctx','sort_order'].includes(field)) data[field] = parseInt(value) || 0;
     else data[field] = value;
     await fetch(`/api/admin/models/${id}`, { method: 'PUT', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify(data) });
     await loadAdminModels();
@@ -305,6 +307,7 @@ function renderSettings() {
         'default_version': '默认版本',
         'default_mode': '默认模式',
         'default_quant': '默认量化版本',
+        'nl_default_text': '自然语言默认提示文本',
         'llm_enabled': '启用LLM解析 (true/false)',
         'llm_api_url': 'LLM API URL',
         'llm_api_key': 'LLM API Key',

static/js/main.js

@@ -6,6 +6,7 @@ let state = {
     mode: 'gpu', gpuSlots: [], showHidden: false,
     currentTab: 'common', paramSearchText: '',
     selectedModel: null, defaultQuant: 'Q4_K_M',
+    lastEstimate: null,
 };
 
 // ===== Init =====
@@ -13,6 +14,7 @@ async function init() {
     await loadVersions();
     await loadGpus();
     await loadModelsGrouped();
+    await loadNlDefaultText();
     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];
@@ -20,6 +22,15 @@ async function init() {
     renderParams(); generateCommand(); updateEstimate();
 }
 
+async function loadNlDefaultText() {
+    const res = await fetch('/api/settings/public');
+    const data = await res.json();
+    const ta = document.getElementById('nl-input');
+    if (data.nl_default_text) {
+        ta.placeholder = data.nl_default_text;
+    }
+}
+
 // ===== Load Data =====
 async function loadVersions() {
     const res = await fetch('/api/versions'); state.versions = await res.json();
@@ -34,7 +45,6 @@ async function loadModelsGrouped() {
     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();
 }
@@ -49,14 +59,14 @@ async function loadParams(versionId) {
     renderParams(); generateCommand(); updateEstimate();
 }
 
-// ===== Version Change =====
 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.getElementById('memory-panel').classList.toggle('hidden', mode !== 'gpu_cpu');
+    const showRam = mode === 'gpu_cpu';
+    document.getElementById('inline-memory-config').classList.toggle('hidden', !showRam);
+    document.getElementById('sticky-ram-row').classList.toggle('hidden', !showRam);
     generateCommand(); updateEstimate();
 }
 
@@ -73,49 +83,56 @@ function addGpuSlot() { if (state.gpuSlots.length >= 4) return; const dg = state
 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(); }
 
-// ===== Model Selection (two-step) =====
+// ===== Model Selection =====
 function filterBaseModels() {
     const text = document.getElementById('model-search').value.toLowerCase();
-    state.filteredBaseModels = state.baseModelList.filter(n => n.toLowerCase().includes(text));
+    if (!text) {
+        // Show top 5 when input is empty (on focus)
+        state.filteredBaseModels = state.baseModelList.slice(0, 5);
+    } else {
+        state.filteredBaseModels = state.baseModelList.filter(n => n.toLowerCase().includes(text));
+    }
     renderBaseModelDropdown();
     document.getElementById('model-dropdown').style.display = state.filteredBaseModels.length > 0 ? 'block' : 'none';
 }
-
+function onModelSearchFocus() {
+    if (!document.getElementById('model-search').value) {
+        state.filteredBaseModels = state.baseModelList.slice(0, 5);
+        renderBaseModelDropdown();
+        document.getElementById('model-dropdown').style.display = 'block';
+    }
+}
+function onModelSearchBlur() {
+    // Delay to allow click on option
+    setTimeout(() => { document.getElementById('model-dropdown').style.display = 'none'; }, 200);
+}
 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">${quantStr}</span></div>`;
+        const qs = quants.map(q => q.quant).filter(Boolean).join(', ');
+        return `<div class="model-option" onclick="selectBaseModel('${n}')"><span class="model-name">${n}</span><span class="model-meta">${qs}</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);
+    const dv = variants.find(v => v.quant === state.defaultQuant) || variants[0];
+    renderQuantOptions(baseName, variants, dv.id);
+    selectModel(dv.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('');
 }
-
 function selectModel(id) {
     const m = state.models.find(m => m.id === id);
     if (!m) return;
     state.selectedModel = m;
-    // Update quant buttons active state
-    const variants = state.modelsGrouped[m.base_model] || [];
-    renderQuantOptions(m.base_model, variants, id);
-    // Show model detail
+    renderQuantOptions(m.base_model, state.modelsGrouped[m.base_model] || [], id);
     document.getElementById('model-selected-info').innerHTML = `
         <div class="model-detail">
             <span class="detail-item"><b>模型:</b> ${m.name}</span>
@@ -126,9 +143,25 @@ function selectModel(id) {
             <span class="detail-item"><b>Head Dim:</b> ${m.head_dim}</span>
             <span class="detail-item"><b>量化:</b> ${m.quant || 'N/A'}</span>
         </div>`;
+    // Auto-adjust params based on selected model
+    applyModelDefaults(m);
     updateEstimate();
 }
 
+function applyModelDefaults(m) {
+    // Set ctx_size to model's default context if available
+    if (m.default_ctx && m.default_ctx > 0) {
+        state.paramValues['ctx_size'] = String(m.default_ctx);
+    }
+    // Set n_gpu_layers to 'all' for GPU mode
+    if (state.mode === 'gpu') {
+        state.paramValues['n_gpu_layers'] = 'all';
+    }
+    // Update param UI if currently visible
+    renderParams();
+    generateCommand();
+}
+
 // ===== Parameter Rendering =====
 function switchTab(cat) {
     state.currentTab = cat;
@@ -160,6 +193,10 @@ function onParamSearch() {
 }
 
 function getFilteredParams() {
+    // 'modified' is a special tab showing all modified params
+    if (state.currentTab === 'modified') {
+        return state.params.filter(p => isParamModified(p, state.paramValues[p.param_key]));
+    }
     let params = state.params.filter(p => p.category === state.currentTab);
     if (state.paramSearchText) {
         params = state.params.filter(p =>
@@ -175,7 +212,8 @@ function getFilteredParams() {
 function renderParams() {
     const container = document.getElementById('param-container');
     const all = getFilteredParams();
-    if (state.paramSearchText) {
+    // For 'modified' tab, show all (no important/hidden distinction)
+    if (state.currentTab === 'modified' || state.paramSearchText) {
         container.innerHTML = all.map(p => renderParamItem(p)).join('');
         document.getElementById('toggle-advanced-btn').style.display = 'none';
         return;
@@ -192,10 +230,17 @@ function renderParams() {
 
 function renderParamItem(p) {
     const val = state.paramValues[p.param_key];
-    const isMod = isParamModified(p, val);
-    const mc = isMod ? 'modified' : '';
+    const mc = '';
     const vb = p.affects_vram ? '<span class="vram-badge">⚡显存</span>' : '';
-    const flag = p.short_flag || p.long_flag;
+    // Show both short and long flag
+    const shortFlag = p.short_flag || '';
+    const longFlag = p.long_flag || '';
+    let flagHtml;
+    if (shortFlag && longFlag && shortFlag !== longFlag) {
+        flagHtml = `<span class="param-flag">${shortFlag}</span><span class="param-flag-long">${longFlag}</span>`;
+    } else {
+        flagHtml = `<span class="param-flag">${longFlag}</span>`;
+    }
     let inp = '';
     if (p.param_type === 'boolean') {
         inp = `<input type="checkbox" ${val ? 'checked' : ''} onchange="setParam('${p.param_key}', this.checked)">`;
@@ -211,7 +256,7 @@ function renderParamItem(p) {
         inp = `<input type="text" value="${val}" onchange="setParam('${p.param_key}', this.value)" data-key="${p.param_key}">`;
     }
     const uh = p.unit ? `<span class="param-unit">${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>`;
+    return `<div class="param-item ${mc}" data-key="${p.param_key}">${flagHtml}${vb}<span class="param-desc-text">${p.description}</span>${inp}${uh}</div>`;
 }
 
 function isParamModified(p, val) {
@@ -222,8 +267,8 @@ function isParamModified(p, val) {
 
 function setParam(key, value) {
     state.paramValues[key] = value;
-    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'); }
+    // If on 'modified' tab, re-render to update the list
+    if (state.currentTab === 'modified') renderParams();
     generateCommand(); updateEstimate();
 }
 
@@ -234,12 +279,16 @@ async function updateEstimate() {
     const params = collectParamsForEstimate();
     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;
+    // No unit select anymore - always GB, 0 = unlimited
+    const sysGb = sm > 0 ? sm : 0;
+    const isUnlimited = sm === 0;
     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();
+    state.lastEstimate = data;
     renderVramDisplay(data);
-    if (state.mode === 'gpu_cpu') renderRamDisplay(data, sysGb);
+    if (state.mode === 'gpu_cpu') renderRamDisplay(data, sysGb, isUnlimited);
+    updateStickyBar(data, sysGb, isUnlimited);
+    generateCommandHint(data, sysGb, isUnlimited);
 }
 
 function collectParamsForEstimate() {
@@ -256,20 +305,87 @@ function collectParamsForEstimate() {
 }
 
 function renderVramDisplay(data) {
-    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>`;
+    // Update sticky bar instead of inline display
+    updateStickyBar(data, parseFloat(document.getElementById('sys-memory').value) || 0, (parseFloat(document.getElementById('sys-memory').value) || 0) === 0);
 }
 
-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>`;
+function renderRamDisplay(data, sysGb, isUnlimited) {
+    // Update sticky ram bar instead of inline display
+    // (sticky bar is already updated in updateStickyBar)
+}
+
+// ===== Sticky top bar =====
+function updateStickyBar(data, sysGb, isUnlimited) {
+    const vr = document.getElementById('sticky-vram-row');
+    const rr = document.getElementById('sticky-ram-row');
+    // VRAM
+    if (data.total_vram_available_mb) {
+        const pct = data.usage_percent || 0;
+        const bar = document.getElementById('sticky-vram-bar');
+        bar.style.width = Math.min(pct, 100) + '%';
+        bar.className = 'sticky-bar' + (pct > 90 ? ' danger' : pct > 75 ? ' warning' : '');
+        document.getElementById('sticky-vram-text').textContent = `VRAM: ${data.total_gb}GB / ${data.total_vram_available_gb}GB (${pct}%)`;
+    } else {
+        document.getElementById('sticky-vram-text').textContent = 'VRAM: 请选择GPU';
+    }
+    // RAM
+    if (state.mode === 'gpu_cpu') {
+        rr.classList.remove('hidden');
+        if (isUnlimited) {
+            document.getElementById('sticky-ram-bar').style.width = '0%';
+            document.getElementById('sticky-ram-text').textContent = `内存: ${data.cpu_total_gb}GB (无限制)`;
+        } else {
+            const pct = data.cpu_usage_percent || 0;
+            const bar = document.getElementById('sticky-ram-bar');
+            bar.style.width = Math.min(pct, 100) + '%';
+            bar.className = 'sticky-bar ram-sticky-bar' + (pct > 90 ? ' danger' : pct > 75 ? ' warning' : '');
+            document.getElementById('sticky-ram-text').textContent = `内存: ${data.cpu_total_gb}GB / ${sysGb}GB (${pct}%)`;
+        }
+    } else {
+        rr.classList.add('hidden');
+    }
+}
+
+// ===== Command Hint =====
+function generateCommandHint(data, sysGb, isUnlimited) {
+    const hint = document.getElementById('command-hint');
+    let hints = [];
+
+    // Check VRAM
+    if (data.usage_percent && data.usage_percent > 100) {
+        hints.push({ type: 'error', text: `⚠️ 显存不足！预计需要 ${data.total_gb}GB，但仅有 ${data.total_vram_available_gb}GB。建议：减少 GPU 层数(n_gpu_layers)、减小上下文(ctx_size)、使用更低量化版本，或切换到 GPU+CPU 模式。` });
+    } else if (data.usage_percent && data.usage_percent > 90) {
+        hints.push({ type: 'warning', text: `⚡ 显存接近上限 (${data.usage_percent}%)，可能存在 OOM 风险。建议适当减小上下文或 GPU 层数。` });
+    } else if (data.usage_percent && data.usage_percent > 75) {
+        hints.push({ type: 'info', text: `ℹ️ 显存使用率 ${data.usage_percent}%，留有余量但不多。` });
+    } else if (state.selectedModel && data.usage_percent && data.usage_percent <= 75) {
+        hints.push({ type: 'ok', text: `✅ 显存充足，预计占用 ${data.usage_percent}%。` });
+    }
+
+    // Check RAM (GPU+CPU mode)
+    if (state.mode === 'gpu_cpu' && !isUnlimited) {
+        if (data.cpu_usage_percent && data.cpu_usage_percent > 100) {
+            hints.push({ type: 'error', text: `⚠️ 系统内存不足！预计需要 ${data.cpu_total_gb}GB，但上限仅 ${sysGb}GB。建议：增加内存上限、减少 GPU 层数让更多权重留在 GPU、或减小上下文。` });
+        } else if (data.cpu_usage_percent && data.cpu_usage_percent > 90) {
+            hints.push({ type: 'warning', text: `⚡ 内存使用率 ${data.cpu_usage_percent}%，接近上限。` });
+        }
+    }
+
+    // No model selected
+    if (!state.selectedModel) {
+        hints.push({ type: 'info', text: '💡 请在上方选择模型，以便进行准确的显存估算。' });
+    }
+
+    // No GPU selected
+    if (state.gpuSlots.length === 0 || !state.gpuSlots[0].name) {
+        hints.push({ type: 'info', text: '💡 请选择 GPU 显卡。' });
+    }
+
+    if (hints.length === 0) {
+        hints.push({ type: 'ok', text: '✅ 配置看起来没问题，可以复制使用。' });
+    }
+
+    hint.innerHTML = hints.map(h => `<div class="hint-${h.type}">${h.text}</div>`).join('');
 }
 
 // ===== Natural Language =====
@@ -304,7 +420,13 @@ function copyCommand() {
     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 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'); }
+    try {
+        document.execCommand('copy');
+        const btn = document.getElementById('copy-btn');
+        btn.textContent = '✅ 已复制';
+        btn.classList.add('copied');
+        setTimeout(() => { btn.textContent = '📋 复制'; btn.classList.remove('copied'); }, 2000);
+    } catch(e) { alert('复制失败，请手动选择文本复制'); }
     document.body.removeChild(ta);
 }