@kaminos/webgpu-inference-kit 0.1.2 → 0.1.4
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +3 -3
- package/package.json +1 -1
- package/src/kernel-profile.js +1 -1
- package/src/kimodo-route.js +100 -0
- package/src/moge-route.js +80 -0
- package/src/route-boundary.js +98 -0
- package/src/scheduler-backpressure.js +108 -0
- package/src/sf3d-route.js +55 -0
- package/src/sharp-route.js +55 -0
package/README.md
CHANGED
|
@@ -31,7 +31,7 @@ import {
|
|
|
31
31
|
- What route is being invoked, such as MoGE depth/normal or SHARP image-to-splat.
|
|
32
32
|
- Which browser WebGPU adapter, device features, limits, and timestamp capabilities were actually available.
|
|
33
33
|
- Which kernel/profile variant ran, and which stages are required for a useful runtime profile.
|
|
34
|
-
- How a route was scheduled: throughput mode, cooperative/yield posture, phase chunk sizes, submitted-work waits, and unsupported scheduler fields.
|
|
34
|
+
- How a route was scheduled: throughput mode, cooperative/yield posture, breathability spans, yield checkpoints, phase chunk sizes, submitted-work waits, and unsupported scheduler fields.
|
|
35
35
|
- Which artifacts went in and out, so downstream consumers can join routes without losing identity.
|
|
36
36
|
|
|
37
37
|
The immediate goal is practical composition inside Kaminos: MoGE can become a local geometry/depth route, SHARP can emit splat candidates, Kimodo can emit motion clips, SF3D can emit meshes, and pipeline/commoner code can consume those outputs through one route grammar. The longer-term opportunity is a browser-native inference runtime kit that makes future image generators, 3D generators, and possibly language-model routes easier to seat without rebuilding the same WebGPU plumbing from scratch.
|
|
@@ -46,7 +46,7 @@ So the intended stack is:
|
|
|
46
46
|
|
|
47
47
|
1. **Route boundary:** define callable browser-local inference routes with stable input/output roles.
|
|
48
48
|
2. **Runtime profile:** preserve adapter/device/kernel/stage identity for the run that actually happened.
|
|
49
|
-
3. **Scheduler/backpressure profile:** expose whether the route is throughput-oriented, cooperative, furnace-class, warm, cached,
|
|
49
|
+
3. **Scheduler/backpressure profile:** expose whether the route is throughput-oriented, cooperative, furnace-class, warm, cached, frame-tail-sensitive, and where it can honestly yield.
|
|
50
50
|
4. **Receipt and classification:** reject stale, fallback, partial, mismatched, or invalid route output before another system treats it as authoritative.
|
|
51
51
|
|
|
52
52
|
The fourth layer protects the first three. It should not swallow the whole story.
|
|
@@ -63,7 +63,7 @@ The fourth layer protects the first three. It should not swallow the whole story
|
|
|
63
63
|
- `createStagedSubmitProfile(input)`, `addStagedSubmitStage(profile, stage)`, `finishStagedSubmitProfile(profile)`, and `validateStagedSubmitProfile(profile)`: describe staged queue-submit timing in a way that can be compared across routes.
|
|
64
64
|
- `createKernelProfileMetadata(input)` and `createRouteKernelProfileMetadata(input)`: normalize kit version, kernel profile, commit, required stages, and timing-source metadata for route definitions and receipts.
|
|
65
65
|
- `createWebGpuRuntimeProfileInput(input)`, `createWebGpuRuntimeProfile(input)`, and `validateWebGpuRuntimeProfile(profile)`: combine effective backend identity, kernel metadata, staged profile, and route mode into one producer-side runtime profile object.
|
|
66
|
-
- `createWebGpuRouteSchedulerProfile(input)` and `validateWebGpuRouteSchedulerProfile(profile)`: preserve requested versus effective scheduling, including throughput/cooperative mode, route-specific phase chunk sizes, submitted-work waits, yield cadence, and unsupported fields.
|
|
66
|
+
- `createWebGpuRouteSchedulerProfile(input)` and `validateWebGpuRouteSchedulerProfile(profile)`: preserve requested versus effective scheduling, including throughput/cooperative mode, route-specific phase chunk sizes, submitted-work waits, yield cadence, breathability spans, yieldable checkpoints, non-preemptible GPU-submit spans, and unsupported fields.
|
|
67
67
|
- `createWebGpuRouteBackpressureProfile(input)` and `validateWebGpuRouteBackpressureProfile(profile)`: record visible-wait/furnace pressure, warm/cache posture, memory-sharing posture, and frame-tail impact.
|
|
68
68
|
- `defineTensorManifest(input)` and `validateTensorManifest(manifest)`: normalize tensor metadata including dtype sizes and byte lengths.
|
|
69
69
|
- `createWebGpuLocalRouteReceipt(input)`, `createWebGpuRouteReceiptFromArtifacts(input)`, `createRouteReceiptArtifacts(input)`, `finishAndValidateRouteProfile(input)`, `validateRouteReceipt(receipt)`, and `assertAuthoritativeRouteReceipt(receipt)`: shared receipt construction and validation helpers.
|
package/package.json
CHANGED
package/src/kernel-profile.js
CHANGED
package/src/kimodo-route.js
CHANGED
|
@@ -8,6 +8,10 @@ import {
|
|
|
8
8
|
createRouteReceiptInputArtifact,
|
|
9
9
|
createWebGpuRouteReceiptFromArtifacts,
|
|
10
10
|
} from './route-receipt-helper.js';
|
|
11
|
+
import {
|
|
12
|
+
createWebGpuRouteBackpressureProfile,
|
|
13
|
+
createWebGpuRouteSchedulerProfile,
|
|
14
|
+
} from './scheduler-backpressure.js';
|
|
11
15
|
|
|
12
16
|
export const KIMODO_TEXT_TO_MOTION_ROUTE_ID = 'kimodo.text-to-motion.webgpu-local.v0';
|
|
13
17
|
const KIMODO_MODEL_ID = 'NVIDIA/Kimodo-SOMA-RP-v1.1';
|
|
@@ -19,6 +23,100 @@ const OUTPUT_ROLES = [
|
|
|
19
23
|
{ key: 'filmstrip', role: 'filmstrip', required: false },
|
|
20
24
|
];
|
|
21
25
|
|
|
26
|
+
function createDefaultKimodoScheduler() {
|
|
27
|
+
return createWebGpuRouteSchedulerProfile({
|
|
28
|
+
requestedScheduler: {
|
|
29
|
+
mode: 'cooperative',
|
|
30
|
+
yieldMs: 4,
|
|
31
|
+
waitForSubmittedWorkDone: true,
|
|
32
|
+
phaseChunkSize: {
|
|
33
|
+
'text-embedding': 1,
|
|
34
|
+
'ddim-sampling': 1,
|
|
35
|
+
'fk-decode': 1,
|
|
36
|
+
'output-capture': 1,
|
|
37
|
+
},
|
|
38
|
+
},
|
|
39
|
+
effectiveScheduler: {
|
|
40
|
+
mode: 'cooperative',
|
|
41
|
+
yieldMs: 4,
|
|
42
|
+
waitForSubmittedWorkDone: true,
|
|
43
|
+
phaseChunkSize: {
|
|
44
|
+
'text-embedding': 1,
|
|
45
|
+
'ddim-sampling': 1,
|
|
46
|
+
'fk-decode': 1,
|
|
47
|
+
'output-capture': 1,
|
|
48
|
+
},
|
|
49
|
+
unsupportedFields: [],
|
|
50
|
+
},
|
|
51
|
+
verificationState: 'scheduler-unverified',
|
|
52
|
+
breathability: {
|
|
53
|
+
spans: [
|
|
54
|
+
{
|
|
55
|
+
name: 'text-embedding',
|
|
56
|
+
stage: 'text-embedding',
|
|
57
|
+
kind: 'external-bound',
|
|
58
|
+
interruptible: false,
|
|
59
|
+
canYieldBefore: true,
|
|
60
|
+
canYieldAfter: true,
|
|
61
|
+
},
|
|
62
|
+
{
|
|
63
|
+
name: 'ddim-sampling-loop',
|
|
64
|
+
stage: 'ddim-sampling',
|
|
65
|
+
kind: 'gpu-submit-loop',
|
|
66
|
+
interruptible: false,
|
|
67
|
+
canYieldBefore: true,
|
|
68
|
+
canYieldAfter: true,
|
|
69
|
+
nonInterruptibleReason: 'Each diffusion step submit is non-preemptible; cooperative yielding occurs between steps.',
|
|
70
|
+
metadata: { checkpointCadence: 'per-diffusion-step' },
|
|
71
|
+
},
|
|
72
|
+
{
|
|
73
|
+
name: 'fk-decode',
|
|
74
|
+
stage: 'fk-decode',
|
|
75
|
+
kind: 'cpu-bound',
|
|
76
|
+
interruptible: true,
|
|
77
|
+
canYieldBefore: true,
|
|
78
|
+
canYieldAfter: true,
|
|
79
|
+
},
|
|
80
|
+
{
|
|
81
|
+
name: 'output-capture',
|
|
82
|
+
stage: 'output-capture',
|
|
83
|
+
kind: 'readback-bound',
|
|
84
|
+
interruptible: false,
|
|
85
|
+
canYieldBefore: true,
|
|
86
|
+
canYieldAfter: true,
|
|
87
|
+
},
|
|
88
|
+
],
|
|
89
|
+
checkpoints: [
|
|
90
|
+
{
|
|
91
|
+
name: 'between-diffusion-steps',
|
|
92
|
+
kind: 'diffusion-step',
|
|
93
|
+
afterStage: 'ddim-sampling',
|
|
94
|
+
yieldable: true,
|
|
95
|
+
waitsForSubmittedWorkDone: true,
|
|
96
|
+
metadata: { cadence: 'per-step' },
|
|
97
|
+
},
|
|
98
|
+
{
|
|
99
|
+
name: 'after-output-capture',
|
|
100
|
+
kind: 'readback',
|
|
101
|
+
afterStage: 'output-capture',
|
|
102
|
+
yieldable: true,
|
|
103
|
+
waitsForSubmittedWorkDone: true,
|
|
104
|
+
},
|
|
105
|
+
],
|
|
106
|
+
notes: 'Kimodo can expose useful cooperative pressure between diffusion steps; each submitted step remains non-preemptible.',
|
|
107
|
+
},
|
|
108
|
+
});
|
|
109
|
+
}
|
|
110
|
+
|
|
111
|
+
function createDefaultKimodoBackpressure() {
|
|
112
|
+
return createWebGpuRouteBackpressureProfile({
|
|
113
|
+
requestedBudget: 'visible-wait',
|
|
114
|
+
effectiveBudget: 'visible-wait',
|
|
115
|
+
memoryExclusivity: 'shared',
|
|
116
|
+
warmCacheState: 'unknown',
|
|
117
|
+
});
|
|
118
|
+
}
|
|
119
|
+
|
|
22
120
|
export function createKimodoTextToMotionRouteReceipt(input) {
|
|
23
121
|
if (!input || typeof input !== 'object') throw new Error('input must be an object');
|
|
24
122
|
if (!input.input?.artifactId || !input.input?.sha256) {
|
|
@@ -75,6 +173,8 @@ export function createKimodoTextToMotionRouteDefinition(input = {}) {
|
|
|
75
173
|
requiredFeatures: input.requiredFeatures || [],
|
|
76
174
|
requiredStages: routeMetadata.requiredStages,
|
|
77
175
|
timingSource: routeMetadata.timingSource,
|
|
176
|
+
scheduler: input.scheduler || createDefaultKimodoScheduler(),
|
|
177
|
+
backpressure: input.backpressure || createDefaultKimodoBackpressure(),
|
|
78
178
|
worker: input.worker || {
|
|
79
179
|
exportName: 'runKimodoTextToMotionRoute',
|
|
80
180
|
textEmbedding: 'external-llama3-8b',
|
package/src/moge-route.js
CHANGED
|
@@ -8,6 +8,10 @@ import {
|
|
|
8
8
|
createRouteReceiptInputArtifact,
|
|
9
9
|
createWebGpuRouteReceiptFromArtifacts,
|
|
10
10
|
} from './route-receipt-helper.js';
|
|
11
|
+
import {
|
|
12
|
+
createWebGpuRouteBackpressureProfile,
|
|
13
|
+
createWebGpuRouteSchedulerProfile,
|
|
14
|
+
} from './scheduler-backpressure.js';
|
|
11
15
|
|
|
12
16
|
export const MOGE_DEPTH_NORMAL_ROUTE_ID = 'moge.depth-normal.webgpu-local.v0';
|
|
13
17
|
const MOGE_MODEL_ID = 'Ruicheng/moge-2-vitl-normal';
|
|
@@ -20,6 +24,80 @@ const OUTPUT_ROLES = [
|
|
|
20
24
|
{ key: 'mask', role: 'mask', required: false },
|
|
21
25
|
];
|
|
22
26
|
|
|
27
|
+
function createDefaultMogeScheduler() {
|
|
28
|
+
return createWebGpuRouteSchedulerProfile({
|
|
29
|
+
requestedScheduler: {
|
|
30
|
+
mode: 'cooperative',
|
|
31
|
+
yieldMs: 4,
|
|
32
|
+
waitForSubmittedWorkDone: true,
|
|
33
|
+
phaseChunkSize: {
|
|
34
|
+
backbone: 1,
|
|
35
|
+
'decoder-heads': 1,
|
|
36
|
+
'output-readback': 1,
|
|
37
|
+
},
|
|
38
|
+
},
|
|
39
|
+
effectiveScheduler: {
|
|
40
|
+
mode: 'cooperative',
|
|
41
|
+
yieldMs: 4,
|
|
42
|
+
waitForSubmittedWorkDone: true,
|
|
43
|
+
phaseChunkSize: {
|
|
44
|
+
backbone: 1,
|
|
45
|
+
'decoder-heads': 1,
|
|
46
|
+
'output-readback': 1,
|
|
47
|
+
},
|
|
48
|
+
unsupportedFields: [],
|
|
49
|
+
},
|
|
50
|
+
verificationState: 'scheduler-unverified',
|
|
51
|
+
breathability: {
|
|
52
|
+
spans: [
|
|
53
|
+
{
|
|
54
|
+
name: 'backbone-submit',
|
|
55
|
+
stage: 'backbone',
|
|
56
|
+
kind: 'gpu-submit-bound',
|
|
57
|
+
interruptible: false,
|
|
58
|
+
canYieldBefore: true,
|
|
59
|
+
canYieldAfter: true,
|
|
60
|
+
nonInterruptibleReason: 'GPU command buffers cannot be preempted after submit',
|
|
61
|
+
},
|
|
62
|
+
{
|
|
63
|
+
name: 'decoder-heads-submit',
|
|
64
|
+
stage: 'decoder-heads',
|
|
65
|
+
kind: 'gpu-submit-bound',
|
|
66
|
+
interruptible: false,
|
|
67
|
+
canYieldBefore: true,
|
|
68
|
+
canYieldAfter: true,
|
|
69
|
+
nonInterruptibleReason: 'GPU command buffers cannot be preempted after submit',
|
|
70
|
+
},
|
|
71
|
+
{
|
|
72
|
+
name: 'output-readback',
|
|
73
|
+
stage: 'output-readback',
|
|
74
|
+
kind: 'readback-bound',
|
|
75
|
+
interruptible: false,
|
|
76
|
+
canYieldBefore: true,
|
|
77
|
+
canYieldAfter: true,
|
|
78
|
+
},
|
|
79
|
+
],
|
|
80
|
+
checkpoints: REQUIRED_STAGES.map(stage => ({
|
|
81
|
+
name: `after-${stage}`,
|
|
82
|
+
kind: stage === 'output-readback' ? 'readback' : 'stage-boundary',
|
|
83
|
+
afterStage: stage,
|
|
84
|
+
yieldable: true,
|
|
85
|
+
waitsForSubmittedWorkDone: true,
|
|
86
|
+
})),
|
|
87
|
+
notes: 'MoGE can cooperate between staged submits and readback, not inside a submitted GPU pass.',
|
|
88
|
+
},
|
|
89
|
+
});
|
|
90
|
+
}
|
|
91
|
+
|
|
92
|
+
function createDefaultMogeBackpressure() {
|
|
93
|
+
return createWebGpuRouteBackpressureProfile({
|
|
94
|
+
requestedBudget: 'visible-wait',
|
|
95
|
+
effectiveBudget: 'visible-wait',
|
|
96
|
+
memoryExclusivity: 'shared',
|
|
97
|
+
warmCacheState: 'unknown',
|
|
98
|
+
});
|
|
99
|
+
}
|
|
100
|
+
|
|
23
101
|
export function createMogeDepthNormalRouteReceipt(input) {
|
|
24
102
|
if (!input || typeof input !== 'object') throw new Error('input must be an object');
|
|
25
103
|
if (!input.input?.artifactId || !input.input?.sha256) {
|
|
@@ -77,6 +155,8 @@ export function createMogeDepthNormalRouteDefinition(input = {}) {
|
|
|
77
155
|
requiredFeatures: input.requiredFeatures || [],
|
|
78
156
|
requiredStages: routeMetadata.requiredStages,
|
|
79
157
|
timingSource: routeMetadata.timingSource,
|
|
158
|
+
scheduler: input.scheduler || createDefaultMogeScheduler(),
|
|
159
|
+
backpressure: input.backpressure || createDefaultMogeBackpressure(),
|
|
80
160
|
worker: input.worker || {
|
|
81
161
|
exportName: 'runMogeDepthNormalRoute',
|
|
82
162
|
},
|
package/src/route-boundary.js
CHANGED
|
@@ -3,6 +3,10 @@ import {
|
|
|
3
3
|
assertAuthoritativeRouteReceipt,
|
|
4
4
|
validateRouteReceipt,
|
|
5
5
|
} from './route-receipt.js';
|
|
6
|
+
import {
|
|
7
|
+
validateWebGpuRouteBackpressureProfile,
|
|
8
|
+
validateWebGpuRouteSchedulerProfile,
|
|
9
|
+
} from './scheduler-backpressure.js';
|
|
6
10
|
|
|
7
11
|
export const WEBGPU_ROUTE_DEFINITION_SCHEMA = 'kaminos.webgpu-route-definition.v0';
|
|
8
12
|
export const WEBGPU_ROUTE_REQUEST_SCHEMA = 'kaminos.webgpu-route-request.v0';
|
|
@@ -12,6 +16,18 @@ function clone(value) {
|
|
|
12
16
|
return value == null ? value : JSON.parse(JSON.stringify(value));
|
|
13
17
|
}
|
|
14
18
|
|
|
19
|
+
function stableJson(value) {
|
|
20
|
+
if (Array.isArray(value)) return `[${value.map(stableJson).join(',')}]`;
|
|
21
|
+
if (value && typeof value === 'object') {
|
|
22
|
+
return `{${Object.keys(value).sort().map(key => `${JSON.stringify(key)}:${stableJson(value[key])}`).join(',')}}`;
|
|
23
|
+
}
|
|
24
|
+
return JSON.stringify(value);
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
function deepEqual(a, b) {
|
|
28
|
+
return stableJson(a) === stableJson(b);
|
|
29
|
+
}
|
|
30
|
+
|
|
15
31
|
function isNonEmptyString(value) {
|
|
16
32
|
return typeof value === 'string' && value.trim().length > 0;
|
|
17
33
|
}
|
|
@@ -95,6 +111,49 @@ function optionalRoleNames(roles) {
|
|
|
95
111
|
return roles.filter(role => role.required === false).map(role => role.role);
|
|
96
112
|
}
|
|
97
113
|
|
|
114
|
+
function routeTimingStageNames(timings = {}) {
|
|
115
|
+
const names = new Set();
|
|
116
|
+
const addStageName = stage => {
|
|
117
|
+
if (isNonEmptyString(stage)) names.add(stage);
|
|
118
|
+
if (isNonEmptyString(stage?.name)) names.add(stage.name);
|
|
119
|
+
};
|
|
120
|
+
|
|
121
|
+
if (Array.isArray(timings.stages)) {
|
|
122
|
+
timings.stages.forEach(addStageName);
|
|
123
|
+
}
|
|
124
|
+
|
|
125
|
+
const profile = timings.profile;
|
|
126
|
+
if (profile && typeof profile === 'object') {
|
|
127
|
+
if (Array.isArray(profile.stageNames)) {
|
|
128
|
+
profile.stageNames.forEach(addStageName);
|
|
129
|
+
}
|
|
130
|
+
if (Array.isArray(profile.stages)) {
|
|
131
|
+
profile.stages.forEach(addStageName);
|
|
132
|
+
}
|
|
133
|
+
}
|
|
134
|
+
|
|
135
|
+
return names;
|
|
136
|
+
}
|
|
137
|
+
|
|
138
|
+
function validateRouteTiming(errors, receipt, route) {
|
|
139
|
+
const timings = receipt?.timings;
|
|
140
|
+
if (!timings || !route) return;
|
|
141
|
+
|
|
142
|
+
if (isNonEmptyString(route.timingSource) && timings.source !== route.timingSource) {
|
|
143
|
+
errors.push(`receipt.timings.source must be ${route.timingSource}`);
|
|
144
|
+
}
|
|
145
|
+
|
|
146
|
+
const requiredStages = Array.isArray(route.requiredStages) ? route.requiredStages : [];
|
|
147
|
+
if (requiredStages.length === 0) return;
|
|
148
|
+
|
|
149
|
+
const stageNames = routeTimingStageNames(timings);
|
|
150
|
+
for (const stageName of requiredStages) {
|
|
151
|
+
if (!stageNames.has(stageName)) {
|
|
152
|
+
errors.push(`receipt.timings missing required stage ${stageName}`);
|
|
153
|
+
}
|
|
154
|
+
}
|
|
155
|
+
}
|
|
156
|
+
|
|
98
157
|
function validateArtifacts(errors, artifacts, roles, path, { requireHash }) {
|
|
99
158
|
const knownRoles = roleSet(roles);
|
|
100
159
|
const artifactsByRole = new Map();
|
|
@@ -147,6 +206,8 @@ export function defineWebGpuRoute(input) {
|
|
|
147
206
|
requiredFeatures: Array.isArray(input.requiredFeatures) ? [...input.requiredFeatures].map(String).sort() : [],
|
|
148
207
|
requiredStages: Array.isArray(input.requiredStages) ? [...input.requiredStages] : [],
|
|
149
208
|
timingSource: input.timingSource || 'queue-submit-wait',
|
|
209
|
+
scheduler: clone(input.scheduler || null),
|
|
210
|
+
backpressure: clone(input.backpressure || null),
|
|
150
211
|
worker: clone(input.worker || null),
|
|
151
212
|
};
|
|
152
213
|
}
|
|
@@ -194,6 +255,16 @@ export function validateRouteDefinition(route) {
|
|
|
194
255
|
}
|
|
195
256
|
if (!isNonEmptyString(route.timingSource)) errors.push('timingSource must be a non-empty string');
|
|
196
257
|
|
|
258
|
+
if (route.scheduler != null) {
|
|
259
|
+
const schedulerResult = validateWebGpuRouteSchedulerProfile(route.scheduler);
|
|
260
|
+
if (!schedulerResult.ok) errors.push(...schedulerResult.errors.map(error => `scheduler.${error}`));
|
|
261
|
+
}
|
|
262
|
+
|
|
263
|
+
if (route.backpressure != null) {
|
|
264
|
+
const backpressureResult = validateWebGpuRouteBackpressureProfile(route.backpressure);
|
|
265
|
+
if (!backpressureResult.ok) errors.push(...backpressureResult.errors.map(error => `backpressure.${error}`));
|
|
266
|
+
}
|
|
267
|
+
|
|
197
268
|
return { ok: errors.length === 0, errors };
|
|
198
269
|
}
|
|
199
270
|
|
|
@@ -250,6 +321,8 @@ export function createRouteInvocationRequest(route, input) {
|
|
|
250
321
|
routeConfig: clone(input.routeConfig || {}),
|
|
251
322
|
model: clone(route.model),
|
|
252
323
|
kernel: clone(route.kernel),
|
|
324
|
+
scheduler: clone(route.scheduler || null),
|
|
325
|
+
backpressure: clone(route.backpressure || null),
|
|
253
326
|
createdAt: input.createdAt || new Date().toISOString(),
|
|
254
327
|
};
|
|
255
328
|
}
|
|
@@ -272,6 +345,30 @@ export function validateRouteInvocationRequest(request, route) {
|
|
|
272
345
|
validateArtifacts(errors, request.outputs, route.outputRoles, 'outputs', { requireHash: false });
|
|
273
346
|
}
|
|
274
347
|
|
|
348
|
+
if (request.scheduler != null) {
|
|
349
|
+
const schedulerResult = validateWebGpuRouteSchedulerProfile(request.scheduler);
|
|
350
|
+
if (!schedulerResult.ok) errors.push(...schedulerResult.errors.map(error => `scheduler.${error}`));
|
|
351
|
+
}
|
|
352
|
+
if (route?.scheduler != null) {
|
|
353
|
+
if (request.scheduler == null) {
|
|
354
|
+
errors.push('scheduler must match route definition');
|
|
355
|
+
} else if (!deepEqual(request.scheduler, route.scheduler)) {
|
|
356
|
+
errors.push('scheduler must match route definition');
|
|
357
|
+
}
|
|
358
|
+
}
|
|
359
|
+
|
|
360
|
+
if (request.backpressure != null) {
|
|
361
|
+
const backpressureResult = validateWebGpuRouteBackpressureProfile(request.backpressure);
|
|
362
|
+
if (!backpressureResult.ok) errors.push(...backpressureResult.errors.map(error => `backpressure.${error}`));
|
|
363
|
+
}
|
|
364
|
+
if (route?.backpressure != null) {
|
|
365
|
+
if (request.backpressure == null) {
|
|
366
|
+
errors.push('backpressure must match route definition');
|
|
367
|
+
} else if (!deepEqual(request.backpressure, route.backpressure)) {
|
|
368
|
+
errors.push('backpressure must match route definition');
|
|
369
|
+
}
|
|
370
|
+
}
|
|
371
|
+
|
|
275
372
|
return { ok: errors.length === 0, errors };
|
|
276
373
|
}
|
|
277
374
|
|
|
@@ -322,6 +419,7 @@ export function validateRouteWorkerResult(result, route) {
|
|
|
322
419
|
if (result.receipt.effectiveRouteId !== route.routeId) {
|
|
323
420
|
errors.push('receipt.effectiveRouteId must match route definition');
|
|
324
421
|
}
|
|
422
|
+
if (routeResult.ok) validateRouteTiming(errors, result.receipt, route);
|
|
325
423
|
}
|
|
326
424
|
|
|
327
425
|
const backendResult = validateWebGpuBackendIdentity(result.backend);
|
|
@@ -6,6 +6,24 @@ const VERIFICATION_STATES = new Set(['verified', 'scheduler-unverified', 'unsupp
|
|
|
6
6
|
const BUDGETS = new Set(['interactive', 'visible-wait', 'furnace', 'batch', 'unknown']);
|
|
7
7
|
const MEMORY_EXCLUSIVITY = new Set(['shared', 'exclusive', 'unknown']);
|
|
8
8
|
const WARM_CACHE_STATES = new Set(['cold', 'warm', 'hot', 'unknown']);
|
|
9
|
+
const BREATHABILITY_SPAN_KINDS = new Set([
|
|
10
|
+
'gpu-submit-bound',
|
|
11
|
+
'gpu-submit-loop',
|
|
12
|
+
'readback-bound',
|
|
13
|
+
'js-yieldable',
|
|
14
|
+
'cpu-bound',
|
|
15
|
+
'external-bound',
|
|
16
|
+
'unknown',
|
|
17
|
+
]);
|
|
18
|
+
const BREATHABILITY_CHECKPOINT_KINDS = new Set([
|
|
19
|
+
'pre-submit',
|
|
20
|
+
'post-submit',
|
|
21
|
+
'stage-boundary',
|
|
22
|
+
'diffusion-step',
|
|
23
|
+
'readback',
|
|
24
|
+
'external-callback',
|
|
25
|
+
'unknown',
|
|
26
|
+
]);
|
|
9
27
|
|
|
10
28
|
function clone(value) {
|
|
11
29
|
return value == null ? value : JSON.parse(JSON.stringify(value));
|
|
@@ -53,6 +71,41 @@ function normalizeEffectiveScheduler(input = {}, requestedScheduler) {
|
|
|
53
71
|
};
|
|
54
72
|
}
|
|
55
73
|
|
|
74
|
+
function normalizeBreathabilitySpan(input = {}) {
|
|
75
|
+
return {
|
|
76
|
+
name: input.name,
|
|
77
|
+
stage: input.stage || null,
|
|
78
|
+
kind: input.kind || 'unknown',
|
|
79
|
+
interruptible: Boolean(input.interruptible),
|
|
80
|
+
canYieldBefore: Boolean(input.canYieldBefore),
|
|
81
|
+
canYieldAfter: Boolean(input.canYieldAfter),
|
|
82
|
+
nonInterruptibleReason: input.nonInterruptibleReason || null,
|
|
83
|
+
metadata: isPlainObject(input.metadata) ? clone(input.metadata) : {},
|
|
84
|
+
};
|
|
85
|
+
}
|
|
86
|
+
|
|
87
|
+
function normalizeBreathabilityCheckpoint(input = {}) {
|
|
88
|
+
return {
|
|
89
|
+
name: input.name,
|
|
90
|
+
kind: input.kind || 'unknown',
|
|
91
|
+
beforeStage: input.beforeStage || null,
|
|
92
|
+
afterStage: input.afterStage || null,
|
|
93
|
+
yieldable: Boolean(input.yieldable),
|
|
94
|
+
waitsForSubmittedWorkDone: Boolean(input.waitsForSubmittedWorkDone),
|
|
95
|
+
metadata: isPlainObject(input.metadata) ? clone(input.metadata) : {},
|
|
96
|
+
};
|
|
97
|
+
}
|
|
98
|
+
|
|
99
|
+
function normalizeBreathability(input = {}) {
|
|
100
|
+
return {
|
|
101
|
+
spans: Array.isArray(input.spans) ? input.spans.map(normalizeBreathabilitySpan) : [],
|
|
102
|
+
checkpoints: Array.isArray(input.checkpoints)
|
|
103
|
+
? input.checkpoints.map(normalizeBreathabilityCheckpoint)
|
|
104
|
+
: [],
|
|
105
|
+
notes: input.notes || null,
|
|
106
|
+
};
|
|
107
|
+
}
|
|
108
|
+
|
|
56
109
|
function validateSchedulerShape(errors, scheduler, label) {
|
|
57
110
|
if (!isPlainObject(scheduler)) {
|
|
58
111
|
errors.push(`${label} must be an object`);
|
|
@@ -79,6 +132,59 @@ function validateSchedulerShape(errors, scheduler, label) {
|
|
|
79
132
|
}
|
|
80
133
|
}
|
|
81
134
|
|
|
135
|
+
function validateBreathability(errors, breathability) {
|
|
136
|
+
if (breathability == null) return;
|
|
137
|
+
if (!isPlainObject(breathability)) {
|
|
138
|
+
errors.push('breathability must be an object');
|
|
139
|
+
return;
|
|
140
|
+
}
|
|
141
|
+
if (!Array.isArray(breathability.spans)) {
|
|
142
|
+
errors.push('breathability.spans must be an array');
|
|
143
|
+
} else {
|
|
144
|
+
breathability.spans.forEach((span, index) => {
|
|
145
|
+
const path = `breathability.spans[${index}]`;
|
|
146
|
+
if (!isNonEmptyString(span?.name)) errors.push(`${path}.name must be a non-empty string`);
|
|
147
|
+
if (span?.stage != null && !isNonEmptyString(span.stage)) errors.push(`${path}.stage must be null or a non-empty string`);
|
|
148
|
+
if (!BREATHABILITY_SPAN_KINDS.has(span?.kind)) errors.push(`${path}.kind has unsupported value`);
|
|
149
|
+
if (typeof span?.interruptible !== 'boolean') errors.push(`${path}.interruptible must be a boolean`);
|
|
150
|
+
if (typeof span?.canYieldBefore !== 'boolean') errors.push(`${path}.canYieldBefore must be a boolean`);
|
|
151
|
+
if (typeof span?.canYieldAfter !== 'boolean') errors.push(`${path}.canYieldAfter must be a boolean`);
|
|
152
|
+
if (span?.nonInterruptibleReason != null && !isNonEmptyString(span.nonInterruptibleReason)) {
|
|
153
|
+
errors.push(`${path}.nonInterruptibleReason must be null or a non-empty string`);
|
|
154
|
+
}
|
|
155
|
+
if (!isPlainObject(span?.metadata)) errors.push(`${path}.metadata must be an object`);
|
|
156
|
+
if ((span?.kind === 'gpu-submit-bound' || span?.kind === 'gpu-submit-loop') && span.interruptible) {
|
|
157
|
+
errors.push(`${path}.${span.kind} cannot be interruptible after GPU submit`);
|
|
158
|
+
}
|
|
159
|
+
});
|
|
160
|
+
}
|
|
161
|
+
|
|
162
|
+
if (!Array.isArray(breathability.checkpoints)) {
|
|
163
|
+
errors.push('breathability.checkpoints must be an array');
|
|
164
|
+
} else {
|
|
165
|
+
breathability.checkpoints.forEach((checkpoint, index) => {
|
|
166
|
+
const path = `breathability.checkpoints[${index}]`;
|
|
167
|
+
if (!isNonEmptyString(checkpoint?.name)) errors.push(`${path}.name must be a non-empty string`);
|
|
168
|
+
if (!BREATHABILITY_CHECKPOINT_KINDS.has(checkpoint?.kind)) errors.push(`${path}.kind has unsupported value`);
|
|
169
|
+
if (checkpoint?.beforeStage != null && !isNonEmptyString(checkpoint.beforeStage)) {
|
|
170
|
+
errors.push(`${path}.beforeStage must be null or a non-empty string`);
|
|
171
|
+
}
|
|
172
|
+
if (checkpoint?.afterStage != null && !isNonEmptyString(checkpoint.afterStage)) {
|
|
173
|
+
errors.push(`${path}.afterStage must be null or a non-empty string`);
|
|
174
|
+
}
|
|
175
|
+
if (typeof checkpoint?.yieldable !== 'boolean') errors.push(`${path}.yieldable must be a boolean`);
|
|
176
|
+
if (typeof checkpoint?.waitsForSubmittedWorkDone !== 'boolean') {
|
|
177
|
+
errors.push(`${path}.waitsForSubmittedWorkDone must be a boolean`);
|
|
178
|
+
}
|
|
179
|
+
if (!isPlainObject(checkpoint?.metadata)) errors.push(`${path}.metadata must be an object`);
|
|
180
|
+
});
|
|
181
|
+
}
|
|
182
|
+
|
|
183
|
+
if (breathability.notes != null && !isNonEmptyString(breathability.notes)) {
|
|
184
|
+
errors.push('breathability.notes must be null or a non-empty string');
|
|
185
|
+
}
|
|
186
|
+
}
|
|
187
|
+
|
|
82
188
|
function missingUnsupportedField(effectiveScheduler, field) {
|
|
83
189
|
return !effectiveScheduler.unsupportedFields.includes(field)
|
|
84
190
|
&& !effectiveScheduler.unsupportedFields.includes('phaseChunkSize');
|
|
@@ -98,6 +204,7 @@ export function createWebGpuRouteSchedulerProfile(input = {}) {
|
|
|
98
204
|
requestedScheduler,
|
|
99
205
|
effectiveScheduler,
|
|
100
206
|
verificationState,
|
|
207
|
+
breathability: normalizeBreathability(input.breathability),
|
|
101
208
|
};
|
|
102
209
|
|
|
103
210
|
const result = validateWebGpuRouteSchedulerProfile(profile);
|
|
@@ -115,6 +222,7 @@ export function validateWebGpuRouteSchedulerProfile(profile) {
|
|
|
115
222
|
}
|
|
116
223
|
validateSchedulerShape(errors, profile.requestedScheduler, 'requestedScheduler');
|
|
117
224
|
validateSchedulerShape(errors, profile.effectiveScheduler, 'effectiveScheduler');
|
|
225
|
+
validateBreathability(errors, profile.breathability);
|
|
118
226
|
|
|
119
227
|
if (!Array.isArray(profile.effectiveScheduler?.unsupportedFields)) {
|
|
120
228
|
errors.push('effectiveScheduler.unsupportedFields must be an array');
|
package/src/sf3d-route.js
CHANGED
|
@@ -8,6 +8,10 @@ import {
|
|
|
8
8
|
createRouteReceiptInputArtifact,
|
|
9
9
|
createWebGpuRouteReceiptFromArtifacts,
|
|
10
10
|
} from './route-receipt-helper.js';
|
|
11
|
+
import {
|
|
12
|
+
createWebGpuRouteBackpressureProfile,
|
|
13
|
+
createWebGpuRouteSchedulerProfile,
|
|
14
|
+
} from './scheduler-backpressure.js';
|
|
11
15
|
|
|
12
16
|
export const SF3D_IMAGE_TO_MESH_ROUTE_ID = 'sf3d.image-to-mesh.webgpu-local.v0';
|
|
13
17
|
const SF3D_MODEL_ID = 'stabilityai/stable-fast-3d';
|
|
@@ -28,6 +32,55 @@ const OUTPUT_ROLES = [
|
|
|
28
32
|
{ key: 'meshObj', role: 'mesh-obj', required: false },
|
|
29
33
|
];
|
|
30
34
|
|
|
35
|
+
function createDefaultSf3dScheduler() {
|
|
36
|
+
return createWebGpuRouteSchedulerProfile({
|
|
37
|
+
requestedScheduler: {
|
|
38
|
+
mode: 'throughput',
|
|
39
|
+
yieldMs: 0,
|
|
40
|
+
waitForSubmittedWorkDone: true,
|
|
41
|
+
phaseChunkSize: {},
|
|
42
|
+
},
|
|
43
|
+
effectiveScheduler: {
|
|
44
|
+
mode: 'throughput',
|
|
45
|
+
yieldMs: 0,
|
|
46
|
+
waitForSubmittedWorkDone: true,
|
|
47
|
+
phaseChunkSize: {},
|
|
48
|
+
unsupportedFields: [],
|
|
49
|
+
},
|
|
50
|
+
verificationState: 'scheduler-unverified',
|
|
51
|
+
breathability: {
|
|
52
|
+
spans: REQUIRED_STAGES.map(stage => ({
|
|
53
|
+
name: `${stage}-phase`,
|
|
54
|
+
stage,
|
|
55
|
+
kind: stage === 'glb-export' ? 'cpu-bound' : 'gpu-submit-bound',
|
|
56
|
+
interruptible: false,
|
|
57
|
+
canYieldBefore: true,
|
|
58
|
+
canYieldAfter: true,
|
|
59
|
+
nonInterruptibleReason: stage === 'glb-export'
|
|
60
|
+
? null
|
|
61
|
+
: 'SF3D browser phases are treated as non-preemptible GPU work until finer receipts prove step boundaries.',
|
|
62
|
+
})),
|
|
63
|
+
checkpoints: REQUIRED_STAGES.map(stage => ({
|
|
64
|
+
name: `after-${stage}`,
|
|
65
|
+
kind: 'stage-boundary',
|
|
66
|
+
afterStage: stage,
|
|
67
|
+
yieldable: true,
|
|
68
|
+
waitsForSubmittedWorkDone: stage !== 'glb-export',
|
|
69
|
+
})),
|
|
70
|
+
notes: 'SF3D remains furnace-class until attention/triplane/marching phases expose finer cooperative boundaries.',
|
|
71
|
+
},
|
|
72
|
+
});
|
|
73
|
+
}
|
|
74
|
+
|
|
75
|
+
function createDefaultSf3dBackpressure() {
|
|
76
|
+
return createWebGpuRouteBackpressureProfile({
|
|
77
|
+
requestedBudget: 'visible-wait',
|
|
78
|
+
effectiveBudget: 'furnace',
|
|
79
|
+
memoryExclusivity: 'exclusive',
|
|
80
|
+
warmCacheState: 'unknown',
|
|
81
|
+
});
|
|
82
|
+
}
|
|
83
|
+
|
|
31
84
|
export function createSf3dImageToMeshRouteReceipt(input) {
|
|
32
85
|
if (!input || typeof input !== 'object') throw new Error('input must be an object');
|
|
33
86
|
if (!input.input?.artifactId || !input.input?.sha256) {
|
|
@@ -86,6 +139,8 @@ export function createSf3dImageToMeshRouteDefinition(input = {}) {
|
|
|
86
139
|
requiredFeatures: input.requiredFeatures || [],
|
|
87
140
|
requiredStages: routeMetadata.requiredStages,
|
|
88
141
|
timingSource: routeMetadata.timingSource,
|
|
142
|
+
scheduler: input.scheduler || createDefaultSf3dScheduler(),
|
|
143
|
+
backpressure: input.backpressure || createDefaultSf3dBackpressure(),
|
|
89
144
|
worker: input.worker || {
|
|
90
145
|
exportName: 'runSf3dImageToMeshRoute',
|
|
91
146
|
meshFormat: 'glb',
|
package/src/sharp-route.js
CHANGED
|
@@ -8,6 +8,10 @@ import {
|
|
|
8
8
|
createRouteReceiptInputArtifact,
|
|
9
9
|
createWebGpuRouteReceiptFromArtifacts,
|
|
10
10
|
} from './route-receipt-helper.js';
|
|
11
|
+
import {
|
|
12
|
+
createWebGpuRouteBackpressureProfile,
|
|
13
|
+
createWebGpuRouteSchedulerProfile,
|
|
14
|
+
} from './scheduler-backpressure.js';
|
|
11
15
|
|
|
12
16
|
export const SHARP_IMAGE_TO_SPLAT_ROUTE_ID = 'sharp.image-to-splat.webgpu-local.v0';
|
|
13
17
|
const SHARP_MODEL_ID = 'apple/ml-sharp';
|
|
@@ -20,6 +24,55 @@ const OUTPUT_ROLES = [
|
|
|
20
24
|
{ key: 'autoCropEvidence', role: 'splat-autocrop-evidence', required: false },
|
|
21
25
|
];
|
|
22
26
|
|
|
27
|
+
function createDefaultSharpScheduler() {
|
|
28
|
+
return createWebGpuRouteSchedulerProfile({
|
|
29
|
+
requestedScheduler: {
|
|
30
|
+
mode: 'throughput',
|
|
31
|
+
yieldMs: 0,
|
|
32
|
+
waitForSubmittedWorkDone: true,
|
|
33
|
+
phaseChunkSize: {},
|
|
34
|
+
},
|
|
35
|
+
effectiveScheduler: {
|
|
36
|
+
mode: 'throughput',
|
|
37
|
+
yieldMs: 0,
|
|
38
|
+
waitForSubmittedWorkDone: true,
|
|
39
|
+
phaseChunkSize: {},
|
|
40
|
+
unsupportedFields: [],
|
|
41
|
+
},
|
|
42
|
+
verificationState: 'scheduler-unverified',
|
|
43
|
+
breathability: {
|
|
44
|
+
spans: REQUIRED_STAGES.map(stage => ({
|
|
45
|
+
name: `${stage}-phase`,
|
|
46
|
+
stage,
|
|
47
|
+
kind: stage === 'output-capture' ? 'readback-bound' : 'gpu-submit-bound',
|
|
48
|
+
interruptible: false,
|
|
49
|
+
canYieldBefore: true,
|
|
50
|
+
canYieldAfter: true,
|
|
51
|
+
nonInterruptibleReason: stage === 'output-capture'
|
|
52
|
+
? null
|
|
53
|
+
: 'Browser WebGPU cannot preempt a submitted SHARP adapter phase.',
|
|
54
|
+
})),
|
|
55
|
+
checkpoints: REQUIRED_STAGES.map(stage => ({
|
|
56
|
+
name: `after-${stage}`,
|
|
57
|
+
kind: stage === 'output-capture' ? 'readback' : 'stage-boundary',
|
|
58
|
+
afterStage: stage,
|
|
59
|
+
yieldable: true,
|
|
60
|
+
waitsForSubmittedWorkDone: true,
|
|
61
|
+
})),
|
|
62
|
+
notes: 'SHARP is furnace-class until finer adapter receipts prove smaller cooperative boundaries.',
|
|
63
|
+
},
|
|
64
|
+
});
|
|
65
|
+
}
|
|
66
|
+
|
|
67
|
+
function createDefaultSharpBackpressure() {
|
|
68
|
+
return createWebGpuRouteBackpressureProfile({
|
|
69
|
+
requestedBudget: 'visible-wait',
|
|
70
|
+
effectiveBudget: 'furnace',
|
|
71
|
+
memoryExclusivity: 'exclusive',
|
|
72
|
+
warmCacheState: 'unknown',
|
|
73
|
+
});
|
|
74
|
+
}
|
|
75
|
+
|
|
23
76
|
export function createSharpImageToSplatRouteReceipt(input) {
|
|
24
77
|
if (!input || typeof input !== 'object') throw new Error('input must be an object');
|
|
25
78
|
if (!input.input?.artifactId || !input.input?.sha256) {
|
|
@@ -78,6 +131,8 @@ export function createSharpImageToSplatRouteDefinition(input = {}) {
|
|
|
78
131
|
requiredFeatures: input.requiredFeatures || [],
|
|
79
132
|
requiredStages: routeMetadata.requiredStages,
|
|
80
133
|
timingSource: routeMetadata.timingSource,
|
|
134
|
+
scheduler: input.scheduler || createDefaultSharpScheduler(),
|
|
135
|
+
backpressure: input.backpressure || createDefaultSharpBackpressure(),
|
|
81
136
|
worker: input.worker || {
|
|
82
137
|
exportName: 'runSharpImageToSplatRoute',
|
|
83
138
|
adapterReportSchema: 'kaminos.sharp-webgpu-adapter-report.v0',
|