enhancement: gpu list based on backend

This commit is contained in:
Faisal Amir 2025-07-03 23:18:38 +07:00
parent 40f1fd4ffd
commit 19fc399ae1
3 changed files with 363 additions and 51 deletions

View File

@ -22,6 +22,7 @@ export interface GPU {
vendor: string
uuid: string
driver_version: string
activated?: boolean
nvidia_info: {
index: number
compute_capability: string
@ -99,6 +100,9 @@ interface HardwareStore {
// Update entire hardware data at once
setHardwareData: (data: HardwareData) => void
// Update hardware data while preserving GPU order
updateHardwareDataPreservingGpuOrder: (data: HardwareData) => void
// Update individual GPU
updateGPU: (index: number, gpu: GPU) => void
@ -119,6 +123,12 @@ interface HardwareStore {
// Reorder GPUs
reorderGPUs: (oldIndex: number, newIndex: number) => void
// Get activated GPU device string
getActivatedDeviceString: (backendType?: string) => string
// Update GPU activation states from device string
updateGPUActivationFromDeviceString: (deviceString: string) => void
}
export const useHardware = create<HardwareStore>()(
@ -172,7 +182,64 @@ export const useHardware = create<HardwareStore>()(
setHardwareData: (data) =>
set({
hardwareData: data,
hardwareData: {
...data,
gpus: data.gpus.map(gpu => ({
...gpu,
activated: gpu.activated ?? false
}))
},
}),
updateHardwareDataPreservingGpuOrder: (data) =>
set((state) => {
// If we have existing GPU data, preserve the order and activation state
if (state.hardwareData.gpus.length > 0) {
// Reorder fresh GPU data to match existing order, adding new GPUs at the end
const reorderedGpus: GPU[] = []
const processedUuids = new Set()
// First, add existing GPUs in their current order, preserving activation state
state.hardwareData.gpus.forEach(existingGpu => {
const freshGpu = data.gpus.find(gpu => gpu.uuid === existingGpu.uuid)
if (freshGpu) {
reorderedGpus.push({
...freshGpu,
activated: existingGpu.activated ?? false
})
processedUuids.add(freshGpu.uuid)
}
})
// Then, add any new GPUs that weren't in the existing order (default to inactive)
data.gpus.forEach(freshGpu => {
if (!processedUuids.has(freshGpu.uuid)) {
reorderedGpus.push({
...freshGpu,
activated: false
})
}
})
return {
hardwareData: {
...data,
gpus: reorderedGpus
}
}
} else {
// No existing GPU data, initialize all GPUs as inactive
return {
hardwareData: {
...data,
gpus: data.gpus.map(gpu => ({
...gpu,
activated: false
}))
}
}
}
}),
updateGPU: (index, gpu) =>
@ -195,35 +262,75 @@ export const useHardware = create<HardwareStore>()(
})),
toggleGPUActivation: async (index) => {
const { pausePolling, setGpuLoading } = get()
const { pausePolling, resumePolling, setGpuLoading } = get()
pausePolling()
setGpuLoading(index, true)
// try {
// await new Promise((resolve) => setTimeout(resolve, 200)) // Simulate async, replace with real API if needed
// set((state) => {
// const newGPUs = [...state.hardwareData.gpus]
// if (index >= 0 && index < newGPUs.length) {
// newGPUs[index] = {
// ...newGPUs[index],
// activated: !newGPUs[index].activated,
// }
// }
// setActiveGpus({
// gpus: newGPUs
// .filter((e) => e.activated)
// .map((e) => parseInt(e.id)),
// })
// return {
// hardwareData: {
// ...state.hardwareData,
// gpus: newGPUs,
// },
// }
// })
// } finally {
// setGpuLoading(index, false)
// setTimeout(resumePolling, 1000) // Resume polling after 1s
// }
try {
await new Promise((resolve) => setTimeout(resolve, 200)) // Simulate async operation
set((state) => {
const newGPUs = [...state.hardwareData.gpus]
if (index >= 0 && index < newGPUs.length) {
newGPUs[index] = {
...newGPUs[index],
activated: !newGPUs[index].activated,
}
}
return {
hardwareData: {
...state.hardwareData,
gpus: newGPUs,
},
}
})
// Update the device setting after state change
const updatedState = get()
// Import and get backend type
const { useModelProvider } = await import('./useModelProvider')
const { updateProvider, getProviderByName } = useModelProvider.getState()
const llamacppProvider = getProviderByName('llamacpp')
const backendType = llamacppProvider?.settings.find(s => s.key === 'version_backend')?.controller_props.value as string
const deviceString = updatedState.getActivatedDeviceString(backendType)
console.log(`GPU ${index} activation toggled. Backend: "${backendType}", New device string: "${deviceString}"`)
console.log('Activated GPUs:', updatedState.hardwareData.gpus.filter(gpu => gpu.activated).map((gpu, i) => ({
name: gpu.name,
nvidia: gpu.nvidia_info?.index,
vulkan: gpu.vulkan_info?.index,
activated: gpu.activated
})))
if (llamacppProvider) {
const updatedSettings = llamacppProvider.settings.map(setting => {
if (setting.key === 'device') {
return {
...setting,
controller_props: {
...setting.controller_props,
value: deviceString
}
}
}
return setting
})
updateProvider('llamacpp', {
settings: updatedSettings
})
console.log(`Updated llamacpp device setting to: "${deviceString}"`)
}
} finally {
setGpuLoading(index, false)
setTimeout(resumePolling, 1000) // Resume polling after 1s
}
},
reorderGPUs: (oldIndex, newIndex) =>
@ -246,6 +353,93 @@ export const useHardware = create<HardwareStore>()(
},
}
}),
getActivatedDeviceString: (backendType?: string) => {
const { hardwareData } = get()
// Get activated GPUs and generate appropriate device format based on backend
const activatedDevices = hardwareData.gpus
.filter(gpu => gpu.activated)
.map(gpu => {
const isCudaBackend = backendType?.includes('cuda')
const isVulkanBackend = backendType?.includes('vulkan')
// Handle different backend scenarios
if (isCudaBackend && isVulkanBackend) {
// Mixed backend - prefer CUDA for NVIDIA GPUs, Vulkan for others
if (gpu.nvidia_info) {
return `cuda:${gpu.nvidia_info.index}`
} else if (gpu.vulkan_info) {
return `vulkan:${gpu.vulkan_info.index}`
}
} else if (isCudaBackend && gpu.nvidia_info) {
// CUDA backend - only use CUDA-compatible GPUs
return `cuda:${gpu.nvidia_info.index}`
} else if (isVulkanBackend && gpu.vulkan_info) {
// Vulkan backend - only use Vulkan-compatible GPUs
return `vulkan:${gpu.vulkan_info.index}`
} else if (!backendType) {
// No backend specified, use GPU's preferred type
if (gpu.nvidia_info) {
return `cuda:${gpu.nvidia_info.index}`
} else if (gpu.vulkan_info) {
return `vulkan:${gpu.vulkan_info.index}`
}
}
return null
})
.filter(device => device !== null) as string[]
const deviceString = activatedDevices.join(',')
return deviceString
},
updateGPUActivationFromDeviceString: (deviceString: string) => {
set((state) => {
const newGPUs = [...state.hardwareData.gpus]
// Parse device string to get active device indices
const activeDevices = deviceString
.split(',')
.map(device => device.trim())
.filter(device => device.length > 0)
.map(device => {
const match = device.match(/^(cuda|vulkan):(\d+)$/)
if (match) {
return {
type: match[1] as 'cuda' | 'vulkan',
index: parseInt(match[2])
}
}
return null
})
.filter(device => device !== null) as Array<{type: 'cuda' | 'vulkan', index: number}>
// Update GPU activation states
newGPUs.forEach((gpu, gpuIndex) => {
const shouldBeActive = activeDevices.some(device => {
if (device.type === 'cuda' && gpu.nvidia_info) {
return gpu.nvidia_info.index === device.index
} else if (device.type === 'vulkan' && gpu.vulkan_info) {
return gpu.vulkan_info.index === device.index
}
return false
})
newGPUs[gpuIndex] = {
...gpu,
activated: shouldBeActive
}
})
return {
hardwareData: {
...state.hardwareData,
gpus: newGPUs
}
}
})
},
}),
{
name: localStorageKey.settingHardware,

View File

@ -7,9 +7,9 @@ import { Switch } from '@/components/ui/switch'
import { Progress } from '@/components/ui/progress'
import { useTranslation } from '@/i18n/react-i18next-compat'
import { useHardware } from '@/hooks/useHardware'
import { useVulkan } from '@/hooks/useVulkan'
// import { useVulkan } from '@/hooks/useVulkan'
import type { GPU, HardwareData } from '@/hooks/useHardware'
import { useEffect } from 'react'
import { useEffect, useState } from 'react'
import {
DndContext,
closestCenter,
@ -34,13 +34,14 @@ import { WebviewWindow } from '@tauri-apps/api/webviewWindow'
import { formatMegaBytes } from '@/lib/utils'
import { windowKey } from '@/constants/windows'
import { toNumber } from '@/utils/number'
import { useModelProvider } from '@/hooks/useModelProvider'
// eslint-disable-next-line @typescript-eslint/no-explicit-any
export const Route = createFileRoute(route.settings.hardware as any)({
component: Hardware,
})
function SortableGPUItem({ gpu, index }: { gpu: GPU; index: number }) {
function SortableGPUItem({ gpu, index, isCompatible, isActivated }: { gpu: GPU; index: number; isCompatible: boolean; isActivated: boolean }) {
const {
attributes,
listeners,
@ -63,7 +64,7 @@ function SortableGPUItem({ gpu, index }: { gpu: GPU; index: number }) {
}
return (
<div ref={setNodeRef} style={style} className="mb-4 last:mb-0">
<div ref={setNodeRef} style={style} className={`mb-4 last:mb-0 ${!isCompatible ? 'opacity-60' : ''}`}>
<CardItem
title={
<div className="flex items-center gap-2">
@ -75,13 +76,18 @@ function SortableGPUItem({ gpu, index }: { gpu: GPU; index: number }) {
<IconGripVertical size={18} className="text-main-view-fg/60" />
</div>
<span className="text-main-view-fg/80">{gpu.name}</span>
{!isCompatible && (
<span className="text-xs bg-destructive/10 text-destructive px-2 py-1 rounded-sm">
Incompatible with current backend
</span>
)}
</div>
}
actions={
<div className="flex items-center gap-4">
<Switch
checked={true}
disabled={!!gpuLoading[index]}
checked={isActivated}
disabled={!!gpuLoading[index] || !isCompatible}
onCheckedChange={() => toggleGPUActivation(index)}
/>
</div>
@ -126,17 +132,109 @@ function Hardware() {
hardwareData,
systemUsage,
setHardwareData,
updateHardwareDataPreservingGpuOrder,
updateSystemUsage,
reorderGPUs,
pollingPaused,
} = useHardware()
const { vulkanEnabled, setVulkanEnabled } = useVulkan()
// const { vulkanEnabled, setVulkanEnabled } = useVulkan()
const { providers } = useModelProvider()
const llamacpp = providers.find((p) => p.provider === 'llamacpp')
const versionBackend = llamacpp?.settings.find((s) => s.key === "version_backend")?.controller_props.value
// Determine backend type and filter GPUs accordingly
const isCudaBackend = typeof versionBackend === 'string' && versionBackend.includes('cuda')
const isVulkanBackend = typeof versionBackend === 'string' && versionBackend.includes('vulkan')
// Filter and prepare GPUs based on backend
const getFilteredGPUs = () => {
// Always show all GPUs, but compatibility will be determined by isGPUActive
return hardwareData.gpus
}
const filteredGPUs = getFilteredGPUs()
// Check if GPU should be active based on backend compatibility
const isGPUCompatible = (gpu: GPU) => {
if (isCudaBackend) {
return gpu.nvidia_info !== null
} else if (isVulkanBackend) {
return gpu.vulkan_info !== null
} else {
// No valid backend - all GPUs are inactive
return false
}
}
// Check if GPU is actually activated
const isGPUActive = (gpu: GPU) => {
return isGPUCompatible(gpu) && (gpu.activated ?? false)
}
useEffect(() => {
getHardwareInfo().then((data) =>
setHardwareData(data as unknown as HardwareData)
)
}, [setHardwareData])
getHardwareInfo().then((freshData) => {
const data = freshData as unknown as HardwareData
updateHardwareDataPreservingGpuOrder(data)
})
}, [updateHardwareDataPreservingGpuOrder])
// Hardware and provider sync logic
const { getActivatedDeviceString, updateGPUActivationFromDeviceString } = useHardware()
const { updateProvider, getProviderByName } = useModelProvider()
const [isInitialized, setIsInitialized] = useState(false)
// Initialize GPU activations from device setting on first load
useEffect(() => {
if (hardwareData.gpus.length > 0 && !isInitialized) {
const llamacppProvider = getProviderByName('llamacpp')
const currentDeviceSetting = llamacppProvider?.settings.find(s => s.key === 'device')?.controller_props.value as string
if (currentDeviceSetting) {
console.log(`Initializing GPU activations from device setting: "${currentDeviceSetting}"`)
updateGPUActivationFromDeviceString(currentDeviceSetting)
}
setIsInitialized(true)
}
}, [hardwareData.gpus.length, isInitialized, getProviderByName, updateGPUActivationFromDeviceString])
// Sync device setting when GPU activations change (only after initialization)
const gpuActivationStates = hardwareData.gpus.map(gpu => gpu.activated)
useEffect(() => {
if (isInitialized && hardwareData.gpus.length > 0) {
const llamacppProvider = getProviderByName('llamacpp')
const backendType = llamacppProvider?.settings.find(s => s.key === 'version_backend')?.controller_props.value as string
const deviceString = getActivatedDeviceString(backendType)
if (llamacppProvider) {
const currentDeviceSetting = llamacppProvider.settings.find(s => s.key === 'device')
// Sync device string when GPU activations change (only after initialization)
if (currentDeviceSetting && currentDeviceSetting.controller_props.value !== deviceString) {
console.log(`Syncing device string from "${currentDeviceSetting.controller_props.value}" to "${deviceString}"`)
const updatedSettings = llamacppProvider.settings.map(setting => {
if (setting.key === 'device') {
return {
...setting,
controller_props: {
...setting.controller_props,
value: deviceString
}
}
}
return setting
})
updateProvider('llamacpp', {
settings: updatedSettings
})
}
}
}
}, [isInitialized, gpuActivationStates, versionBackend, getActivatedDeviceString, updateProvider, getProviderByName, hardwareData.gpus.length])
// Set up DnD sensors
const sensors = useSensors(
@ -149,13 +247,12 @@ function Hardware() {
const { active, over } = event
if (over && active.id !== over.id) {
// Find the indices of the dragged item and the drop target
const oldIndex = hardwareData.gpus.findIndex(
(_, index) => index === active.id
)
const newIndex = hardwareData.gpus.findIndex(
(_, index) => index === over.id
)
// Find the actual indices in the original hardwareData.gpus array
const activeGpu = filteredGPUs[active.id as number]
const overGpu = filteredGPUs[over.id as number]
const oldIndex = hardwareData.gpus.findIndex(gpu => gpu.uuid === activeGpu.uuid)
const newIndex = hardwareData.gpus.findIndex(gpu => gpu.uuid === overGpu.uuid)
if (oldIndex !== -1 && newIndex !== -1) {
reorderGPUs(oldIndex, newIndex)
@ -356,7 +453,7 @@ function Hardware() {
</Card>
{/* Vulkan Settings */}
{hardwareData.gpus.length > 0 && (
{/* {hardwareData.gpus.length > 0 && (
<Card title={t('settings:hardware.vulkan')}>
<CardItem
title={t('settings:hardware.enableVulkan')}
@ -376,11 +473,13 @@ function Hardware() {
}
/>
</Card>
)}
)} */}
{/* GPU Information */}
{!IS_MACOS ? (
<Card title={t('settings:hardware.gpus')}>
{hardwareData.gpus.length > 0 ? (
<DndContext
sensors={sensors}
@ -388,11 +487,17 @@ function Hardware() {
onDragEnd={handleDragEnd}
>
<SortableContext
items={hardwareData.gpus.map((_, index) => index)}
items={filteredGPUs.map((_, index) => index)}
strategy={verticalListSortingStrategy}
>
{hardwareData.gpus.map((gpu, index) => (
<SortableGPUItem key={index} gpu={gpu} index={index} />
{filteredGPUs.map((gpu, index) => (
<SortableGPUItem
key={index}
gpu={gpu}
index={index}
isCompatible={isGPUCompatible(gpu)}
isActivated={isGPUActive(gpu)}
/>
))}
</SortableContext>
</DndContext>

View File

@ -2,6 +2,7 @@ import { Card, CardItem } from '@/containers/Card'
import HeaderPage from '@/containers/HeaderPage'
import SettingsMenu from '@/containers/SettingsMenu'
import { useModelProvider } from '@/hooks/useModelProvider'
import { useHardware } from '@/hooks/useHardware'
import { cn, getProviderTitle } from '@/lib/utils'
import { open } from '@tauri-apps/plugin-dialog'
import {
@ -77,6 +78,7 @@ function ProviderDetail() {
const [refreshingModels, setRefreshingModels] = useState(false)
const { providerName } = useParams({ from: Route.id })
const { getProviderByName, setProviders, updateProvider } = useModelProvider()
const { updateGPUActivationFromDeviceString } = useHardware()
const provider = getProviderByName(providerName)
const isSetup = step === 'setup_remote_provider'
const navigate = useNavigate()
@ -282,6 +284,17 @@ function ProviderDetail() {
) {
updateObj.base_url = newValue
}
// Special handling for device setting changes
if (
settingKey === 'device' &&
typeof newValue === 'string' &&
provider.provider === 'llamacpp'
) {
console.log(`Device setting manually changed to: "${newValue}"`)
updateGPUActivationFromDeviceString(newValue)
}
updateSettings(
providerName,
updateObj.settings ?? []