feat: 调优创建支持模型量化配置

新增训练时量化（QLoRA 4/8bit）与训练后导出量化模型开关，支持 bitsandbytes、GPTQ、AWQ、GGUF 四种方法及对应位数、分组、导出格式配置，命令预览与提交参数同步接入，常量补充三类量化选项。

frontend/src/constants/index.ts

@@ -76,6 +76,29 @@ export const LR_SCHEDULER_OPTIONS = [
   { label: 'constant', value: 'constant' },
 ]
 
+/** 训练时量化位数（QLoRA） */
+export const QUANTIZATION_BIT_OPTIONS = [
+  { label: '不量化', value: 0 },
+  { label: '8 bit', value: 8 },
+  { label: '4 bit (QLoRA)', value: 4 },
+]
+
+/** 训练后导出量化方法 */
+export const QUANT_METHOD_OPTIONS = [
+  { label: 'bitsandbytes', value: 'bnb' },
+  { label: 'GPTQ', value: 'gptq' },
+  { label: 'AWQ', value: 'awq' },
+  { label: 'GGUF (llama.cpp)', value: 'gguf' },
+]
+
+/** GGUF 量化格式 */
+export const GGUF_FORMAT_OPTIONS = [
+  { label: 'Q4_K_M（推荐）', value: 'Q4_K_M' },
+  { label: 'Q5_K_M', value: 'Q5_K_M' },
+  { label: 'Q8_0', value: 'Q8_0' },
+  { label: 'F16（不量化）', value: 'F16' },
+]
+
 /** 训练模板分组（按模型系列） */
 export const TEMPLATE_GROUPS: { label: string; options: { label: string; value: string }[] }[] = [
   {

frontend/src/views/fine-tune/FineTuneCreateView.vue

@@ -13,7 +13,7 @@ import {
 import { getModelList } from '@/api/modules/model'
 import { getDatasetList } from '@/api/modules/dataset'
 import { getSystemInfo } from '@/api/modules/system'
-import { TEMPLATE_GROUPS, LR_SCHEDULER_OPTIONS } from '@/constants'
+import { TEMPLATE_GROUPS, LR_SCHEDULER_OPTIONS, QUANTIZATION_BIT_OPTIONS, QUANT_METHOD_OPTIONS, GGUF_FORMAT_OPTIONS } from '@/constants'
 import type { ModelItem, DatasetItem, GpuInfo } from '@/types'
 
 const router = useRouter()
@@ -48,6 +48,13 @@ const form = reactive({
   lora_alpha: 16, // 修复原项目 lora_alpha 默认值不一致 bug
   lora_dropout: 0.1,
   lora_rank: 8,
+  // 量化参数
+  quantization_bit: 0, // 训练时量化（QLoRA）：0=不量化
+  export_quantized: false, // 训练后是否导出量化模型
+  quant_method: 'bnb', // 导出量化方法
+  quant_bits: 4, // 导出量化位数
+  quant_group_size: 128, // 分组大小（GPTQ/AWQ）
+  export_format: 'Q4_K_M', // GGUF 导出格式
 })
 
 const rules: FormRules = {
@@ -95,6 +102,19 @@ const commandPreview = computed(() => {
     cmd += ` \\\n  --lora_dropout ${form.lora_dropout}`
     cmd += ` \\\n  --lora_rank ${form.lora_rank}`
   }
+  if (showLoraParams.value && form.quantization_bit) {
+    cmd += ` \\\n  --quantization_bit ${form.quantization_bit}`
+  }
+  if (form.export_quantized) {
+    const method = form.quant_method
+    const bits = form.quant_bits
+    cmd += ` \\\n  # 训练后导出量化模型：${method} ${bits}bit`
+    if (method === 'gguf') {
+      cmd += ` \\\n  # export_format=${form.export_format}`
+    } else if (method === 'gptq' || method === 'awq') {
+      cmd += ` \\\n  # group_size=${form.quant_group_size}`
+    }
+  }
   return cmd
 })
 
@@ -132,6 +152,12 @@ function resetParams() {
     lora_alpha: 16,
     lora_dropout: 0.1,
     lora_rank: 8,
+    quantization_bit: 0,
+    export_quantized: false,
+    quant_method: 'bnb',
+    quant_bits: 4,
+    quant_group_size: 128,
+    export_format: 'Q4_K_M',
   })
 }
 
@@ -230,6 +256,12 @@ async function handleSubmit() {
         lora_alpha: form.lora_alpha,
         lora_dropout: form.lora_dropout,
         lora_rank: form.lora_rank,
+        quantization_bit: form.train_method === 'lora' ? form.quantization_bit : 0,
+        export_quantized: form.export_quantized,
+        quant_method: form.export_quantized ? form.quant_method : '',
+        quant_bits: form.export_quantized ? form.quant_bits : 0,
+        quant_group_size: form.export_quantized ? form.quant_group_size : 0,
+        export_format: form.export_quantized && form.quant_method === 'gguf' ? form.export_format : '',
         status: 'pending',
         progress: 0,
       }
@@ -260,6 +292,12 @@ async function handleSubmit() {
           lora_alpha: form.lora_alpha,
           lora_dropout: form.lora_dropout,
           lora_rank: form.lora_rank,
+          quantization_bit: form.train_method === 'lora' ? form.quantization_bit : 0,
+          export_quantized: form.export_quantized,
+          quant_method: form.export_quantized ? form.quant_method : '',
+          quant_bits: form.export_quantized ? form.quant_bits : 0,
+          quant_group_size: form.export_quantized ? form.quant_group_size : 0,
+          export_format: form.export_quantized && form.quant_method === 'gguf' ? form.export_format : '',
         })
         ElMessage.success('训练任务已创建并启动')
       } catch (e) {
@@ -437,6 +475,74 @@ onMounted(() => {
           </el-form-item>
         </template>
 
+        <!-- 模型量化 -->
+        <el-divider content-position="left">模型量化</el-divider>
+        <!-- 训练时量化（QLoRA）：仅 LoRA 训练时可用 -->
+        <el-form-item v-if="form.train_method === 'lora'" label="训练时量化">
+          <el-select v-model="form.quantization_bit" style="width: 420px;">
+            <el-option
+              v-for="opt in QUANTIZATION_BIT_OPTIONS"
+              :key="opt.value"
+              :label="opt.label"
+              :value="opt.value"
+            />
+          </el-select>
+        </el-form-item>
+
+        <!-- 训练后导出量化模型 -->
+        <el-form-item label="导出量化模型">
+          <el-select v-model="form.export_quantized" style="width: 420px;">
+            <el-option label="否" :value="false" />
+            <el-option label="是" :value="true" />
+          </el-select>
+        </el-form-item>
+
+        <template v-if="form.export_quantized">
+          <el-form-item label="量化方法">
+            <el-select v-model="form.quant_method" style="width: 420px;">
+              <el-option
+                v-for="opt in QUANT_METHOD_OPTIONS"
+                :key="opt.value"
+                :label="opt.label"
+                :value="opt.value"
+              />
+            </el-select>
+          </el-form-item>
+
+          <el-form-item label="量化位数">
+            <el-input-number
+              v-model="form.quant_bits"
+              :min="form.quant_method === 'bnb' ? 4 : 2"
+              :max="form.quant_method === 'bnb' ? 8 : 16"
+              :step="1"
+              controls-position="right"
+              style="width: 200px"
+            />
+          </el-form-item>
+
+          <el-form-item v-if="form.quant_method === 'gptq' || form.quant_method === 'awq'" label="分组大小">
+            <el-input-number
+              v-model="form.quant_group_size"
+              :min="32"
+              :max="1024"
+              :step="32"
+              controls-position="right"
+              style="width: 200px"
+            />
+          </el-form-item>
+
+          <el-form-item v-if="form.quant_method === 'gguf'" label="导出格式">
+            <el-select v-model="form.export_format" style="width: 420px;">
+              <el-option
+                v-for="opt in GGUF_FORMAT_OPTIONS"
+                :key="opt.value"
+                :label="opt.label"
+                :value="opt.value"
+              />
+            </el-select>
+          </el-form-item>
+        </template>
+
         <!-- 训练命令预览 -->
         <el-divider content-position="left">训练命令预览</el-divider>
         <div class="command-preview-wrapper">
@@ -606,6 +712,13 @@ onMounted(() => {
   margin-bottom: 22px;
 }
 
+.field-tip {
+  color: #909399;
+  font-size: 12px;
+  line-height: 20px;
+  margin-top: 4px;
+}
+
 .form-actions-wrapper {
   position: fixed;
   bottom: 0;