Files
ComfyUI_frontend/src/platform/missingModel/missingModelScan.ts
jaeone94 8d82944441 fix: limit workflow models to metadata enrichment (#12990)
## Summary

This PR intentionally narrows workflow-embedded model metadata handling
so root-level `models[]` and node-level embedded model metadata can
enrich existing missing-model candidates, but can no longer create new
candidates by themselves.

## Why this PR exists

ADR 0009, **Subgraph promoted widgets use linked inputs**, changes
promoted value ownership for subgraphs. That design was implemented by
[#12197](https://github.com/Comfy-Org/ComfyUI_frontend/pull/12197),
**Subgraph Link Only Promotion (ADR 0009)**.

Under ADR 0009, a promoted widget is represented as a standard linked
`SubgraphInput` on the host `SubgraphNode`. The host boundary owns the
promoted value identity through the host node locator plus
`SubgraphInput.name`. The interior source widget remains the provider of
schema, type, options, tooltip, defaults, diagnostics, and migration
metadata, but it is not the persistence owner of the promoted value.

This PR is a preparatory cleanup discovered while working on the missing
model detection follow-up required by that ADR 0009 /
[#12197](https://github.com/Comfy-Org/ComfyUI_frontend/pull/12197)
behavior. The follow-up needs missing model detection to respect the new
subgraph promoted-widget ownership model. While reviewing that path, we
found that the existing embedded model metadata fallback in
`enrichWithEmbeddedMetadata` was doing more than metadata enrichment.

The important finding was that this fallback was not just attaching
metadata to candidates that had already been detected from live node
widgets. It could also synthesize brand-new `MissingModelCandidate`
entries from workflow JSON metadata, including root-level `models[]`
entries, when no live candidate existed.

That behavior is inaccurate for the missing model system for two
reasons.

First, the normal missing model lifecycle is anchored to a real
node/widget binding. A candidate found from a COMBO or asset widget has
a concrete `nodeId + widgetName` reference. That reference is what lets
the UI surface the error, cache it as a pending warning, and later clear
or resolve it when the underlying node/widget value is fixed. A
root-level `models[]` entry does not reliably provide that anchor. If
metadata-only fallback creates a candidate without a real live widget
reference, the resulting error can be detected but cannot reliably
travel through the existing clearing path. In practice, that can become
an effectively unremovable missing model warning unless the user
downloads exactly the same model referenced by the stale metadata.

Second, a missing model error is meant to mean that a model-selecting
widget on an active node references a value that is not available.
Workflow JSON metadata by itself is not the same source of truth. If a
model only appears in root workflow metadata, or appears in node
metadata that is not represented by an active COMBO or asset widget
candidate, that is a different kind of state from the existing missing
model error model. Treating that metadata as a candidate creates a
second, less reliable detector that is not aligned with the scan/clear
lifecycle.

This is especially important before the ADR 0009 missing-model
follow-up. With linked-input promoted widgets, the host promoted value
is the value that matters. The interior source widget may still carry
stale or default metadata, and it must not become a second source of
truth for missing model errors. A detection path that can create
candidates directly from workflow metadata would make it harder to
reason about which value actually produced the warning.

For those reasons, this PR removes metadata-only candidate synthesis and
keeps embedded metadata in the role it can perform safely: metadata
enrichment. If the live widget/asset scan produces a candidate, embedded
metadata may fill in `directory`, `url`, `hash`, and `hashType`. If no
live candidate exists, the metadata is not enough to create a missing
model warning.

This PR is intended to land before the child PR that updates runtime
missing model detection for ADR 0009 linked-input promoted widgets.

## Changes

- **What**: Restrict `enrichWithEmbeddedMetadata` to enriching existing
candidates instead of creating fallback candidates from unmatched root
`models[]` or embedded model metadata.
- **What**: Remove the now-unused installed-model check callback and
asset-support callback from `enrichWithEmbeddedMetadata`.
- **What**: Remove the now-unnecessary `modelStore.loadModelFolders()`
path from the missing model pipeline, since embedded metadata no longer
performs installed-model fallback detection.
- **What**: Remove dead source-tracking metadata
(`EmbeddedModelWithSource`, source node/widget fields, and widget-name
lookup) that only existed to support metadata-only synthesis.
- **What**: Update missing model tests so they assert the new contract:
metadata enriches live candidates, but does not create candidates
without a live scan result.
- **What**: Delete obsolete fixtures that only covered the removed
metadata-only synthesis path.
- **Breaking**: None expected. This is an intentional narrowing of an
inaccurate fallback detector, not a public API change.
- **Dependencies**: None.

## Review Focus

Please focus on whether the candidate lifecycle now has a single source
of truth: live COMBO/asset widget scanning creates candidates, while
workflow metadata only enriches those candidates.

The intended behavioral change is that a model present only in
workflow-level metadata, with no active node widget candidate
referencing it, no longer appears as a missing model. This avoids
surfacing warnings that cannot be cleared through the normal `nodeId +
widgetName` path.

The expected retained behavior is that active widget-referenced missing
models are still detected by `scanAllModelCandidates`, and metadata from
root `models[]` or node `properties.models` still supplies
download-related fields for those live candidates.

## Screenshots (if applicable)

Not applicable. This is a detection/pipeline behavior change covered by
unit tests.

## Validation

- `pnpm test:unit src/platform/missingModel/missingModelScan.test.ts
src/platform/missingModel/missingModelPipeline.test.ts`
- `pnpm exec eslint src/platform/missingModel/missingModelScan.ts
src/platform/missingModel/missingModelScan.test.ts
src/platform/missingModel/missingModelPipeline.ts
src/platform/missingModel/missingModelPipeline.test.ts
src/platform/missingModel/types.ts`
- `pnpm exec oxfmt --check src/platform/missingModel/missingModelScan.ts
src/platform/missingModel/missingModelScan.test.ts
src/platform/missingModel/missingModelPipeline.ts
src/platform/missingModel/missingModelPipeline.test.ts
src/platform/missingModel/types.ts`
- `pnpm typecheck`
- pre-push hook: `knip --cache`
2026-06-23 07:44:43 +00:00

398 lines
12 KiB
TypeScript

import type { ModelFile } from '@/platform/workflow/validation/schemas/workflowSchema'
import type { FlattenableWorkflowGraph } from '@/platform/workflow/core/utils/workflowFlattening'
import { flattenWorkflowNodes } from '@/platform/workflow/core/utils/workflowFlattening'
import type { MissingModelCandidate, MissingModelViewModel } from './types'
import { getAssetFilename } from '@/platform/assets/utils/assetMetadataUtils'
import type { AssetItem } from '@/platform/assets/schemas/assetSchema'
// eslint-disable-next-line import-x/no-restricted-paths
import { getSelectedModelsMetadata } from '@/workbench/utils/modelMetadataUtil'
import type { LGraph } from '@/lib/litegraph/src/LGraph'
import type { LGraphNode } from '@/lib/litegraph/src/LGraphNode'
import type {
IAssetWidget,
IBaseWidget,
IComboWidget
} from '@/lib/litegraph/src/types/widgets'
import {
collectAllNodes,
getExecutionIdByNode
} from '@/utils/graphTraversalUtil'
import { LGraphEventMode } from '@/lib/litegraph/src/types/globalEnums'
import { resolveComboValues } from '@/utils/litegraphUtil'
import { getParentExecutionIds } from '@/types/nodeIdentification'
export type MissingModelWorkflowData = FlattenableWorkflowGraph & {
models?: ModelFile[]
}
function isComboWidget(widget: IBaseWidget): widget is IComboWidget {
return widget.type === 'combo'
}
/**
* Fills url/hash/directory onto a candidate from the node's embedded
* `properties.models` metadata when the names match. The full pipeline
* does this via enrichWithEmbeddedMetadata + graphData.models, but the
* realtime single-node scan (paste, un-bypass) otherwise loses these
* fields — making the Missing Model row's download/copy-url buttons
* disappear after a bypass/un-bypass cycle.
*/
function enrichCandidateFromNodeProperties(
candidate: MissingModelCandidate,
embeddedModels: readonly ModelFile[] | undefined
): MissingModelCandidate {
if (!embeddedModels?.length) return candidate
// Require directory agreement when the candidate already has one —
// a single node can reference two models with the same name under
// different directories (e.g. a LoRA present in multiple folders);
// name-only matching would stamp the wrong url/hash onto the
// candidate. Mirrors the directory check in enrichWithEmbeddedMetadata.
const match = embeddedModels.find(
(m) =>
m.name === candidate.name &&
(!candidate.directory || candidate.directory === m.directory)
)
if (!match) return candidate
return {
...candidate,
directory: candidate.directory ?? match.directory,
url: candidate.url ?? match.url,
hash: candidate.hash ?? match.hash,
hashType: candidate.hashType ?? match.hash_type
}
}
function isAssetWidget(widget: IBaseWidget): widget is IAssetWidget {
return widget.type === 'asset'
}
function isInactiveMode(mode: number | undefined): boolean {
return mode === LGraphEventMode.NEVER || mode === LGraphEventMode.BYPASS
}
// Full set of model file extensions used for scanning candidate widgets.
// Intentionally broader than ALLOWED_SUFFIXES in missingModelDownload.ts,
// which restricts which files are eligible for download.
export const MODEL_FILE_EXTENSIONS = new Set([
'.safetensors',
'.ckpt',
'.pt',
'.pth',
'.bin',
'.sft',
'.onnx',
'.gguf'
])
export function isModelFileName(name: string): boolean {
const lower = name.toLowerCase()
return Array.from(MODEL_FILE_EXTENSIONS).some((ext) => lower.endsWith(ext))
}
/**
* Scan COMBO and asset widgets on configured graph nodes for model-like values.
* Must be called after `graph.configure()` so widget name/value mappings are accurate.
*
* Non-asset-supported nodes: `isMissing` resolved immediately via widget options.
* Asset-supported nodes: `isMissing` left `undefined` for async verification.
*/
export function scanAllModelCandidates(
rootGraph: LGraph,
isAssetSupported: (nodeType: string, widgetName: string) => boolean,
getDirectory?: (nodeType: string) => string | undefined
): MissingModelCandidate[] {
if (!rootGraph) return []
const allNodes = collectAllNodes(rootGraph)
const candidates: MissingModelCandidate[] = []
for (const node of allNodes) {
if (!node.widgets?.length) continue
// Skip subgraph container nodes: their promoted widgets are synthetic
// views of interior widgets, which are already scanned via recursion.
if (node.isSubgraphNode?.()) continue
if (isInactiveMode(node.mode)) continue
candidates.push(
...scanNodeModelCandidates(
rootGraph,
node,
isAssetSupported,
getDirectory
)
)
}
return candidates
}
/** Scan a single node's widgets for missing model candidates (OSS immediate resolution). */
export function scanNodeModelCandidates(
rootGraph: LGraph,
node: LGraphNode,
isAssetSupported: (nodeType: string, widgetName: string) => boolean,
getDirectory?: (nodeType: string) => string | undefined
): MissingModelCandidate[] {
if (!node.widgets?.length) return []
const executionId = getExecutionIdByNode(rootGraph, node)
if (!executionId) return []
const candidates: MissingModelCandidate[] = []
const embeddedModels = (node as { properties?: { models?: ModelFile[] } })
.properties?.models
for (const widget of node.widgets) {
let candidate: MissingModelCandidate | null = null
if (isAssetWidget(widget)) {
candidate = scanAssetWidget(node, widget, executionId, getDirectory)
} else if (isComboWidget(widget)) {
candidate = scanComboWidget(
node,
widget,
executionId,
isAssetSupported,
getDirectory
)
}
if (candidate) {
candidates.push(
enrichCandidateFromNodeProperties(candidate, embeddedModels)
)
}
}
return candidates
}
function scanAssetWidget(
node: { type: string },
widget: IAssetWidget,
executionId: string,
getDirectory: ((nodeType: string) => string | undefined) | undefined
): MissingModelCandidate | null {
const value = widget.value
if (typeof value !== 'string' || !value.trim()) return null
if (!isModelFileName(value)) return null
return {
nodeId: executionId,
nodeType: node.type,
widgetName: widget.name,
isAssetSupported: true,
name: value,
directory: getDirectory?.(node.type),
isMissing: undefined
}
}
function scanComboWidget(
node: { type: string },
widget: IComboWidget,
executionId: string,
isAssetSupported: (nodeType: string, widgetName: string) => boolean,
getDirectory: ((nodeType: string) => string | undefined) | undefined
): MissingModelCandidate | null {
const value = widget.value
if (typeof value !== 'string' || !value.trim()) return null
if (!isModelFileName(value)) return null
const nodeIsAssetSupported = isAssetSupported(node.type, widget.name)
const options = resolveComboValues(widget)
const inOptions = options.includes(value)
return {
nodeId: executionId,
nodeType: node.type,
widgetName: widget.name,
isAssetSupported: nodeIsAssetSupported,
name: value,
directory: getDirectory?.(node.type),
isMissing: nodeIsAssetSupported ? undefined : !inOptions
}
}
export function enrichWithEmbeddedMetadata(
candidates: readonly MissingModelCandidate[],
graphData: MissingModelWorkflowData
): MissingModelCandidate[] {
const allNodes = flattenWorkflowNodes(graphData)
const embeddedModels = collectEmbeddedModels(allNodes, graphData)
const enriched = candidates.map((c) => ({ ...c }))
const candidatesByKey = new Map<string, MissingModelCandidate[]>()
for (const c of enriched) {
const dirKey = `${c.name}::${c.directory ?? ''}`
const dirList = candidatesByKey.get(dirKey)
if (dirList) dirList.push(c)
else candidatesByKey.set(dirKey, [c])
const nameKey = c.name
const nameList = candidatesByKey.get(nameKey)
if (nameList) nameList.push(c)
else candidatesByKey.set(nameKey, [c])
}
const deduped: ModelFile[] = []
const enrichedKeys = new Set<string>()
for (const model of embeddedModels) {
const dedupeKey = `${model.name}::${model.directory}`
if (enrichedKeys.has(dedupeKey)) continue
enrichedKeys.add(dedupeKey)
deduped.push(model)
}
for (const model of deduped) {
const dirKey = `${model.name}::${model.directory}`
const exact = candidatesByKey.get(dirKey)
const fallback = candidatesByKey.get(model.name)
const existing = exact?.length ? exact : fallback
if (!existing) continue
for (const c of existing) {
if (c.directory && c.directory !== model.directory) continue
c.directory ??= model.directory
c.url ??= model.url
c.hash ??= model.hash
c.hashType ??= model.hash_type
}
}
return enriched
}
function collectEmbeddedModels(
allNodes: ReturnType<typeof flattenWorkflowNodes>,
graphData: MissingModelWorkflowData
): ModelFile[] {
const result: ModelFile[] = []
const nodesById = new Map(allNodes.map((node) => [String(node.id), node]))
for (const node of allNodes) {
if (!isNodeAndAncestorsActive(node, nodesById)) continue
const selected = getSelectedModelsMetadata(node)
if (!selected?.length) continue
result.push(...selected)
}
if (graphData.models?.length) result.push(...graphData.models)
return result
}
function isNodeAndAncestorsActive(
node: ReturnType<typeof flattenWorkflowNodes>[number],
nodesById: ReadonlyMap<
string,
ReturnType<typeof flattenWorkflowNodes>[number]
>
): boolean {
if (isInactiveMode(node.mode)) return false
for (const ancestorId of getParentExecutionIds(String(node.id))) {
const ancestor = nodesById.get(ancestorId)
if (isInactiveMode(ancestor?.mode)) return false
}
return true
}
interface AssetVerifier {
updateModelsForNodeType: (nodeType: string) => Promise<void>
getAssets: (nodeType: string) => AssetItem[] | undefined
}
export async function verifyAssetSupportedCandidates(
candidates: MissingModelCandidate[],
signal?: AbortSignal,
assetsStore?: AssetVerifier
): Promise<void> {
if (signal?.aborted) return
const pendingCandidates = candidates.filter(
(c) => c.isAssetSupported && c.isMissing === undefined
)
if (pendingCandidates.length === 0) return
const pendingNodeTypes = new Set(
pendingCandidates.map((candidate) => candidate.nodeType)
)
const store =
assetsStore ?? (await import('@/stores/assetsStore')).useAssetsStore()
const failedNodeTypes = new Set<string>()
await Promise.allSettled(
[...pendingNodeTypes].map(async (nodeType) => {
if (signal?.aborted) return
try {
await store.updateModelsForNodeType(nodeType)
} catch (err) {
failedNodeTypes.add(nodeType)
console.warn(
`[Missing Model Pipeline] Failed to load assets for ${nodeType}:`,
err
)
}
})
)
if (signal?.aborted) return
for (const c of candidates) {
if (!c.isAssetSupported || c.isMissing !== undefined) continue
if (failedNodeTypes.has(c.nodeType)) continue
const assets = store.getAssets(c.nodeType) ?? []
c.isMissing = !isAssetInstalled(c, assets)
}
}
function normalizePath(path: string): string {
return path.replace(/\\/g, '/')
}
function isAssetInstalled(
candidate: MissingModelCandidate,
assets: AssetItem[]
): boolean {
if (candidate.hash && candidate.hashType) {
const candidateHash = `${candidate.hashType}:${candidate.hash}`
if (assets.some((a) => a.hash === candidateHash)) return true
}
const normalizedName = normalizePath(candidate.name)
return assets.some((a) => {
const f = normalizePath(getAssetFilename(a))
return f === normalizedName || f.endsWith('/' + normalizedName)
})
}
export function groupCandidatesByName(
candidates: MissingModelCandidate[]
): MissingModelViewModel[] {
const map = new Map<string, MissingModelViewModel>()
for (const c of candidates) {
const existing = map.get(c.name)
if (existing) {
if (c.nodeId) {
existing.referencingNodes.push({
nodeId: c.nodeId,
widgetName: c.widgetName
})
}
} else {
map.set(c.name, {
name: c.name,
representative: c,
referencingNodes: c.nodeId
? [{ nodeId: c.nodeId, widgetName: c.widgetName }]
: []
})
}
}
return Array.from(map.values())
}