feat: add extended_kv_cache

src/components/seal-form/seal-cascader.tsx

@@ -1,3 +1,4 @@
+import IconFont from '@/components/icon-font';
 import { isNotEmptyValue } from '@/utils/index';
 import { useIntl } from '@umijs/max';
 import type { CascaderAutoProps } from 'antd';
@@ -33,10 +34,11 @@ const renderTag = (props: any) => {
 
 const OptionNodes = (props: {
   data: any;
+  notFoundContent?: React.ReactNode;
   optionNode: React.FC<{ data: any }>;
 }) => {
   const intl = useIntl();
-  const { data, optionNode: OptionNode } = props;
+  const { data, optionNode: OptionNode, notFoundContent } = props;
   if (data.value === '__EMPTY__') {
     return (
       <Empty
@@ -48,9 +50,12 @@ const OptionNodes = (props: {
           justifyContent: 'center',
           alignItems: 'center'
         }}
-        description={intl.formatMessage({
+        description={
-          id: 'common.search.empty'
+          notFoundContent ||
-        })}
+          intl.formatMessage({
+            id: 'common.search.empty'
+          })
+        }
       ></Empty>
     );
   }
@@ -85,6 +90,7 @@ const SealCascader: React.FC<
     allowNull,
     isInFormItems = true,
     optionNode,
+    notFoundContent,
     tagRender,
     ...rest
   } = props;
@@ -168,11 +174,13 @@ const SealCascader: React.FC<
       >
         <Cascader
           {...rest}
+          suffixIcon={<IconFont type="icon-down"></IconFont>}
           optionRender={
             optionNode
               ? (data) => (
                   <OptionNodes
                     data={data}
+                    notFoundContent={notFoundContent}
                     optionNode={optionNode}
                   ></OptionNodes>
                 )

src/components/seal-form/wrapper/index.tsx

@@ -158,6 +158,7 @@ const Inner = styled.div.attrs({
   className: '__inner__'
 })`
   width: 100%;
+  display: flex;
 `;
 
 const Extra = styled.div`

src/components/seal-form/wrapper/slider.ts

@@ -10,6 +10,9 @@ const SliderWrapper = styled.div`
   .borderless {
     background-color: transparent;
   }
+  .ant-slider {
+    flex: 1;
+  }
   padding-block: 0;
   padding-inline: 2px;
   input.ant-input-number-input {

src/locales/en-US/models.ts

@@ -182,5 +182,10 @@ export default {
   'models.table.users.all': 'All Users',
   'models.table.users.selected': 'Selected Users',
   'models.table.nouserFound': 'No users found',
-  'models.form.performance': 'Performance'
+  'models.form.performance': 'Performance',
+  'models.form.gpus.notfound': 'No GPUs found',
+  'models.form.extendedkvcache': 'Enable Extended KV Cache',
+  'models.form.chunkSize': 'Size Of Cache Chunks',
+  'models.form.maxCPUSize': 'Maximum CPU Cache Size (GiB)',
+  'models.form.remoteURL': 'Remote Storage URL'
 };

src/locales/ja-JP/models.ts

@@ -179,7 +179,12 @@ export default {
   'models.table.users.all': 'All Users',
   'models.table.users.selected': 'Selected Users',
   'models.table.nouserFound': 'No users found',
-  'models.form.performance': 'Performance'
+  'models.form.performance': 'Performance',
+  'models.form.gpus.notfound': 'No GPUs found',
+  'models.form.extendedkvcache': 'Enable Extended KV Cache',
+  'models.form.chunkSize': 'Size Of Cache Chunks',
+  'models.form.maxCPUSize': 'Maximum CPU Cache Size (GiB)',
+  'models.form.remoteURL': 'Remote Storage URL'
 };
 
 // ========== To-Do: Translate Keys (Remove After Translation) ==========
@@ -202,5 +207,11 @@ export default {
 // 17. 'models.table.users.all': 'All Users',
 // 18. 'models.table.users.selected': 'Selected Users',
 // 19. 'models.table.nouserFound': 'No users found',
-// 20. 'models.form.performance': 'Performance'
+// 20. 'models.form.performance': 'Performance',
+// 21. 'models.form.gpus.notfound': 'No GPUs found',
+// 22. 'models.form.extendedkvcache': 'Enable Extended KV Cache',
+// 23. 'models.form.chunkSize': 'Size Of Cache Chunks',
+// 24. 'models.form.maxCPUSize': 'Maximum CPU Cache Size (GiB)',
+// 25. 'models.form.remoteURL': 'Remote Storage URL'
+
 // ========== End of To-Do List ==========

src/locales/ru-RU/models.ts

@@ -182,7 +182,12 @@ export default {
   'models.table.users.all': 'All users',
   'models.table.users.selected': 'Selected users',
   'models.table.nouserFound': 'No users found',
-  'models.form.performance': 'Performance'
+  'models.form.performance': 'Performance',
+  'models.form.gpus.notfound': 'No GPUs found',
+  'models.form.extendedkvcache': 'Enable Extended KV Cache',
+  'models.form.chunkSize': 'Size Of Cache Chunks',
+  'models.form.maxCPUSize': 'Maximum CPU Cache Size (GiB)',
+  'models.form.remoteURL': 'Remote Storage URL'
 };
 
 // ========== To-Do: Translate Keys (Remove After Translation) ==========
@@ -197,5 +202,10 @@ export default {
 // 9. 'models.table.users.all': 'All users',
 // 10. 'models.table.users.selected': 'Selected users',
 // 11. 'models.table.nouserFound': 'No users found',
-// 12. 'models.form.performance': 'Performance'
+// 12. 'models.form.performance': 'Performance',
+// 13. 'models.form.gpus.notfound': 'No GPUs found',
+// 14. 'models.form.extendedkvcache': 'Enable Extended KV Cache',
+// 15. 'models.form.chunkSize': 'Size Of Cache Chunks',
+// 16. 'models.form.maxCPUSize': 'Maximum CPU Cache Size (GiB)',
+// 17. 'models.form.remoteURL': 'Remote Storage URL'
 // ========== End of To-Do List ==========

src/locales/zh-CN/models.ts

@@ -172,5 +172,10 @@ export default {
   'models.table.users.all': '所有用户',
   'models.table.users.selected': '已选用户',
   'models.table.nouserFound': '未找到用户',
-  'models.form.performance': '性能'
+  'models.form.performance': '性能',
+  'models.form.gpus.notfound': '未找到 GPU',
+  'models.form.extendedkvcache': '启用扩展 KV 缓存',
+  'models.form.chunkSize': '缓存分块大小',
+  'models.form.maxCPUSize': 'CPU 缓存最大占用 (GiB)',
+  'models.form.remoteURL': '远程存储地址'
 };

src/pages/llmodels/config/types.ts

@@ -76,6 +76,12 @@ export interface FormData {
   optimize_long_prompt: boolean;
   enable_speculative_decoding: boolean;
   cluster_id: number;
+  extended_kv_cache: {
+    enabled: boolean;
+    chunk_size: number;
+    max_local_cpu_size: number;
+    remote_url: string;
+  };
 }
 
 interface ComputedResourceClaim {

src/pages/llmodels/forms/advance-config.tsx

@@ -16,6 +16,7 @@ import { useFormContext } from '../config/form-context';
 import { FormData } from '../config/types';
 import Backend from '../forms/backend';
 import BackendParametersList from '../forms/backend-parameters-list';
+import ScheduleTypeForm from '../forms/schedule-type';
 
 const placementStrategyTips = [
   {
@@ -88,6 +89,7 @@ const AdvanceConfig = () => {
           options={modelCategories}
         ></SealSelect>
       </Form.Item>
+      <ScheduleTypeForm></ScheduleTypeForm>
       <Backend></Backend>
       {scheduleType === ScheduleValueMap.Auto && (
         <>
@@ -169,7 +171,7 @@ const AdvanceConfig = () => {
         [backendOptionsMap.vllm, backendOptionsMap.ascendMindie].includes(
           backend
         ) && (
-          <div style={{ paddingBottom: 22, paddingLeft: 10 }}>
+          <div style={{ paddingBottom: 22 }}>
             <Form.Item<FormData>
               name="distributed_inference_across_workers"
               valuePropName="checked"
@@ -187,7 +189,7 @@ const AdvanceConfig = () => {
             </Form.Item>
           </div>
         )}
-      <div style={{ paddingBottom: 22, paddingLeft: 10 }}>
+      <div style={{ paddingBottom: 22 }}>
         <Form.Item<FormData>
           name="restart_on_error"
           valuePropName="checked"

src/pages/llmodels/forms/index.tsx

@@ -191,6 +191,12 @@ const DataForm: React.FC<DataFormProps> = forwardRef((props, ref) => {
           categories: null,
           restart_on_error: true,
           distributed_inference_across_workers: true,
+          extended_kv_cache: {
+            enabled: false,
+            chunk_size: 256,
+            max_local_cpu_size: 10,
+            remote_url: ''
+          },
           ...initialValues
         }}
       >
@@ -254,6 +260,28 @@ const DataForm: React.FC<DataFormProps> = forwardRef((props, ref) => {
             ></SealSelect>
           }
         </Form.Item>
+        <Form.Item<FormData>
+          name="replicas"
+          rules={[
+            {
+              required: true,
+              message: getRuleMessage('input', 'models.form.replicas')
+            }
+          ]}
+        >
+          <SealInput.Number
+            style={{ width: '100%' }}
+            label={intl.formatMessage({
+              id: 'models.form.replicas'
+            })}
+            required
+            description={intl.formatMessage(
+              { id: 'models.form.replicas.tips' },
+              { api: `${window.location.origin}/v1` }
+            )}
+            min={0}
+          ></SealInput.Number>
+        </Form.Item>
         <CatalogFrom></CatalogFrom>
         <Form.Item<FormData> name="description">
           <SealInput.TextArea

src/pages/llmodels/forms/kv-cache.tsx

@@ -0,0 +1,60 @@
+import CheckboxField from '@/components/seal-form/checkbox-field';
+import SealInputNumber from '@/components/seal-form/input-number';
+import SealInput from '@/components/seal-form/seal-input';
+import { useIntl } from '@umijs/max';
+import { Form } from 'antd';
+import { FormData } from '../config/types';
+
+const KVCacheForm = () => {
+  const intl = useIntl();
+  const form = Form.useFormInstance();
+  const kvCacheEnabled = Form.useWatch(['extended_kv_cache', 'enabled'], form);
+
+  return (
+    <>
+      <div style={{ paddingBottom: 22 }}>
+        <Form.Item<FormData>
+          name={['extended_kv_cache', 'enabled']}
+          valuePropName="checked"
+          style={{ padding: '0 10px', marginBottom: 0 }}
+          noStyle
+        >
+          <CheckboxField
+            label={intl.formatMessage({ id: 'models.form.extendedkvcache' })}
+          ></CheckboxField>
+        </Form.Item>
+      </div>
+      {kvCacheEnabled && (
+        <>
+          <Form.Item<FormData>
+            name={['extended_kv_cache', 'max_local_cpu_size']}
+          >
+            <SealInputNumber
+              label={intl.formatMessage({ id: 'models.form.maxCPUSize' })}
+              min={0}
+              step={1}
+              precision={0}
+            />
+          </Form.Item>
+          <Form.Item<FormData> name={['extended_kv_cache', 'chunk_size']}>
+            <SealInputNumber
+              label={intl.formatMessage({ id: 'models.form.chunkSize' })}
+              min={0}
+              step={1}
+            />
+          </Form.Item>
+          <Form.Item<FormData> name={['extended_kv_cache', 'remote_url']}>
+            <SealInput.Input
+              label={intl.formatMessage({ id: 'models.form.remoteURL' })}
+              min={0}
+              step={1}
+              placeholder="protocol://host:port"
+            />
+          </Form.Item>
+        </>
+      )}
+    </>
+  );
+};
+
+export default KVCacheForm;

src/pages/llmodels/forms/performance.tsx

@@ -1,143 +1,15 @@
-import SealCascader from '@/components/seal-form/seal-cascader';
-import SealInput from '@/components/seal-form/seal-input';
-import SealSelect from '@/components/seal-form/seal-select';
-import TooltipList from '@/components/tooltip-list';
-import useAppUtils from '@/hooks/use-app-utils';
 import { useIntl } from '@umijs/max';
 import { Form } from 'antd';
 import React from 'react';
-import GPUCard from '../components/gpu-card';
+import KVCacheForm from './kv-cache';
-import { scheduleList, ScheduleValueMap } from '../config';
-import { backendOptionsMap } from '../config/backend-parameters';
-import { useCatalogFormContext, useFormContext } from '../config/form-context';
-import { FormData } from '../config/types';
-
-const scheduleTypeTips = [
-  {
-    title: {
-      text: 'models.form.scheduletype.auto',
-      locale: true
-    },
-    tips: 'models.form.scheduletype.auto.tips'
-  },
-  {
-    title: {
-      text: 'models.form.scheduletype.manual',
-      locale: true
-    },
-    tips: 'models.form.scheduletype.manual.tips'
-  }
-];
 
 const Performance: React.FC = () => {
   const intl = useIntl();
-  const { onValuesChange, gpuOptions } = useFormContext();
-  const { onQuantizationChange } = useCatalogFormContext();
-  const { getRuleMessage } = useAppUtils();
   const form = Form.useFormInstance();
-  const scheduleType = Form.useWatch('scheduleType', form);
-
-  const handleScheduleTypeChange = (value: string) => {
-    if (value === ScheduleValueMap.Auto) {
-      onValuesChange?.({}, form.getFieldsValue());
-    }
-  };
-
-  const handleOnQuantizationChange = (val: any) => {
-    onQuantizationChange?.(val);
-  };
-
-  const handleBeforeGpuSelectorChange = (gpuIds: any[]) => {};
-
-  const handleGpuSelectorChange = (value: any[]) => {
-    handleBeforeGpuSelectorChange(value);
-    onValuesChange?.({}, form.getFieldsValue());
-  };
 
   return (
     <>
-      <Form.Item<FormData>
+      <KVCacheForm></KVCacheForm>
-        name="replicas"
-        rules={[
-          {
-            required: true,
-            message: getRuleMessage('input', 'models.form.replicas')
-          }
-        ]}
-      >
-        <SealInput.Number
-          style={{ width: '100%' }}
-          label={intl.formatMessage({
-            id: 'models.form.replicas'
-          })}
-          required
-          description={intl.formatMessage(
-            { id: 'models.form.replicas.tips' },
-            { api: `${window.location.origin}/v1` }
-          )}
-          min={0}
-        ></SealInput.Number>
-      </Form.Item>
-      <Form.Item name="scheduleType">
-        <SealSelect
-          onChange={handleScheduleTypeChange}
-          label={intl.formatMessage({ id: 'models.form.scheduletype' })}
-          description={<TooltipList list={scheduleTypeTips}></TooltipList>}
-          options={scheduleList}
-        ></SealSelect>
-      </Form.Item>
-      {scheduleType === ScheduleValueMap.SpecificGPUType && (
-        <>
-          <Form.Item name={['gpu_selector', 'gpu_type']}>
-            <SealSelect
-              label={intl.formatMessage({ id: 'models.form.gpuType' })}
-              options={[]}
-            ></SealSelect>
-          </Form.Item>
-          <Form.Item name={['gpu_selector', 'gpu_count']}>
-            <SealSelect
-              label={intl.formatMessage({ id: 'models.form.gpuCount' })}
-              options={[]}
-            ></SealSelect>
-          </Form.Item>
-        </>
-      )}
-      {scheduleType === ScheduleValueMap.Manual &&
-        !form.getFieldValue('fix_gpu_type') && (
-          <>
-            <Form.Item
-              name={['gpu_selector', 'gpu_ids']}
-              rules={[
-                {
-                  required: true,
-                  message: getRuleMessage('select', 'models.form.gpuselector')
-                }
-              ]}
-            >
-              <SealCascader
-                required
-                showSearch
-                expandTrigger="hover"
-                multiple={
-                  form.getFieldValue('backend') !== backendOptionsMap.voxBox
-                }
-                classNames={{
-                  popup: {
-                    root: 'cascader-popup-wrapper gpu-selector'
-                  }
-                }}
-                maxTagCount={1}
-                label={intl.formatMessage({ id: 'models.form.gpuselector' })}
-                options={gpuOptions}
-                showCheckedStrategy="SHOW_CHILD"
-                value={form.getFieldValue(['gpu_selector', 'gpu_ids'])}
-                optionNode={GPUCard}
-                getPopupContainer={(triggerNode) => triggerNode.parentNode}
-                onChange={handleGpuSelectorChange}
-              ></SealCascader>
-            </Form.Item>
-          </>
-        )}
       {/* <div style={{ paddingBottom: 22, paddingLeft: 10 }}>
         <Form.Item<FormData>
           name="optimize_long_prompt"

src/pages/llmodels/forms/schedule-type.tsx

@@ -0,0 +1,124 @@
+import SealCascader from '@/components/seal-form/seal-cascader';
+import SealSelect from '@/components/seal-form/seal-select';
+import TooltipList from '@/components/tooltip-list';
+import useAppUtils from '@/hooks/use-app-utils';
+import { useIntl } from '@umijs/max';
+import { Form } from 'antd';
+import React from 'react';
+import GPUCard from '../components/gpu-card';
+import { scheduleList, ScheduleValueMap } from '../config';
+import { backendOptionsMap } from '../config/backend-parameters';
+import { useCatalogFormContext, useFormContext } from '../config/form-context';
+
+const scheduleTypeTips = [
+  {
+    title: {
+      text: 'models.form.scheduletype.auto',
+      locale: true
+    },
+    tips: 'models.form.scheduletype.auto.tips'
+  },
+  {
+    title: {
+      text: 'models.form.scheduletype.manual',
+      locale: true
+    },
+    tips: 'models.form.scheduletype.manual.tips'
+  }
+];
+
+const ScheduleTypeForm: React.FC = () => {
+  const intl = useIntl();
+  const { onValuesChange, gpuOptions } = useFormContext();
+  const { onQuantizationChange } = useCatalogFormContext();
+  const { getRuleMessage } = useAppUtils();
+  const form = Form.useFormInstance();
+  const scheduleType = Form.useWatch('scheduleType', form);
+
+  const handleScheduleTypeChange = (value: string) => {
+    if (value === ScheduleValueMap.Auto) {
+      onValuesChange?.({}, form.getFieldsValue());
+    }
+  };
+
+  const handleOnQuantizationChange = (val: any) => {
+    onQuantizationChange?.(val);
+  };
+
+  const handleBeforeGpuSelectorChange = (gpuIds: any[]) => {};
+
+  const handleGpuSelectorChange = (value: any[]) => {
+    handleBeforeGpuSelectorChange(value);
+    onValuesChange?.({}, form.getFieldsValue());
+  };
+
+  return (
+    <>
+      <Form.Item name="scheduleType">
+        <SealSelect
+          onChange={handleScheduleTypeChange}
+          label={intl.formatMessage({ id: 'models.form.scheduletype' })}
+          description={<TooltipList list={scheduleTypeTips}></TooltipList>}
+          options={scheduleList}
+        ></SealSelect>
+      </Form.Item>
+      {scheduleType === ScheduleValueMap.SpecificGPUType && (
+        <>
+          <Form.Item name={['gpu_selector', 'gpu_type']}>
+            <SealSelect
+              label={intl.formatMessage({ id: 'models.form.gpuType' })}
+              options={[]}
+            ></SealSelect>
+          </Form.Item>
+          <Form.Item name={['gpu_selector', 'gpu_count']}>
+            <SealSelect
+              label={intl.formatMessage({ id: 'models.form.gpuCount' })}
+              options={[]}
+            ></SealSelect>
+          </Form.Item>
+        </>
+      )}
+      {scheduleType === ScheduleValueMap.Manual &&
+        !form.getFieldValue('fix_gpu_type') && (
+          <>
+            <Form.Item
+              name={['gpu_selector', 'gpu_ids']}
+              rules={[
+                {
+                  required: true,
+                  message: getRuleMessage('select', 'models.form.gpuselector')
+                }
+              ]}
+            >
+              <SealCascader
+                required
+                showSearch
+                expandTrigger="hover"
+                multiple={
+                  form.getFieldValue('backend') !== backendOptionsMap.voxBox
+                }
+                classNames={{
+                  popup: {
+                    root: 'cascader-popup-wrapper gpu-selector'
+                  }
+                }}
+                maxTagCount={1}
+                label={intl.formatMessage({ id: 'models.form.gpuselector' })}
+                options={gpuOptions}
+                showCheckedStrategy="SHOW_CHILD"
+                value={form.getFieldValue(['gpu_selector', 'gpu_ids'])}
+                optionNode={GPUCard}
+                getPopupContainer={(triggerNode) => triggerNode.parentNode}
+                notFoundContent={intl.formatMessage({
+                  id: 'models.form.gpus.notfound'
+                })}
+                onChange={handleGpuSelectorChange}
+              ></SealCascader>
+            </Form.Item>
+          </>
+        )}
+    </>
+  );
+};
+
+export default ScheduleTypeForm;