From 8cb94f6da65576b06a04e6ed9d14832da6b5fa99 Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Thu, 16 Apr 2026 03:49:22 +0000
Subject: [PATCH 01/13] feat(ascend): add ResourceCoreName to support hami-core
 compute partitioning

Signed-off-by: ashergaga <1214443299@qq.com>
---
 .../templates/scheduler/device-configmap.yaml |   2 +
 charts/hami/values.yaml                       |   4 +
 pkg/device/ascend/device.go                   | 123 ++++++++++++++----
 pkg/device/ascend/vnpu.go                     |   1 +
 4 files changed, 107 insertions(+), 23 deletions(-)

diff --git a/charts/hami/templates/scheduler/device-configmap.yaml b/charts/hami/templates/scheduler/device-configmap.yaml
index aaab6b664b..23db1d21ac 100644
--- a/charts/hami/templates/scheduler/device-configmap.yaml
+++ b/charts/hami/templates/scheduler/device-configmap.yaml
@@ -333,6 +333,7 @@ data:
       commonWord: Ascend910B3
       resourceName: huawei.com/Ascend910B3
       resourceMemoryName: huawei.com/Ascend910B3-memory
+      resourceCoreName: huawei.com/Ascend910B3-core
       memoryAllocatable: 65536
       memoryCapacity: 65536
       memoryFactor: 1
@@ -413,6 +414,7 @@ data:
       commonWord: Ascend910C
       resourceName: huawei.com/Ascend910C
       resourceMemoryName: huawei.com/Ascend910C-memory
+      resourceCoreName: huawei.com/Ascend910C-core
       memoryAllocatable: 65536
       memoryCapacity: 65536
       aiCore: 20
diff --git a/charts/hami/values.yaml b/charts/hami/values.yaml
index 4f821ac2ee..35c97a9da8 100644
--- a/charts/hami/values.yaml
+++ b/charts/hami/values.yaml
@@ -486,12 +486,16 @@ devices:
       - huawei.com/Ascend910B2-memory
       - huawei.com/Ascend910B3
       - huawei.com/Ascend910B3-memory
+      - huawei.com/Ascend910B3-core
       - huawei.com/Ascend910B4
       - huawei.com/Ascend910B4-memory
       - huawei.com/Ascend910B4-1
       - huawei.com/Ascend910B4-1-memory
       - huawei.com/Ascend310P
       - huawei.com/Ascend310P-memory
+      - huawei.com/Ascend910C
+      - huawei.com/Ascend910C-memory
+      - huawei.com/Ascend910C-core
   iluvatar:
     enabled: false
     customresources:
diff --git a/pkg/device/ascend/device.go b/pkg/device/ascend/device.go
index 5ac32325cd..021f912deb 100644
--- a/pkg/device/ascend/device.go
+++ b/pkg/device/ascend/device.go
@@ -41,6 +41,8 @@ const (
 	Ascend910Prefix        = "Ascend910"
 	Ascend910CType         = "Ascend910C"
 	Ascend910NetworkWeight = 10
+	VNPUModeAnnotation     = "huawei.com/vnpu-mode"
+    VNPUModeHamiCore       = "hami-core"
 )
 
 type Devices struct {
@@ -52,8 +54,10 @@ type Devices struct {
 }
 
 type RuntimeInfo struct {
-	UUID string `json:"UUID,omitempty"`
-	Temp string `json:"temp,omitempty"`
+	UUID     string `json:"UUID,omitempty"`
+	Temp     string `json:"temp,omitempty"`
+	Memory   int64  `json:"memory,omitempty"` 
+	Core     int32  `json:"core,omitempty"` 
 }
 
 var (
@@ -111,11 +115,11 @@ func (dev *Devices) CommonWord() string {
 }
 
 func (dev *Devices) MutateAdmission(ctr *corev1.Container, p *corev1.Pod) (bool, error) {
-	count, ok := ctr.Resources.Limits[corev1.ResourceName(dev.config.ResourceName)]
+    count, ok := ctr.Resources.Limits[corev1.ResourceName(dev.config.ResourceName)]
 	if !ok {
 		return false, nil
 	}
-
+ 
 	reqNum := count.Value()
 	if dev.config.CommonWord == Ascend910CType {
 		if reqNum == 1 {
@@ -134,12 +138,41 @@ func (dev *Devices) MutateAdmission(ctr *corev1.Container, p *corev1.Pod) (bool,
 		}
 	}
 
+	// Check if hami-core is declared
+	vnpuMode := p.Annotations[VNPUModeAnnotation]
+    isHAMiCore := (vnpuMode == VNPUModeHamiCore)
+
+	if isHAMiCore {
+        klog.V(3).Infof("Ascend core resource detected, injecting postStart lifecycle for container %s", ctr.Name)
+        
+        if ctr.Lifecycle == nil {
+            ctr.Lifecycle = &corev1.Lifecycle{}
+        }
+        
+        // Inject PostStart hook to start the limiter process
+        if ctr.Lifecycle.PostStart == nil {
+            ctr.Lifecycle.PostStart = &corev1.LifecycleHandler{
+                Exec: &corev1.ExecAction{
+                    Command: []string{
+                        "bash",
+                        "-c",
+                        "export RUST_LOG=info\n/hami-vnpu-core/target/debug/limiter > /tmp/inst1_manager.log 2>&1 &",
+                    },
+                },
+            }
+        }
+    }
+ 
 	trimMem := dev.config.MemoryAllocatable
 	memory, ok := ctr.Resources.Limits[corev1.ResourceName(dev.config.ResourceMemoryName)]
 	if ok {
-		trimMem, _ = dev.trimMemory(memory.Value())
-		if trimMem <= 0 {
-			return false, fmt.Errorf("%s %d is invalid", dev.config.ResourceMemoryName, memory.Value())
+		if isHAMiCore {
+			trimMem = memory.Value()
+		} else {
+			trimMem, _ = dev.trimMemory(memory.Value())
+			if trimMem <= 0 {
+				return false, fmt.Errorf("%s %d is invalid", dev.config.ResourceMemoryName, memory.Value())
+			}
 		}
 	}
 	if count.Value() > 1 {
@@ -179,15 +212,27 @@ func (dev *Devices) PatchAnnotations(pod *corev1.Pod, annoInput *map[string]stri
 		(*annoInput)[device.InRequestDevices[commonWord]] = device.EncodePodSingleDevice(devList)
 		(*annoInput)[device.SupportDevices[commonWord]] = device.EncodePodSingleDevice(devList)
 		(*annoInput)["predicate-time"] = strconv.FormatInt(time.Now().Unix(), 10)
+
+		// Check if hami-core is declared
+		vnpuMode := pod.Annotations[VNPUModeAnnotation]
+		klog.V(4).Infof("Patching pod %s with vnpu mode: %s", pod.Name, vnpuMode)
+
 		allocateStr := fmt.Sprintf("huawei.com/%s", dev.CommonWord())
 		var rtInfo []RuntimeInfo
 		for _, dp := range devList {
 			for _, val := range dp {
-				_, temp := dev.trimMemory(int64(val.Usedmem))
-				rtInfo = append(rtInfo, RuntimeInfo{
-					UUID: val.UUID,
-					Temp: temp,
-				})
+				info := RuntimeInfo{UUID: val.UUID}
+
+				// If is hami core, populate Memory and Core directly without using Temp
+				if vnpuMode == VNPUModeHamiCore {
+					info.Memory = int64(val.Usedmem)
+					info.Core = val.Usedcores
+				} else {
+					_, temp := dev.trimMemory(int64(val.Usedmem))
+					info.Temp = temp
+				}
+	
+				rtInfo = append(rtInfo, info)
 			}
 		}
 		s, err := json.Marshal(rtInfo)
@@ -247,6 +292,8 @@ func (dev *Devices) CheckHealth(devType string, n *corev1.Node) (bool, bool) {
 func (dev *Devices) GenerateResourceRequests(ctr *corev1.Container) device.ContainerDeviceRequest {
 	ascendResourceCount := corev1.ResourceName(dev.config.ResourceName)
 	ascendResourceMem := corev1.ResourceName(dev.config.ResourceMemoryName)
+	ascendResourceCore := corev1.ResourceName(dev.config.ResourceCoreName)
+
 	v, ok := ctr.Resources.Limits[ascendResourceCount]
 	if !ok {
 		v, ok = ctr.Resources.Requests[ascendResourceCount]
@@ -261,18 +308,35 @@ func (dev *Devices) GenerateResourceRequests(ctr *corev1.Container) device.Conta
 				mem, ok = ctr.Resources.Requests[ascendResourceMem]
 			}
 			if ok {
-				memnums, ok := mem.AsInt64()
-				if ok {
-					if dev.config.MemoryFactor > 1 {
-						rawMemnums := memnums
-						memnums = memnums * int64(dev.config.MemoryFactor)
-						klog.V(4).Infof("Update Ascend memory request. before %d, after %d, factor %d", rawMemnums, memnums, dev.config.MemoryFactor)
+				memnums, _ := mem.AsInt64()
+				
+				// If "core" is requested, it explicitly indicates the use of soft-partitioning.
+				isCoreRequested := false
+				if ascendResourceCore != "" {
+					_, isCoreRequested = ctr.Resources.Limits[ascendResourceCore]
+					if !isCoreRequested {
+						_, isCoreRequested = ctr.Resources.Requests[ascendResourceCore]
 					}
+				}
+
+				if isCoreRequested {
+					// Soft-partitioning: Use the raw value directly.
+					memnum = int(memnums)
+				} else {
 					m, _ := dev.trimMemory(memnums)
 					memnum = int(m)
 				}
 			}
+			
+			// Process Core Resources
 			corenum := int32(0)
+			if ascendResourceCore != "" {
+				if cv, ok := ctr.Resources.Limits[ascendResourceCore]; ok {
+					corenum = int32(cv.Value())
+				} else if cv, ok := ctr.Resources.Requests[ascendResourceCore]; ok {
+					corenum = int32(cv.Value())
+				}
+			}
 
 			mempnum := 0
 			if memnum == 0 {
@@ -340,7 +404,7 @@ func (dev *Devices) GetResourceNames() device.ResourceNames {
 	return device.ResourceNames{
 		ResourceCountName:  dev.config.ResourceName,
 		ResourceMemoryName: dev.config.ResourceMemoryName,
-		ResourceCoreName:   "",
+		ResourceCoreName:   dev.config.ResourceCoreName,
 	}
 }
 
@@ -349,6 +413,13 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 	originReq := k.Nums
 	prevnuma := -1
 	klog.InfoS("Allocating device for container request", "pod", klog.KObj(pod), "card request", k)
+	
+	vnpuMode := ""
+	if pod != nil && pod.Annotations != nil {
+		vnpuMode = pod.Annotations[VNPUModeAnnotation]
+	}
+	klog.V(4).InfoS("Fit: vnpu-mode annotation", "pod", pod.Name, "vnpuMode", vnpuMode)
+
 	var tmpDevs map[string]device.ContainerDevices
 	tmpDevs = make(map[string]device.ContainerDevices)
 	reason := make(map[string]int)
@@ -405,19 +476,25 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 			klog.V(5).InfoS(common.CardInsufficientMemory, "pod", klog.KObj(pod), "device", dev.ID, "device index", i, "device total memory", dev.Totalmem, "device used memory", dev.Usedmem, "request memory", memreq)
 			continue
 		}
-		if dev.Totalcore-dev.Usedcores < k.Coresreq {
+		// Set dev.Totalcore to 100 if vnpuMode is hami-core
+		effectiveTotalCore := dev.Totalcore
+		if vnpuMode == VNPUModeHamiCore {
+			effectiveTotalCore = 100
+		}
+
+		if effectiveTotalCore-dev.Usedcores < k.Coresreq {
 			reason[common.CardInsufficientCore]++
-			klog.V(5).InfoS(common.CardInsufficientCore, "pod", klog.KObj(pod), "device", dev.ID, "device index", i, "device total core", dev.Totalcore, "device used core", dev.Usedcores, "request cores", k.Coresreq)
+			klog.V(5).InfoS(common.CardInsufficientCore, "pod", klog.KObj(pod), "device", dev.ID, "device index", i, "device total core", effectiveTotalCore, "device used core", dev.Usedcores, "request cores", k.Coresreq)
 			continue
 		}
 		// Coresreq=100 indicates it want this card exclusively
-		if dev.Totalcore == 100 && k.Coresreq == 100 && dev.Used > 0 {
+		if effectiveTotalCore == 100 && k.Coresreq == 100 && dev.Used > 0 {
 			reason[common.ExclusiveDeviceAllocateConflict]++
 			klog.V(5).InfoS(common.ExclusiveDeviceAllocateConflict, "pod", klog.KObj(pod), "device", dev.ID, "device index", i, "used", dev.Used)
 			continue
 		}
 		// You can't allocate core=0 job to an already full GPU
-		if dev.Totalcore != 0 && dev.Usedcores == dev.Totalcore && k.Coresreq == 0 {
+		if effectiveTotalCore != 0 && dev.Usedcores == effectiveTotalCore && k.Coresreq == 0 {
 			reason[common.CardComputeUnitsExhausted]++
 			klog.V(5).InfoS(common.CardComputeUnitsExhausted, "pod", klog.KObj(pod), "device", dev.ID, "device index", i)
 			continue
diff --git a/pkg/device/ascend/vnpu.go b/pkg/device/ascend/vnpu.go
index 26783e376d..4c558780dc 100644
--- a/pkg/device/ascend/vnpu.go
+++ b/pkg/device/ascend/vnpu.go
@@ -28,6 +28,7 @@ type VNPUConfig struct {
 	ChipName           string     `yaml:"chipName"`
 	ResourceName       string     `yaml:"resourceName"`
 	ResourceMemoryName string     `yaml:"resourceMemoryName"`
+	ResourceCoreName   string     `yaml:"resourceCoreName"`
 	MemoryAllocatable  int64      `yaml:"memoryAllocatable"`
 	MemoryCapacity     int64      `yaml:"memoryCapacity"`
 	MemoryFactor       int32      `yaml:"memoryFactor"`

From 8072c6e1ff453dd8ccf8c3faffe82672e5263136 Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Fri, 17 Apr 2026 08:06:16 +0000
Subject: [PATCH 02/13] feat: add docs ascend910-hami-vnpu-support

Signed-off-by: ashergaga <1214443299@qq.com>
---
 docs/ascend910-hami-vnpu-core-support.md    | 104 +++++++++++++++++++
 docs/ascend910-hami-vnpu-core-support_cn.md | 108 ++++++++++++++++++++
 2 files changed, 212 insertions(+)
 create mode 100644 docs/ascend910-hami-vnpu-core-support.md
 create mode 100644 docs/ascend910-hami-vnpu-core-support_cn.md

diff --git a/docs/ascend910-hami-vnpu-core-support.md b/docs/ascend910-hami-vnpu-core-support.md
new file mode 100644
index 0000000000..4b7755fe9f
--- /dev/null
+++ b/docs/ascend910-hami-vnpu-core-support.md
@@ -0,0 +1,104 @@
+# Introduction to Ascend910B Series and Ascend910C Support
+
+Implements  a soft slicing mechanism based on `libvnpu.so` interception and `limiter` token scheduling, enabling fine-grained resource sharing.  For detailed information, check [hami-vnpu-core](https://github.com/Project-HAMi/hami-vnpu-core).
+
+* **_NPU sharing_**: Each task can allocate a portion of Ascend NPU instead of a whole NLU card, thus NPU can be shared among multiple tasks.
+* **_Device Memory Control_**: Ascend NPUs can be allocated with certain device memory size and guarantee it that it does not exceed the boundary.
+* **_Device Core Control_**: Ascend NPUs can be allocated with certain compute cores and guarantee it that it does not exceed the boundary.
+
+## Prerequisites
+
+* Ascend docker runtime
+* driver version > 25.5
+* Ascend device type: 910B3、910C
+
+## Enabling NPU Sharing
+
+* [Ascend docker runtime](https://gitcode.com/Ascend/mind-cluster/tree/master/component/ascend-docker-runtime)
+
+* [ascend-device-plugin](https://github.com/Project-HAMi/ascend-device-plugin)
+* **Chip Mode**: enable `device-share` mode on Ascend chips for virtualization
+
+### Host Environment Preparation
+
+Before launching any containers, the **Global Shared Memory (SHM) Region** must be initialized on the host to allow inter-Pod coordination.
+
+#### 1. Create the Shared Directory
+
+```bash
+sudo mkdir -p /tmp/hami-shared-region
+sudo chmod 777 /tmp/hami-shared-region
+```
+
+#### 2. Deploy hami-vnpu-core Components
+
+Place the following files in a fixed host path (`/usr/local/hami-vnpu-core/`) for mounting into containers:
+
+```
+/usr/local/hami-vnpu-core/
+├── limiter              # Manager daemon binary (compiled from hami-vnpu-core)
+├── libvnpu.so           # Interception library for LD_PRELOAD
+└── ld.so.preload        # Global preload config 
+```
+
+  ### Create device-config.yaml
+
+The content is as follows:
+
+  ```yaml
+vnpus:
+- chipName: 910B3
+  commonWord: Ascend910B3
+  resourceName: huawei.com/Ascend910B3
+  resourceMemoryName: huawei.com/Ascend910B3-memory
+  resourceCoreName: huawei.com/Ascend910B3-core
+  memoryAllocatable: 65536
+  memoryCapacity: 65536
+  aiCore: 20
+  aiCPU: 7
+  templates:
+    - name: vir05_1c_16g
+      memory: 16384
+      aiCore: 5
+      aiCPU: 1
+    - name: vir10_3c_32g
+      memory: 32768
+      aiCore: 10
+      aiCPU: 3
+- chipName: Ascend910
+      commonWord: Ascend910C
+      resourceName: huawei.com/Ascend910C
+      resourceMemoryName: huawei.com/Ascend910C-memory
+      resourceCoreName: huawei.com/Ascend910C-core
+      memoryAllocatable: 65536
+      memoryCapacity: 65536
+      aiCore: 20
+      aiCPU: 7
+  ```
+
+  ### Install and update with Helm
+
+  Helm installation and updates will be based on the configuration in this file, overwriting the built-in configuration of Helm.
+
+## Running NPU Workloads
+
+You can request Ascend 910B series resources using the `huawei.com/ascend910Bx` ,`huawei.com/ascend910Bx-memory`  and   `huawei.com/ascend910Bx-core` resource types:
+
+```yaml
+apiVersion: v1
+kind: Pod
+metadata:
+  name: ascend-soft-slice-pod
+  annotations:
+    huawei.com/vnpu-mode: 'hami-core' # Enables hami-vnpu-core soft-segmentation for this pod
+spec:
+  containers:
+    - name: npu_pod
+      ...
+      resources:
+        limits:
+          huawei.com/Ascend910B3: "1"           # Request 1 physical NPU
+          huawei.com/Ascend910B3-memory: "28672"     # Request 28Gi memory
+          huawei.com/Ascend910B3-core: "40"     # Request 40% core
+```
+
diff --git a/docs/ascend910-hami-vnpu-core-support_cn.md b/docs/ascend910-hami-vnpu-core-support_cn.md
new file mode 100644
index 0000000000..3e867d0f14
--- /dev/null
+++ b/docs/ascend910-hami-vnpu-core-support_cn.md
@@ -0,0 +1,108 @@
+# huawei.com/Ascend910B、huawei.com/Ascend910C系列软切分支持简介
+
+实现了基于 `libvnpu.so` 拦截和limiter令牌调度的软切分机制，能够实现精细化的资源共享。详细信息请参阅 [hami-vnpu-core](https://www.google.com/search?q=链接)。
+
+* **_NPU 共享_**: 每个任务可以只占用一部分显卡，多个任务可以共享一张显卡；
+
+* **_可限制分配的显存大小_**: 你可以用显存值（例如 3000M）来分配 NPU，本组件会确保任务使用的显存不会超过分配数值；
+
+* **_可限制分配的算力大小_**: 你可以用固定数量来分配 NPU 的 AI 核心和 AI CPU 核心，本组件会确保任务使用的算力不会超过分配数值。
+
+## 节点需求
+
+* Ascend docker runtime
+* driver version > 25.5
+* Ascend device type: 910B3、910C
+
+## 开启 NPU 复用
+
+* 部署 [Ascend docker runtime](https://gitcode.com/Ascend/mind-cluster/tree/master/component/ascend-docker-runtime)
+
+* 部署 [ascend-device-plugin](https://github.com/Project-HAMi/ascend-device-plugin)
+* 芯片模式：在昇腾芯片上开启 `device-share` 模式以支持虚拟化。
+
+## 宿主机环境准备
+
+在启动任何容器之前，必须在宿主机上初始化 **全局共享内存 (SHM) 区域**，以便进行 Pod 间的协同。
+
+1. **创建共享目录**
+
+   ```
+   sudo mkdir -p /tmp/hami-shared-region
+   sudo chmod 777 /tmp/hami-shared-region
+   ```
+
+2. **部署 hami-vnpu-core 组件** 
+
+   将以下文件放置在固定的宿主机路径（`/usr/local/hami-vnpu-core/`）中，以便挂载到容器内： 
+
+   ```
+   /usr/local/hami-vnpu-core/
+   ├── limiter              # Manager daemon binary (compiled from hami-vnpu-core)
+   ├── libvnpu.so           # Interception library for LD_PRELOAD
+   └── ld.so.preload        # Global preload config 
+   ```
+
+  ### 在 files 目录下创建 device-config.yaml
+
+  配置文件如下所示，可以按需调整：
+
+  ```yaml
+  vnpus:
+  - chipName: 910B3
+    commonWord: Ascend910B3
+    resourceName: huawei.com/Ascend910B3
+    resourceMemoryName: huawei.com/Ascend910B3-memory
+    resourceCoreName: huawei.com/Ascend910B3-core
+    memoryAllocatable: 65536
+    memoryCapacity: 65536
+    aiCore: 20
+    aiCPU: 7
+    templates:
+      - name: vir05_1c_16g
+        memory: 16384
+        aiCore: 5
+        aiCPU: 1
+      - name: vir10_3c_32g
+        memory: 32768
+        aiCore: 10
+        aiCPU: 3
+  - chipName: Ascend910
+        commonWord: Ascend910C
+        resourceName: huawei.com/Ascend910C
+        resourceMemoryName: huawei.com/Ascend910C-memory
+        resourceCoreName: huawei.com/Ascend910C-core
+        memoryAllocatable: 65536
+        memoryCapacity: 65536
+        aiCore: 20
+        aiCPU: 7
+  ```
+
+  ### Helm 安装和更新
+
+  Helm 安装、更新将基于该配置文件，覆盖默认的配置文件
+
+</details>
+
+## 运行 NPU 任务
+
+可通过使用 `huawei.com/ascend910Bx` 、 `huawei.com/ascend910Bx-memory` 和 `huawei.com/ascend910Bx-core`资源类型，来请求 Ascend 910B 系列设备：
+
+```yaml
+apiVersion: v1
+kind: Pod
+metadata:
+  name: ascend-soft-slice-pod
+  annotations:
+    huawei.com/vnpu-mode: 'hami-core' # 添加该注解的走hami-vnpu-core软切分
+spec:
+  containers:
+    - name: npu_pod
+      ...
+      resources:
+        limits:
+          huawei.com/Ascend910B3: "1"          # 请求 1 块物理 NPU
+          huawei.com/Ascend910B3-memory: "28672" # 请求 28Gi 显存
+          huawei.com/Ascend910B3-core: "40"      # 请求 40% 的算力
+```
+

From 14d132d5e492dd5eb871721fba33ff4e0b5a186d Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Fri, 17 Apr 2026 08:32:15 +0000
Subject: [PATCH 03/13] feat: fix ai suggestion

Signed-off-by: ashergaga <1214443299@qq.com>
---
 pkg/device/ascend/device.go | 39 ++++++++++++++++++++++---------------
 1 file changed, 23 insertions(+), 16 deletions(-)

diff --git a/pkg/device/ascend/device.go b/pkg/device/ascend/device.go
index 021f912deb..7bcf78149b 100644
--- a/pkg/device/ascend/device.go
+++ b/pkg/device/ascend/device.go
@@ -42,7 +42,7 @@ const (
 	Ascend910CType         = "Ascend910C"
 	Ascend910NetworkWeight = 10
 	VNPUModeAnnotation     = "huawei.com/vnpu-mode"
-    VNPUModeHamiCore       = "hami-core"
+	VNPUModeHamiCore       = "hami-core"
 )
 
 type Devices struct {
@@ -115,7 +115,7 @@ func (dev *Devices) CommonWord() string {
 }
 
 func (dev *Devices) MutateAdmission(ctr *corev1.Container, p *corev1.Pod) (bool, error) {
-    count, ok := ctr.Resources.Limits[corev1.ResourceName(dev.config.ResourceName)]
+	count, ok := ctr.Resources.Limits[corev1.ResourceName(dev.config.ResourceName)]
 	if !ok {
 		return false, nil
 	}
@@ -156,7 +156,7 @@ func (dev *Devices) MutateAdmission(ctr *corev1.Container, p *corev1.Pod) (bool,
                     Command: []string{
                         "bash",
                         "-c",
-                        "export RUST_LOG=info\n/hami-vnpu-core/target/debug/limiter > /tmp/inst1_manager.log 2>&1 &",
+                        "export RUST_LOG=info\n/hami-vnpu-core/limiter > /tmp/limiter_manager.log 2>&1 &",
                     },
                 },
             }
@@ -309,21 +309,28 @@ func (dev *Devices) GenerateResourceRequests(ctr *corev1.Container) device.Conta
 			}
 			if ok {
 				memnums, _ := mem.AsInt64()
-				
-				// If "core" is requested, it explicitly indicates the use of soft-partitioning.
-				isCoreRequested := false
-				if ascendResourceCore != "" {
-					_, isCoreRequested = ctr.Resources.Limits[ascendResourceCore]
-					if !isCoreRequested {
-						_, isCoreRequested = ctr.Resources.Requests[ascendResourceCore]
+				if ok {
+					if dev.config.MemoryFactor > 1 {
+						rawMemnums := memnums
+						memnums = memnums * int64(dev.config.MemoryFactor)
+						klog.V(4).Infof("Update Ascend memory request. before %d, after %d, factor %d", rawMemnums, memnums, dev.config.MemoryFactor)
+					}
+					// If "core" is requested, it explicitly indicates the use of soft-partitioning.
+					isCoreRequested := false
+					if ascendResourceCore != "" {
+						_, isCoreRequested = ctr.Resources.Limits[ascendResourceCore]
+						if !isCoreRequested {
+							_, isCoreRequested = ctr.Resources.Requests[ascendResourceCore]
+						}
 					}
-				}
 
-				if isCoreRequested {
-					// Soft-partitioning: Use the raw value directly.
-					memnum = int(memnums)
-				} else {
-					m, _ := dev.trimMemory(memnums)
+					if isCoreRequested {
+						// Soft-partitioning: Use the raw value directly.
+						memnum = int(memnums)
+					} else {
+						m, _ := dev.trimMemory(memnums)
+						memnum = int(m)
+					}
 					memnum = int(m)
 				}
 			}

From 0663b976eb70150b171f04afa29eab770c061bec Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Mon, 20 Apr 2026 06:13:17 +0000
Subject: [PATCH 04/13] feat: fix golangci lint

Signed-off-by: ashergaga <1214443299@qq.com>
---
 pkg/device/ascend/device.go | 1 -
 1 file changed, 1 deletion(-)

diff --git a/pkg/device/ascend/device.go b/pkg/device/ascend/device.go
index 7bcf78149b..61373f8e78 100644
--- a/pkg/device/ascend/device.go
+++ b/pkg/device/ascend/device.go
@@ -331,7 +331,6 @@ func (dev *Devices) GenerateResourceRequests(ctr *corev1.Container) device.Conta
 						m, _ := dev.trimMemory(memnums)
 						memnum = int(m)
 					}
-					memnum = int(m)
 				}
 			}
 			

From b1cce71960edd154512160f82f1d7e8444dc6e56 Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Mon, 20 Apr 2026 06:35:25 +0000
Subject: [PATCH 05/13] feat: fix format

Signed-off-by: ashergaga <1214443299@qq.com>
---
 pkg/device/ascend/device.go | 60 ++++++++++++++++++-------------------
 1 file changed, 30 insertions(+), 30 deletions(-)

diff --git a/pkg/device/ascend/device.go b/pkg/device/ascend/device.go
index 61373f8e78..9e33010586 100644
--- a/pkg/device/ascend/device.go
+++ b/pkg/device/ascend/device.go
@@ -54,10 +54,10 @@ type Devices struct {
 }
 
 type RuntimeInfo struct {
-	UUID     string `json:"UUID,omitempty"`
-	Temp     string `json:"temp,omitempty"`
-	Memory   int64  `json:"memory,omitempty"` 
-	Core     int32  `json:"core,omitempty"` 
+	UUID   string `json:"UUID,omitempty"`
+	Temp   string `json:"temp,omitempty"`
+	Memory int64  `json:"memory,omitempty"`
+	Core   int32  `json:"core,omitempty"`
 }
 
 var (
@@ -119,7 +119,7 @@ func (dev *Devices) MutateAdmission(ctr *corev1.Container, p *corev1.Pod) (bool,
 	if !ok {
 		return false, nil
 	}
- 
+
 	reqNum := count.Value()
 	if dev.config.CommonWord == Ascend910CType {
 		if reqNum == 1 {
@@ -140,29 +140,29 @@ func (dev *Devices) MutateAdmission(ctr *corev1.Container, p *corev1.Pod) (bool,
 
 	// Check if hami-core is declared
 	vnpuMode := p.Annotations[VNPUModeAnnotation]
-    isHAMiCore := (vnpuMode == VNPUModeHamiCore)
+	isHAMiCore := (vnpuMode == VNPUModeHamiCore)
 
 	if isHAMiCore {
-        klog.V(3).Infof("Ascend core resource detected, injecting postStart lifecycle for container %s", ctr.Name)
-        
-        if ctr.Lifecycle == nil {
-            ctr.Lifecycle = &corev1.Lifecycle{}
-        }
-        
-        // Inject PostStart hook to start the limiter process
-        if ctr.Lifecycle.PostStart == nil {
-            ctr.Lifecycle.PostStart = &corev1.LifecycleHandler{
-                Exec: &corev1.ExecAction{
-                    Command: []string{
-                        "bash",
-                        "-c",
-                        "export RUST_LOG=info\n/hami-vnpu-core/limiter > /tmp/limiter_manager.log 2>&1 &",
-                    },
-                },
-            }
-        }
-    }
- 
+		klog.V(3).Infof("Ascend core resource detected, injecting postStart lifecycle for container %s", ctr.Name)
+
+		if ctr.Lifecycle == nil {
+			ctr.Lifecycle = &corev1.Lifecycle{}
+		}
+
+		// Inject PostStart hook to start the limiter process
+		if ctr.Lifecycle.PostStart == nil {
+			ctr.Lifecycle.PostStart = &corev1.LifecycleHandler{
+				Exec: &corev1.ExecAction{
+					Command: []string{
+						"bash",
+						"-c",
+						"export RUST_LOG=info\n/hami-vnpu-core/limiter > /tmp/limiter_manager.log 2>&1 &",
+					},
+				},
+			}
+		}
+	}
+
 	trimMem := dev.config.MemoryAllocatable
 	memory, ok := ctr.Resources.Limits[corev1.ResourceName(dev.config.ResourceMemoryName)]
 	if ok {
@@ -231,7 +231,7 @@ func (dev *Devices) PatchAnnotations(pod *corev1.Pod, annoInput *map[string]stri
 					_, temp := dev.trimMemory(int64(val.Usedmem))
 					info.Temp = temp
 				}
-	
+
 				rtInfo = append(rtInfo, info)
 			}
 		}
@@ -308,7 +308,7 @@ func (dev *Devices) GenerateResourceRequests(ctr *corev1.Container) device.Conta
 				mem, ok = ctr.Resources.Requests[ascendResourceMem]
 			}
 			if ok {
-				memnums, _ := mem.AsInt64()
+				memnums, ok := mem.AsInt64()
 				if ok {
 					if dev.config.MemoryFactor > 1 {
 						rawMemnums := memnums
@@ -333,7 +333,7 @@ func (dev *Devices) GenerateResourceRequests(ctr *corev1.Container) device.Conta
 					}
 				}
 			}
-			
+
 			// Process Core Resources
 			corenum := int32(0)
 			if ascendResourceCore != "" {
@@ -419,7 +419,7 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 	originReq := k.Nums
 	prevnuma := -1
 	klog.InfoS("Allocating device for container request", "pod", klog.KObj(pod), "card request", k)
-	
+
 	vnpuMode := ""
 	if pod != nil && pod.Annotations != nil {
 		vnpuMode = pod.Annotations[VNPUModeAnnotation]

From b8f8ff8db7d7c775fe39af0efd4420b82ef1115b Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Mon, 20 Apr 2026 08:00:21 +0000
Subject: [PATCH 06/13] feat: add unit test & proposal

Signed-off-by: ashergaga <1214443299@qq.com>
---
 docs/ascend910-hami-vnpu-core-support.md      |   4 +-
 docs/ascend910-hami-vnpu-core-support_cn.md   |   4 +-
 ...ng_hami_vnpu_core_for_virtualization_cn.md | 174 +++++++++++++++++
 pkg/device/ascend/device_test.go              | 184 ++++++++++++++++++
 4 files changed, 362 insertions(+), 4 deletions(-)
 create mode 100644 docs/proposals/integrating_hami_vnpu_core_for_virtualization_cn.md

diff --git a/docs/ascend910-hami-vnpu-core-support.md b/docs/ascend910-hami-vnpu-core-support.md
index 4b7755fe9f..7ce3ebe250 100644
--- a/docs/ascend910-hami-vnpu-core-support.md
+++ b/docs/ascend910-hami-vnpu-core-support.md
@@ -26,8 +26,8 @@ Before launching any containers, the **Global Shared Memory (SHM) Region** must
 #### 1. Create the Shared Directory
 
 ```bash
-sudo mkdir -p /tmp/hami-shared-region
-sudo chmod 777 /tmp/hami-shared-region
+sudo mkdir -p /usr/local/hami-shared-region
+sudo chmod 777 /usr/local/hami-shared-region
 ```
 
 #### 2. Deploy hami-vnpu-core Components
diff --git a/docs/ascend910-hami-vnpu-core-support_cn.md b/docs/ascend910-hami-vnpu-core-support_cn.md
index 3e867d0f14..435cdd407a 100644
--- a/docs/ascend910-hami-vnpu-core-support_cn.md
+++ b/docs/ascend910-hami-vnpu-core-support_cn.md
@@ -28,8 +28,8 @@
 1. **创建共享目录**
 
    ```
-   sudo mkdir -p /tmp/hami-shared-region
-   sudo chmod 777 /tmp/hami-shared-region
+   sudo mkdir -p /usr/local/hami-shared-region
+   sudo chmod 777 /usr/local/hami-shared-region
    ```
 
 2. **部署 hami-vnpu-core 组件** 
diff --git a/docs/proposals/integrating_hami_vnpu_core_for_virtualization_cn.md b/docs/proposals/integrating_hami_vnpu_core_for_virtualization_cn.md
new file mode 100644
index 0000000000..4432fecbb8
--- /dev/null
+++ b/docs/proposals/integrating_hami_vnpu_core_for_virtualization_cn.md
@@ -0,0 +1,174 @@
+HAMi-vnpu-core 是一个Rust语言编写的昇腾 NPU 容器内资源控制器，通过 libvnpu.so（拦截器） + Limiter（管理器）实现用户态拦截，通过NPU_MEM_QUOTA 和 NPU_PRIORITY 两个环境变量分别声明显存限制和调度优先级。本方案在 HAMi 调度中集成该能力，以支持昇腾 NPU 的显存虚拟化与算力时间片软切分。
+
+## 一、前置条件
+
+昇腾驱动版本25.5以上，芯片开启 device-share 模式:
+
+#### 功能说明
+
+**npu-smi set -t device-share -i** *id* **-d** value 用于设置指定设备的所有芯片的容器共享模式。
+
+#### 参数说明
+
+| 类型    | 描述                                                        |
+| ------- | ----------------------------------------------------------- |
+| *id*    | 设备ID。通过**npu-smi info -l**命令查出的NPU ID即为设备ID。 |
+| *value* | 容器使能状态：分为禁用、使能。默认禁用。0：禁用 1：使能     |
+
+## 二、HAMi-scheduler改动点
+
+### 2\.1 扩展资源名
+
+复用 huawei.com/Ascend910B3-memory ：显存分配逻辑可复用原逻辑；
+
+新增 huawei.com/Ascend910B3-core：声明了该资源名的请求走vnpu软切分，否则走原硬切分逻辑。
+| 资源名称 | 单位 | 含义 | 示例 |
+|--------------|----------|----------|----------|
+| huawei.com/Ascend910B3 | 整数 | NPU卡数 | 1 |
+| huawei.com/Ascend910B3-memory | Mi | 显存配额 | 28672 (28Gi) |
+| huawei.com/Ascend910B3-core | 整数 | 百分比 | 20, 40 |
+
+
+
+### 2\.2 Filter阶段
+
+**Filter阶段，**修改 **Fit** 函数核心逻辑，确保单张卡总算力不超过 100；同时修改 PatchAnnotation 注入的注解。
+
+**Pod配额预期分配结果格式（Pod Annotation）:**
+
+```shell {"data-theme":"githubLight"}
+{
+  "huawei.com/Ascend910B3": "[
+    {
+      "UUID":"xxx",
+      "memory": 28672,   
+      "core": 20,
+    }
+  ]"
+}
+```
+
+**PatchAnnotation**判断逻辑 vnpu软切分新增显存（memory）和核心（core）字段
+
+```go {"data-title":"pkg/device/ascend/device.go","data-theme":"githubLight"}
+func (dev *Devices) PatchAnnotations(pod *corev1.Pod, annoInput *map[string]string, pd device.PodDevices) map[string]string {
+	commonWord := dev.CommonWord()
+	devList, ok := pd[commonWord]
+	if ok && len(devList) > 0 {
+		...
+		for _, dp := range devList {
+			for _, val := range dp {
+              ...
+    
+				// 解析 Pod 的memQuota、 priority 资源声明 构建 RuntimeInfo
+				rtInfo = append(rtInfo, RuntimeInfo{
+					UUID:     val.UUID,
+					Temp:     tempName,        
+					MemQuota: memory,        
+					Priority: core,        
+				})
+			}
+		}
+       ...
+	}
+	return *annoInput
+}
+```
+
+### 2\.3 容器limiter进程启动
+
+通过K8S生命周期钩子 PostStart 执行自定义动作，启动limiter进程。pkg\\device\\ascend\\device.go **MutateAdmission**注入 postStart 钩子
+
+```ruby {"data-theme":"githubLight"}
+lifecycle:
+  postStart:
+    exec:
+      command:
+        - "bash"
+        - "-c"
+        - |
+          export RUST_LOG=info
+          /usr/local/hami-vnpu-core/limiter > /usr/local/hami-vnpu-core/inst1_manager.log 2>&1 &
+```
+
+**libvnpu.so改动点：**因为 postStart 无法保证在entrypoint之前完成。在业务真正开启分配内存或启动算力之前，**循环检测确保 lilmiter已经准备就绪**。
+
+```ruby {"data-theme":"githubLight"}
+impl SchedulerClient {
+    pub fn new() -> Self {
+        let pid = std::process::id();
+        let shmem_name = local_shmem_name();
+        
+        // ================== 新增：自动化等待逻辑 ==================
+        let shm_path = format!("/dev/shm/{}", shmem_name);
+        let mut retry_count = 0;
+        
+        info!("[Worker PID:{}] Checking for limiter at {}...", pid, shm_path);
+        
+        while !std::path::Path::new(&shm_path).exists() {
+            if retry_count % 50 == 0 { // 每 5 秒提醒一次
+                warn!("[Worker PID:{}] Still waiting for limiter process to create shared memory...", pid);
+            }
+            
+            std::thread::sleep(std::time::Duration::from_millis(100));
+            retry_count += 1;
+            
+            // 如果 1 分钟还没起来就崩溃
+            if retry_count > 600 { 
+                panic!("[Scheduler] FATAL: Limiter not found after 60 seconds.");
+            }
+        }
+        info!("[Worker PID:{}] Limiter detected. Connecting to shared memory.", pid);
+        // ========================================================
+
+        // 此时文件一定存在了，安全打开
+        let shmem = shmem::shm_setup::open_shmem::<LocalContainerShmem>(shmem_name.as_str());
+    }
+}
+```
+
+## 三、hami-ascend-device-plugin改动点
+
+### 3\.1 宿主机固定路径存放limiter和libvnpu.so
+
+宿主机固定路径存放limiter和libvnpu.so，为后续映射limiter和libvnpu.so到容器内准备：
+
+```shell {"data-theme":"githubLight"}
+/usr/local/hami-vnpu-core/
+├── limiter          
+└── libvnpu.so       
+└── ld.so.preload  // 动态链接预加载配置文件
+```
+
+**ld.so.preload文件内容**
+
+```plain-text {"data-theme":"githubLight"}
+/hami-vnpu-core/target/debug/libvnpu.so
+```
+
+### 3\.2 宿主机创建共享内存区域
+
+```shell {"data-theme":"githubLight"}
+sudo mkdir -p /usr/local/hami-shared-region
+sudo chmod 777 /usr/local/hami-shared-region
+```
+
+### 3\.3 Allocate函数增强
+
+```go {"data-theme":"githubLight"}
+func (ps *PluginServer) Allocate(ctx context.Context, reqs *v1beta1.AllocateRequest)
+   /*
+   1. (可省略)处理设备注入 可见设备环境变量设置后 设备文件会自动注入)
+   2. 处理存储卷注入 (-v)
+      A. 华为驱动与 SMI 工具链 
+      B. vnpu-core代码路径 : /usr/local/hami-vnpu-core
+      C. 注入 HAMi 劫持库路径 (对应脚本中的 LIBRARY_PATH)，通过/etc/ld.so.preload 挂载
+      D. HAMi 算力切分专用共享目录 : /usr/local/hami-shared-region:/hami-shared-region
+   3. 环境变量注入 
+      A. 可见设备IDs
+      B. 共享内存通信路径：NPU_GLOBAL_SHM_PATH = /hami-shared-region/{ID}_global_registry
+      C. 显存配额NPU_MEM_QUOTA：从Annotation读取 memory = 28672
+      D. 优先级NPU_PRIORITY：从Annotation读取 core = 20
+   */
+}
+```
\ No newline at end of file
diff --git a/pkg/device/ascend/device_test.go b/pkg/device/ascend/device_test.go
index fc9e59da84..013c1d6bdf 100644
--- a/pkg/device/ascend/device_test.go
+++ b/pkg/device/ascend/device_test.go
@@ -345,6 +345,62 @@ func Test_PatchAnnotations(t *testing.T) {
 	}
 }
 
+func Test_PatchAnnotations_VNPUCoreMode(t *testing.T) {
+	dev := Devices{
+		config: VNPUConfig{
+			CommonWord:   "Ascend910B3",
+			ResourceName: "huawei.com/Ascend910B3",
+		},
+	}
+
+	tests := []struct {
+		name string
+		pod  *corev1.Pod
+		pd   device.PodDevices
+		want map[string]string
+	}{
+		{
+			name: "vNPU-mode patch: check json contains cores and memory instead of template",
+			pod: &corev1.Pod{
+				ObjectMeta: metav1.ObjectMeta{
+					Annotations: map[string]string{
+						"huawei.com/vnpu-mode": "hami-core",
+					},
+				},
+			},
+			pd: device.PodDevices{
+				"Ascend910B3": device.PodSingleDevice{
+					[]device.ContainerDevice{
+						{
+							Idx:       0,
+							UUID:      "ascend-uuid-1",
+							Type:      "Ascend",
+							Usedcores: 5,
+							Usedmem:   8192,
+						},
+					},
+				},
+			},
+			want: map[string]string{
+				"huawei.com/Ascend910B3": "[{\"UUID\":\"ascend-uuid-1\",\"core\":5,\"memory\":8192}]",
+			},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			annoInput := make(map[string]string)
+			result := dev.PatchAnnotations(test.pod, &annoInput, test.pd)
+
+			val, ok := result["huawei.com/Ascend910B3"]
+			assert.Assert(t, ok)
+			assert.Assert(t, strings.Contains(val, "\"core\":5"))
+			assert.Assert(t, strings.Contains(val, "\"memory\":8192"))
+			assert.Assert(t, !strings.Contains(val, "\"temp\""))
+		})
+	}
+}
+
 func Test_checkType(t *testing.T) {
 	dev := Devices{
 		config: VNPUConfig{
@@ -739,6 +795,88 @@ func Test_MutateAdmission910C(t *testing.T) {
 	}
 }
 
+func Test_MutateAdmission_VNPUCoreMode(t *testing.T) {
+	const VNPUModeAnnotation = "huawei.com/vnpu-mode"
+	const VNPUModeHamiCore = "hami-core"
+
+	tests := []struct {
+		name string
+		args struct {
+			ctr corev1.Container
+			pod corev1.Pod
+		}
+		wantPostStart bool
+		wantMem       int64
+		wantCore      int64
+	}{
+		{
+			name: "vNPU-mode hami-core: inject postStart and keep raw memory",
+			args: struct {
+				ctr corev1.Container
+				pod corev1.Pod
+			}{
+				ctr: corev1.Container{
+					Name: "test-container",
+					Resources: corev1.ResourceRequirements{
+						Limits: corev1.ResourceList{
+							"huawei.com/Ascend910B3":        resource.MustParse("1"),
+							"huawei.com/Ascend910B3-memory": resource.MustParse("15360"),
+							"huawei.com/Ascend910B3-core":   resource.MustParse("20"),
+						},
+						Requests: corev1.ResourceList{
+							"huawei.com/Ascend910B3":        resource.MustParse("1"),
+							"huawei.com/Ascend910B3-memory": resource.MustParse("15360"),
+							"huawei.com/Ascend910B3-core":   resource.MustParse("20"),
+						},
+					},
+				},
+				pod: corev1.Pod{
+					ObjectMeta: metav1.ObjectMeta{
+						Annotations: map[string]string{
+							VNPUModeAnnotation: VNPUModeHamiCore,
+						},
+					},
+				},
+			},
+			wantPostStart: true,
+			wantMem:       15360,
+			wantCore:      20,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			dev := Devices{
+				config: VNPUConfig{
+					CommonWord:         "Ascend910B3",
+					ResourceName:       "huawei.com/Ascend910B3",
+					ResourceMemoryName: "huawei.com/Ascend910B3-memory",
+					ResourceCoreName:   "huawei.com/Ascend910B3-core",
+					MemoryAllocatable:  32768,
+				},
+			}
+
+			ok, err := dev.MutateAdmission(&test.args.ctr, &test.args.pod)
+			assert.NilError(t, err)
+			assert.Equal(t, ok, true)
+
+			if test.wantPostStart {
+				assert.Assert(t, test.args.ctr.Lifecycle != nil, "Lifecycle should not be nil")
+				assert.Assert(t, test.args.ctr.Lifecycle.PostStart != nil, "PostStart should not be nil")
+				assert.Assert(t, test.args.ctr.Lifecycle.PostStart.Exec != nil)
+				commandStr := strings.Join(test.args.ctr.Lifecycle.PostStart.Exec.Command, " ")
+				assert.Assert(t, strings.Contains(commandStr, "/hami-vnpu-core/limiter"))
+			}
+
+			memLimit := test.args.ctr.Resources.Limits[corev1.ResourceName(dev.config.ResourceMemoryName)]
+			assert.Equal(t, memLimit.Value(), test.wantMem)
+
+			coreLimit := test.args.ctr.Resources.Limits[corev1.ResourceName(dev.config.ResourceCoreName)]
+			assert.Equal(t, coreLimit.Value(), test.wantCore)
+		})
+	}
+}
+
 func Test_GenerateResourceRequests(t *testing.T) {
 	tests := []struct {
 		name string
@@ -835,6 +973,52 @@ func Test_GenerateResourceRequests(t *testing.T) {
 	}
 }
 
+func Test_GenerateResourceRequests_VNPUCoreMode(t *testing.T) {
+	tests := []struct {
+		name string
+		args corev1.Container
+		want device.ContainerDeviceRequest
+	}{
+		{
+			name: "parse custom vNPU core and memory resources",
+			args: corev1.Container{
+				Resources: corev1.ResourceRequirements{
+					Limits: corev1.ResourceList{
+						"huawei.com/Ascend910B3":        resource.MustParse("1"),
+						"huawei.com/Ascend910B3-core":   resource.MustParse("10"),
+						"huawei.com/Ascend910B3-memory": resource.MustParse("15360"),
+					},
+				},
+			},
+			want: device.ContainerDeviceRequest{
+				Nums:             int32(1),
+				Type:             "Ascend910B3",
+				Memreq:           int32(15360),
+				Coresreq:         int32(10),
+				MemPercentagereq: int32(0),
+			},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			dev := Devices{
+				config: VNPUConfig{
+					CommonWord:         "Ascend910B3",
+					ResourceName:       "huawei.com/Ascend910B3",
+					ResourceCoreName:   "huawei.com/Ascend910B3-core",
+					ResourceMemoryName: "huawei.com/Ascend910B3-memory",
+				},
+			}
+			result := dev.GenerateResourceRequests(&test.args)
+
+			assert.Equal(t, result.Memreq, test.want.Memreq)
+			assert.Equal(t, result.Coresreq, test.want.Coresreq)
+			assert.Equal(t, result.Type, test.want.Type)
+		})
+	}
+}
+
 func Test_GenerateResourceRequestsFactor(t *testing.T) {
 	req := corev1.Container{
 		Resources: corev1.ResourceRequirements{

From 1be08037b55e764e425d5311e60cc6fdc72e87fd Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Tue, 21 Apr 2026 02:46:27 +0000
Subject: [PATCH 07/13] feat: hami-vnpu-core ReadMe modify

Signed-off-by: ashergaga <1214443299@qq.com>
---
 docs/ascend910-hami-vnpu-core-support.md    | 22 ---------------------
 docs/ascend910-hami-vnpu-core-support_cn.md | 22 ---------------------
 2 files changed, 44 deletions(-)

diff --git a/docs/ascend910-hami-vnpu-core-support.md b/docs/ascend910-hami-vnpu-core-support.md
index 7ce3ebe250..bdd344e5da 100644
--- a/docs/ascend910-hami-vnpu-core-support.md
+++ b/docs/ascend910-hami-vnpu-core-support.md
@@ -19,28 +19,6 @@ Implements  a soft slicing mechanism based on `libvnpu.so` interception and `lim
 * [ascend-device-plugin](https://github.com/Project-HAMi/ascend-device-plugin)
 * **Chip Mode**: enable `device-share` mode on Ascend chips for virtualization
 
-### Host Environment Preparation
-
-Before launching any containers, the **Global Shared Memory (SHM) Region** must be initialized on the host to allow inter-Pod coordination.
-
-#### 1. Create the Shared Directory
-
-```bash
-sudo mkdir -p /usr/local/hami-shared-region
-sudo chmod 777 /usr/local/hami-shared-region
-```
-
-#### 2. Deploy hami-vnpu-core Components
-
-Place the following files in a fixed host path (`/usr/local/hami-vnpu-core/`) for mounting into containers:
-
-```
-/usr/local/hami-vnpu-core/
-├── limiter              # Manager daemon binary (compiled from hami-vnpu-core)
-├── libvnpu.so           # Interception library for LD_PRELOAD
-└── ld.so.preload        # Global preload config 
-```
-
   ### Create device-config.yaml
 
 The content is as follows:
diff --git a/docs/ascend910-hami-vnpu-core-support_cn.md b/docs/ascend910-hami-vnpu-core-support_cn.md
index 435cdd407a..7271c99201 100644
--- a/docs/ascend910-hami-vnpu-core-support_cn.md
+++ b/docs/ascend910-hami-vnpu-core-support_cn.md
@@ -21,28 +21,6 @@
 * 部署 [ascend-device-plugin](https://github.com/Project-HAMi/ascend-device-plugin)
 * 芯片模式：在昇腾芯片上开启 `device-share` 模式以支持虚拟化。
 
-## 宿主机环境准备
-
-在启动任何容器之前，必须在宿主机上初始化 **全局共享内存 (SHM) 区域**，以便进行 Pod 间的协同。
-
-1. **创建共享目录**
-
-   ```
-   sudo mkdir -p /usr/local/hami-shared-region
-   sudo chmod 777 /usr/local/hami-shared-region
-   ```
-
-2. **部署 hami-vnpu-core 组件** 
-
-   将以下文件放置在固定的宿主机路径（`/usr/local/hami-vnpu-core/`）中，以便挂载到容器内： 
-
-   ```
-   /usr/local/hami-vnpu-core/
-   ├── limiter              # Manager daemon binary (compiled from hami-vnpu-core)
-   ├── libvnpu.so           # Interception library for LD_PRELOAD
-   └── ld.so.preload        # Global preload config 
-   ```
-
   ### 在 files 目录下创建 device-config.yaml
 
   配置文件如下所示，可以按需调整：

From 3b26978a42ebefa3c2ff008694016c9d810e4d6b Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Mon, 27 Apr 2026 08:42:32 +0000
Subject: [PATCH 08/13] feat: add failureReason ModeNotFit

Signed-off-by: ashergaga <1214443299@qq.com>
---
 pkg/device/ascend/device.go | 37 +++++++++++++++++++++++++++++++++----
 pkg/device/common/common.go |  1 +
 2 files changed, 34 insertions(+), 4 deletions(-)

diff --git a/pkg/device/ascend/device.go b/pkg/device/ascend/device.go
index 7e0765e0da..10d61a6d91 100644
--- a/pkg/device/ascend/device.go
+++ b/pkg/device/ascend/device.go
@@ -43,6 +43,7 @@ const (
 	Ascend910NetworkWeight = 10
 	VNPUModeAnnotation     = "huawei.com/vnpu-mode"
 	VNPUModeHamiCore       = "hami-core"
+	VNPUNodeSelectorAnnotation = "hami-vnpu-core"
 )
 
 type Devices struct {
@@ -430,16 +431,44 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 	originReq := k.Nums
 	prevnuma := -1
 	klog.InfoS("Allocating device for container request", "pod", klog.KObj(pod), "card request", k)
+	var tmpDevs map[string]device.ContainerDevices
+	tmpDevs = make(map[string]device.ContainerDevices)
+	reason := make(map[string]int)
 
 	vnpuMode := ""
 	if pod != nil && pod.Annotations != nil {
 		vnpuMode = pod.Annotations[VNPUModeAnnotation]
 	}
+
+	
+	isHAMiCore := (vnpuMode == VNPUModeHamiCore)
+
+	// Werify whether the Node supports hami vnpu core 
+	nodeSupportHamiCore := false
+	if nodeInfo.Node.Annotations != nil {
+		if val, ok := nodeInfo.Node.Annotations[VNPUNodeSelectorAnnotation]; ok && val == "true" {
+			nodeSupportHamiCore = true
+		}
+	}
+
+	var totalMemPerCard int32 = 0
+	if len(devices) > 0 {
+		totalMemPerCard = devices[0].Totalmem
+	}
+
+	if request.Memreq > 0 && request.Memreq < totalMemPerCard && request.Nums > 0 {
+        if !nodeSupportHamiCore && isHAMiCore {
+			reason[common.ModeNotFit]++
+			klog.V(4).InfoS("Node filtered: Node does not support hami-core mode", "node", nodeInfo.Node.Name, "pod", pod.Name)
+			return false, nil, common.GenReason(reason, len(devices))
+		} else if nodeSupportHamiCore && !isHAMiCore {
+			reason[common.ModeNotFit]++
+			klog.V(4).InfoS("Node filtered: Reserved for hami-core but pod is legacy vNPU", "node", nodeInfo.Node.Name, "pod", pod.Name)
+			return false, nil, common.GenReason(reason, len(devices))
+		}
+	}
 	klog.V(4).InfoS("Fit: vnpu-mode annotation", "pod", pod.Name, "vnpuMode", vnpuMode)
 
-	var tmpDevs map[string]device.ContainerDevices
-	tmpDevs = make(map[string]device.ContainerDevices)
-	reason := make(map[string]int)
 	needTopology := false
 	if strings.HasPrefix(npu.CommonWord(), Ascend910Prefix) && hasNetworkID(devices) {
 		klog.V(4).Infof("all devices have NetworkID. device CommonWord %s", npu.CommonWord())
@@ -495,7 +524,7 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 		}
 		// Set dev.Totalcore to 100 if vnpuMode is hami-core
 		effectiveTotalCore := dev.Totalcore
-		if vnpuMode == VNPUModeHamiCore {
+		if isHAMiCore {
 			effectiveTotalCore = 100
 		}
 
diff --git a/pkg/device/common/common.go b/pkg/device/common/common.go
index 65410143b0..c14ec66778 100644
--- a/pkg/device/common/common.go
+++ b/pkg/device/common/common.go
@@ -37,6 +37,7 @@ const (
 	NodeUnfitPod                      = "NodeUnfitPod"
 	NodeFitPod                        = "NodeFitPod"
 	ResourceQuotaNotFit               = "ResourceQuotaNotFit"
+	ModeNotFit					   	  = "ModeNotFit"
 )
 
 func GenReason(reasons map[string]int, cards int) string {

From dd90ef315dc4d31968df702610b5fffacfcee647 Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Tue, 28 Apr 2026 08:02:50 +0000
Subject: [PATCH 09/13] feat: fix ai suggestion

Signed-off-by: ashergaga <1214443299@qq.com>
---
 pkg/device/ascend/device.go | 19 ++++++++-----------
 pkg/device/common/common.go |  2 +-
 2 files changed, 9 insertions(+), 12 deletions(-)

diff --git a/pkg/device/ascend/device.go b/pkg/device/ascend/device.go
index 10d61a6d91..0b148a8aec 100644
--- a/pkg/device/ascend/device.go
+++ b/pkg/device/ascend/device.go
@@ -439,16 +439,13 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 	if pod != nil && pod.Annotations != nil {
 		vnpuMode = pod.Annotations[VNPUModeAnnotation]
 	}
-
 	
 	isHAMiCore := (vnpuMode == VNPUModeHamiCore)
 
-	// Werify whether the Node supports hami vnpu core 
+	// Verify whether the Node supports hami vnpu core 
 	nodeSupportHamiCore := false
 	if nodeInfo.Node.Annotations != nil {
-		if val, ok := nodeInfo.Node.Annotations[VNPUNodeSelectorAnnotation]; ok && val == "true" {
-			nodeSupportHamiCore = true
-		}
+		nodeSupportHamiCore = nodeInfo.Node.Annotations[VNPUNodeSelectorAnnotation] == "true"
 	}
 
 	var totalMemPerCard int32 = 0
@@ -457,13 +454,13 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 	}
 
 	if request.Memreq > 0 && request.Memreq < totalMemPerCard && request.Nums > 0 {
-        if !nodeSupportHamiCore && isHAMiCore {
+        if nodeSupportHamiCore != isHAMiCore {
 			reason[common.ModeNotFit]++
-			klog.V(4).InfoS("Node filtered: Node does not support hami-core mode", "node", nodeInfo.Node.Name, "pod", pod.Name)
-			return false, nil, common.GenReason(reason, len(devices))
-		} else if nodeSupportHamiCore && !isHAMiCore {
-			reason[common.ModeNotFit]++
-			klog.V(4).InfoS("Node filtered: Reserved for hami-core but pod is legacy vNPU", "node", nodeInfo.Node.Name, "pod", pod.Name)
+			logMsg := "Node filtered: Reserved for hami-core but pod is legacy vNPU"
+			if isHAMiCore {
+				logMsg = "Node filtered: Node does not support hami-core mode"
+			}
+			klog.V(4).InfoS(logMsg, "node", nodeInfo.Node.Name, "pod", pod.Name)
 			return false, nil, common.GenReason(reason, len(devices))
 		}
 	}
diff --git a/pkg/device/common/common.go b/pkg/device/common/common.go
index c14ec66778..6405dc8dc7 100644
--- a/pkg/device/common/common.go
+++ b/pkg/device/common/common.go
@@ -37,7 +37,7 @@ const (
 	NodeUnfitPod                      = "NodeUnfitPod"
 	NodeFitPod                        = "NodeFitPod"
 	ResourceQuotaNotFit               = "ResourceQuotaNotFit"
-	ModeNotFit					   	  = "ModeNotFit"
+	ModeNotFit                        = "ModeNotFit"
 )
 
 func GenReason(reasons map[string]int, cards int) string {

From a6c1449a785d113565d6f97a27b6fb83e3076668 Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Tue, 28 Apr 2026 08:10:06 +0000
Subject: [PATCH 10/13] feat: fix cilint

Signed-off-by: ashergaga <1214443299@qq.com>
---
 pkg/device/ascend/device.go | 18 +++++++++---------
 1 file changed, 9 insertions(+), 9 deletions(-)

diff --git a/pkg/device/ascend/device.go b/pkg/device/ascend/device.go
index 0b148a8aec..c612bdf607 100644
--- a/pkg/device/ascend/device.go
+++ b/pkg/device/ascend/device.go
@@ -37,12 +37,12 @@ import (
 )
 
 const (
-	NodeLockAscend         = "hami.io/mutex.lock"
-	Ascend910Prefix        = "Ascend910"
-	Ascend910CType         = "Ascend910C"
-	Ascend910NetworkWeight = 10
-	VNPUModeAnnotation     = "huawei.com/vnpu-mode"
-	VNPUModeHamiCore       = "hami-core"
+	NodeLockAscend             = "hami.io/mutex.lock"
+	Ascend910Prefix            = "Ascend910"
+	Ascend910CType             = "Ascend910C"
+	Ascend910NetworkWeight     = 10
+	VNPUModeAnnotation         = "huawei.com/vnpu-mode"
+	VNPUModeHamiCore           = "hami-core"
 	VNPUNodeSelectorAnnotation = "hami-vnpu-core"
 )
 
@@ -439,10 +439,10 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 	if pod != nil && pod.Annotations != nil {
 		vnpuMode = pod.Annotations[VNPUModeAnnotation]
 	}
-	
+
 	isHAMiCore := (vnpuMode == VNPUModeHamiCore)
 
-	// Verify whether the Node supports hami vnpu core 
+	// Verify whether the Node supports hami vnpu core
 	nodeSupportHamiCore := false
 	if nodeInfo.Node.Annotations != nil {
 		nodeSupportHamiCore = nodeInfo.Node.Annotations[VNPUNodeSelectorAnnotation] == "true"
@@ -454,7 +454,7 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 	}
 
 	if request.Memreq > 0 && request.Memreq < totalMemPerCard && request.Nums > 0 {
-        if nodeSupportHamiCore != isHAMiCore {
+		if nodeSupportHamiCore != isHAMiCore {
 			reason[common.ModeNotFit]++
 			logMsg := "Node filtered: Reserved for hami-core but pod is legacy vNPU"
 			if isHAMiCore {

From 2936d0060164c8787efae572c68910ea37aecc59 Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Tue, 28 Apr 2026 08:41:00 +0000
Subject: [PATCH 11/13] feat: fix unit test error

Signed-off-by: ashergaga <1214443299@qq.com>
---
 pkg/device/ascend/device.go | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

diff --git a/pkg/device/ascend/device.go b/pkg/device/ascend/device.go
index c612bdf607..c4d03ce380 100644
--- a/pkg/device/ascend/device.go
+++ b/pkg/device/ascend/device.go
@@ -444,7 +444,8 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 
 	// Verify whether the Node supports hami vnpu core
 	nodeSupportHamiCore := false
-	if nodeInfo.Node.Annotations != nil {
+
+	if nodeInfo != nil && nodeInfo.Node != nil && nodeInfo.Node.Annotations != nil {
 		nodeSupportHamiCore = nodeInfo.Node.Annotations[VNPUNodeSelectorAnnotation] == "true"
 	}
 
@@ -454,13 +455,13 @@ func (npu *Devices) Fit(devices []*device.DeviceUsage, request device.ContainerD
 	}
 
 	if request.Memreq > 0 && request.Memreq < totalMemPerCard && request.Nums > 0 {
-		if nodeSupportHamiCore != isHAMiCore {
+		if !nodeSupportHamiCore && isHAMiCore {
 			reason[common.ModeNotFit]++
-			logMsg := "Node filtered: Reserved for hami-core but pod is legacy vNPU"
-			if isHAMiCore {
-				logMsg = "Node filtered: Node does not support hami-core mode"
-			}
-			klog.V(4).InfoS(logMsg, "node", nodeInfo.Node.Name, "pod", pod.Name)
+			klog.V(4).InfoS("Node filtered: Node does not support hami-core mode", "node", nodeInfo.Node.Name, "pod", pod.Name)
+			return false, nil, common.GenReason(reason, len(devices))
+		} else if nodeSupportHamiCore && !isHAMiCore {
+			reason[common.ModeNotFit]++
+			klog.V(4).InfoS("Node filtered: Reserved for hami-core but pod is legacy vNPU", "node", nodeInfo.Node.Name, "pod", pod.Name)
 			return false, nil, common.GenReason(reason, len(devices))
 		}
 	}

From 27bef0df9f5601f863a33fa2e876fce0e7341122 Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Tue, 28 Apr 2026 09:25:18 +0000
Subject: [PATCH 12/13] feat: add unit test

Signed-off-by: ashergaga <1214443299@qq.com>
---
 pkg/device/ascend/device_test.go | 58 +++++++++++++++++++++++++++-----
 1 file changed, 50 insertions(+), 8 deletions(-)

diff --git a/pkg/device/ascend/device_test.go b/pkg/device/ascend/device_test.go
index 013c1d6bdf..ac2616710a 100644
--- a/pkg/device/ascend/device_test.go
+++ b/pkg/device/ascend/device_test.go
@@ -1385,14 +1385,15 @@ func TestDevices_Fit(t *testing.T) {
 	devs := InitDevices(config)
 
 	tests := []struct {
-		name       string
-		devices    []*device.DeviceUsage
-		request    device.ContainerDeviceRequest
-		annos      map[string]string
-		wantFit    bool
-		wantLen    int
-		wantDevIDs []string
-		wantReason string
+		name           string
+		devices        []*device.DeviceUsage
+		request        device.ContainerDeviceRequest
+		annos          map[string]string
+		nodeAnnotation map[string]string
+		wantFit        bool
+		wantLen        int
+		wantDevIDs     []string
+		wantReason     string
 	}{
 		{
 			name: "fit success",
@@ -1776,6 +1777,42 @@ func TestDevices_Fit(t *testing.T) {
 			wantDevIDs: []string{"dev-2", "dev-1"},
 			wantReason: "",
 		},
+		{
+			name: "fit fail: hami-core pod on legacy node (ModeNotFit)",
+			devices: []*device.DeviceUsage{{
+				ID: "dev-0", Index: 0, Used: 0, Count: 100,
+				Usedmem: 0, Totalmem: 32768, Totalcore: 100, Usedcores: 0,
+				Numa: 0, Health: true,
+			}},
+			request: device.ContainerDeviceRequest{
+				Nums: 1, Memreq: 15360, MemPercentagereq: 0, Coresreq: 20,
+			},
+			annos: map[string]string{
+				VNPUModeAnnotation: VNPUModeHamiCore,
+			},
+			wantFit:        false,
+			wantLen:        0,
+			wantDevIDs:     []string{},
+			wantReason:     "1/1 ModeNotFit",
+			nodeAnnotation: map[string]string{},
+		},
+		{
+			name: "fit fail: legacy pod on hami-core reserved node (ModeNotFit)",
+			devices: []*device.DeviceUsage{{
+				ID: "dev-0", Index: 0, Used: 0, Count: 100,
+				Usedmem: 0, Totalmem: 32768, Totalcore: 100, Usedcores: 0,
+				Numa: 0, Health: true,
+			}},
+			request: device.ContainerDeviceRequest{
+				Nums: 1, Memreq: 8738, MemPercentagereq: 0, Coresreq: 0, 
+			},
+			annos:          map[string]string{},
+			wantFit:        false,
+			wantLen:        0,
+			wantDevIDs:     []string{},
+			wantReason:     "1/1 ModeNotFit",
+			nodeAnnotation: map[string]string{VNPUNodeSelectorAnnotation: "true"},
+		},
 	}
 
 	for _, dev := range devs {
@@ -1802,6 +1839,11 @@ func TestDevices_Fit(t *testing.T) {
 				}
 				nodeInfo := &device.NodeInfo{
 					ID: "node1",
+					Node: &corev1.Node{
+						ObjectMeta: metav1.ObjectMeta{
+							Annotations: test.nodeAnnotation,
+						},
+					},
 					Devices: map[string][]device.DeviceInfo{
 						dev.config.CommonWord: {
 							{

From 73e6b6edc0f993fb113e550bdd9cf331b8624699 Mon Sep 17 00:00:00 2001
From: ashergaga <1214443299@qq.com>
Date: Tue, 28 Apr 2026 09:31:28 +0000
Subject: [PATCH 13/13] feat: fix cilint

Signed-off-by: ashergaga <1214443299@qq.com>
---
 pkg/device/ascend/device_test.go | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/pkg/device/ascend/device_test.go b/pkg/device/ascend/device_test.go
index ac2616710a..da7588141c 100644
--- a/pkg/device/ascend/device_test.go
+++ b/pkg/device/ascend/device_test.go
@@ -1804,7 +1804,7 @@ func TestDevices_Fit(t *testing.T) {
 				Numa: 0, Health: true,
 			}},
 			request: device.ContainerDeviceRequest{
-				Nums: 1, Memreq: 8738, MemPercentagereq: 0, Coresreq: 0, 
+				Nums: 1, Memreq: 8738, MemPercentagereq: 0, Coresreq: 0,
 			},
 			annos:          map[string]string{},
 			wantFit:        false,