File size: 13,264 Bytes
31c7d49
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
import { StrictMode, useEffect, useRef, useState } from 'react'
import { createRoot } from 'react-dom/client'

import { compareFloat32, type TensorComparison } from './models/triposplat/tensorMath'
import { OrtWorkerClient, type OrtWorkerStatus } from './runtime/OrtWorkerClient'
import { createTensorPayload } from './runtime/tensors'

const DEFAULT_MODEL = '/models/triposplat/dit_step.onnx'
const DEFAULT_FIXTURE = '/fixtures/generated/dit-step-fp32'
const SESSION_ID = 'triposplat/dit-parity'
const FP16_TOLERANCE = { absolute: 0.04, relative: 0.03, minimumCosineSimilarity: 0.9999 }
const FP32_TOLERANCE = { absolute: 0.00002, relative: 0.001, minimumCosineSimilarity: 0.999999999 }

const SHAPES = {
  latent: [1, 8192, 16],
  camera: [1, 1, 5],
  t: [1],
  feature1: [1, 4101, 1280],
  feature2: [1, 4101, 128],
  pred_latent: [1, 8192, 16],
  pred_camera: [1, 1, 5],
} as const

interface OutputGate extends TensorComparison {
  fractionWithinTolerance: number
  maxErrorIndex: number
  referenceAtMaxError: number
  candidateAtMaxError: number
  firstMismatches: Array<{ index: number; reference: number; candidate: number }>
  passed: boolean
}

interface DitLabResult {
  passed: boolean
  executionProvider: string
  modelLoadMs: number
  modelTransferBytes?: number
  inferenceMs: number
  readbackMs: number
  outputs?: {
    pred_latent: OutputGate
    pred_camera: OutputGate
  }
  probes?: Record<string, OutputGate>
  tolerance: { absolute: number; relative: number; minimumCosineSimilarity: number }
  environment: {
    userAgent: string
    crossOriginIsolated: boolean
    webgpu: boolean
  }
}

interface ProbeManifest {
  includeFinalOutputs?: boolean
  outputs: Array<{ name: string; path: string; shape: number[]; elements: number }>
}

declare global {
  interface Window {
    __TRIPOSPLAT_DIT_RESULT__?: DitLabResult
  }
}

async function fetchFloat32(url: string, expectedElements: number): Promise<Float32Array> {
  const response = await fetch(url)
  if (!response.ok) throw new Error(`Could not fetch ${url}: HTTP ${response.status}`)
  const buffer = await response.arrayBuffer()
  if (buffer.byteLength !== expectedElements * 4) {
    throw new Error(`${url} has ${buffer.byteLength} bytes; expected ${expectedElements * 4}.`)
  }
  return new Float32Array(buffer)
}

function elementCount(shape: readonly number[]): number {
  return shape.reduce((product, value) => product * value, 1)
}

type DitTolerance = typeof FP16_TOLERANCE

async function fixtureTolerance(fixtureUrl: string): Promise<DitTolerance> {
  const response = await fetch(`${fixtureUrl}/manifest.json`)
  if (response.status === 404) return FP16_TOLERANCE
  if (!response.ok) throw new Error(`Could not fetch fixture manifest: HTTP ${response.status}`)
  const manifest = await response.json() as { metadata?: { internal_precision?: unknown } }
  return manifest.metadata?.internal_precision === 'fp32' ? FP32_TOLERANCE : FP16_TOLERANCE
}

function gateOutput(
  reference: Float32Array,
  candidate: Float32Array,
  tolerance: DitTolerance,
): OutputGate {
  const comparison = compareFloat32(reference, candidate)
  let within = 0
  let maxErrorIndex = 0
  let maxError = -1
  const firstMismatches: OutputGate['firstMismatches'] = []
  for (let index = 0; index < reference.length; index += 1) {
    const error = Math.abs(reference[index] - candidate[index])
    const isWithin = error <= tolerance.absolute + tolerance.relative * Math.abs(reference[index])
    if (error > maxError) {
      maxError = error
      maxErrorIndex = index
    }
    if (isWithin) {
      within += 1
    } else if (firstMismatches.length < 8) {
      firstMismatches.push({ index, reference: reference[index], candidate: candidate[index] })
    }
  }
  const fractionWithinTolerance = within / reference.length
  return {
    ...comparison,
    fractionWithinTolerance,
    maxErrorIndex,
    referenceAtMaxError: reference[maxErrorIndex],
    candidateAtMaxError: candidate[maxErrorIndex],
    firstMismatches,
    passed: comparison.finite
      && fractionWithinTolerance === 1
      && comparison.cosineSimilarity >= tolerance.minimumCosineSimilarity,
  }
}

async function contentLength(url: string): Promise<number | undefined> {
  try {
    const response = await fetch(url, { method: 'HEAD' })
    if (!response.ok) return undefined
    const value = Number(response.headers.get('content-length'))
    return Number.isFinite(value) && value >= 0 ? value : undefined
  } catch {
    return undefined
  }
}

function initialUrl(parameter: string, fallback: string): string {
  return new URLSearchParams(window.location.search).get(parameter) ?? fallback
}

export function DitLab() {
  const clientRef = useRef<OrtWorkerClient | null>(null)
  const autoRunStartedRef = useRef(false)
  const runRef = useRef<() => Promise<void>>(async () => undefined)
  const [modelUrl, setModelUrl] = useState(() => initialUrl('model', DEFAULT_MODEL))
  const [fixtureUrl, setFixtureUrl] = useState(() => initialUrl('fixture', DEFAULT_FIXTURE))
  const [status, setStatus] = useState('Ready to validate one official DiT invocation.')
  const [busy, setBusy] = useState(false)
  const [error, setError] = useState<string | null>(null)
  const [result, setResult] = useState<DitLabResult | null>(null)

  useEffect(() => () => {
    const client = clientRef.current
    clientRef.current = null
    if (client) void client.dispose()
  }, [])

  const run = async () => {
    if (busy) return
    setBusy(true)
    setError(null)
    setResult(null)
    delete window.__TRIPOSPLAT_DIT_RESULT__
    let client: OrtWorkerClient | undefined
    try {
      if (clientRef.current) await clientRef.current.dispose()
      const onStatus = (event: OrtWorkerStatus) => setStatus(event.message)
      client = new OrtWorkerClient({ onStatus })
      clientRef.current = client
      setStatus('Fetching deterministic official one-step tensors…')
      const tolerance = await fixtureTolerance(fixtureUrl)
      const names = Object.keys(SHAPES) as Array<keyof typeof SHAPES>
      const tensors = Object.fromEntries(await Promise.all(names.map(async (name) => [
        name,
        await fetchFloat32(`${fixtureUrl}/${name}.f32`, elementCount(SHAPES[name])),
      ]))) as Record<keyof typeof SHAPES, Float32Array>
      let probeManifest: ProbeManifest | undefined
      const probeManifestResponse = await fetch(`${fixtureUrl}/probes.json`)
      if (
        probeManifestResponse.ok
        && probeManifestResponse.headers.get('content-type')?.includes('application/json')
      ) {
        probeManifest = await probeManifestResponse.json() as ProbeManifest
      } else if (!probeManifestResponse.ok && probeManifestResponse.status !== 404) {
        throw new Error(`Could not fetch probe manifest: HTTP ${probeManifestResponse.status}`)
      }
      const probeReferences = probeManifest
        ? Object.fromEntries(await Promise.all(probeManifest.outputs.map(async ({ name, path, elements }) => [
            name,
            await fetchFloat32(`${fixtureUrl}/${path}`, elements),
          ]))) as Record<string, Float32Array>
        : undefined

      const sidecarUrl = `${modelUrl}.data`
      const externalDataPath = new URL(modelUrl, document.baseURI).pathname.split('/').at(-1)
      if (!externalDataPath) throw new Error(`Could not derive external-data path from ${modelUrl}.`)
      const transferParts = await Promise.all([contentLength(modelUrl), contentLength(sidecarUrl)])
      const modelTransferBytes = transferParts.every((value) => value !== undefined)
        ? transferParts.reduce<number>((sum, value) => sum + (value ?? 0), 0)
        : undefined
      const loaded = await client.loadSession({
        sessionId: SESSION_ID,
        manifest: {
          graphUrl: modelUrl,
          externalData: [{ path: `${decodeURIComponent(externalDataPath)}.data`, url: sidecarUrl }],
        },
        options: {
          allowWasmFallback: false,
          graphOptimizationLevel: 'disabled',
        },
      })
      if (loaded.executionProvider !== 'webgpu') {
        throw new Error(`Expected WebGPU, loaded ${loaded.executionProvider}.`)
      }

      const response = await client.runSession({
        sessionId: SESSION_ID,
        inputs: {
          latent: createTensorPayload('float32', tensors.latent, SHAPES.latent),
          camera: createTensorPayload('float32', tensors.camera, SHAPES.camera),
          t: createTensorPayload('float32', tensors.t, SHAPES.t),
          feature1: createTensorPayload('float32', tensors.feature1, SHAPES.feature1),
          feature2: createTensorPayload('float32', tensors.feature2, SHAPES.feature2),
        },
        outputs: [
          ...(probeManifest?.includeFinalOutputs === false ? [] : ['pred_latent', 'pred_camera']),
          ...(probeManifest?.outputs.map(({ name }) => name) ?? []),
        ],
        tag: 'official-one-step-parity',
      })
      const outputs = probeManifest?.includeFinalOutputs === false
        ? undefined
        : (() => {
            const predLatent = response.outputs.pred_latent
            const predCamera = response.outputs.pred_camera
            if (!predLatent || predLatent.type !== 'float32' || !predCamera || predCamera.type !== 'float32') {
              throw new Error('DiT did not return both public float32 outputs.')
            }
            return {
              pred_latent: gateOutput(tensors.pred_latent, predLatent.data, tolerance),
              pred_camera: gateOutput(tensors.pred_camera, predCamera.data, tolerance),
            }
          })()
      const probes = probeReferences
        ? Object.fromEntries(Object.entries(probeReferences).map(([name, reference]) => {
            const payload = response.outputs[name]
            if (!payload || payload.type !== 'float32') {
              throw new Error(`DiT probe '${name}' did not return float32 data.`)
            }
            return [name, gateOutput(reference, payload.data, tolerance)]
          }))
        : undefined
      const next: DitLabResult = {
        passed: (outputs === undefined || (outputs.pred_latent.passed && outputs.pred_camera.passed))
          && (probes === undefined || Object.values(probes).every(({ passed }) => passed)),
        executionProvider: loaded.executionProvider,
        modelLoadMs: loaded.loadMs,
        modelTransferBytes,
        inferenceMs: response.timings.inferenceMs,
        readbackMs: response.timings.readbackMs,
        outputs,
        probes,
        tolerance,
        environment: {
          userAgent: navigator.userAgent,
          crossOriginIsolated: self.crossOriginIsolated,
          webgpu: 'gpu' in navigator,
        },
      }
      window.__TRIPOSPLAT_DIT_RESULT__ = next
      setResult(next)
      setStatus(next.passed ? 'PASS: one WebGPU DiT invocation matches official PyTorch.' : 'FAIL: WebGPU DiT output drift exceeds tolerance.')
    } catch (caught) {
      const message = caught instanceof Error ? caught.message : String(caught)
      setError(message)
      setStatus('DiT validation failed.')
    } finally {
      setBusy(false)
    }
  }

  runRef.current = run
  useEffect(() => {
    const timeout = window.setTimeout(() => {
      if (
        new URLSearchParams(window.location.search).get('autorun') === '1'
        && !autoRunStartedRef.current
      ) {
        autoRunStartedRef.current = true
        void runRef.current()
      }
    }, 0)
    return () => window.clearTimeout(timeout)
  }, [])

  return (
    <main>
      <h1>TripoSplat DiT · WebGPU parity</h1>
      <p>Loads the query-chunked official one-step graph and compares one browser invocation with an untouched PyTorch fixture.</p>
      <label>ONNX graph <input value={modelUrl} onChange={(event) => setModelUrl(event.target.value)} /></label>
      <label>Fixture directory <input value={fixtureUrl} onChange={(event) => setFixtureUrl(event.target.value)} /></label>
      <button type="button" disabled={busy} onClick={() => void run()}>{busy ? 'Running…' : 'Run DiT parity gate'}</button>
      <p role="status" data-testid="dit-status">{status}</p>
      {error ? <pre className="error" data-testid="dit-error">{error}</pre> : null}
      {result ? <pre data-testid="dit-result">{JSON.stringify(result, null, 2)}</pre> : null}
    </main>
  )
}

const style = document.createElement('style')
style.textContent = `
  :root { color: #ececf3; background: #101014; font: 15px/1.5 ui-monospace, SFMono-Regular, Menlo, monospace; }
  body { margin: 0; }
  main { max-width: 920px; margin: 0 auto; padding: 48px 24px; }
  h1 { font: 600 28px/1.2 system-ui, sans-serif; }
  label { display: grid; gap: 6px; margin: 18px 0; }
  input { box-sizing: border-box; width: 100%; padding: 10px; color: inherit; background: #1b1b22; border: 1px solid #3a3a48; border-radius: 6px; }
  button { padding: 10px 16px; color: #08080a; background: #f8cf00; border: 0; border-radius: 6px; font-weight: 700; cursor: pointer; }
  button:disabled { opacity: .55; cursor: wait; }
  pre { overflow: auto; padding: 16px; background: #18181f; border-radius: 8px; }
  .error { color: #ff9b9b; }
`
document.head.append(style)

createRoot(document.getElementById('root')!).render(
  <StrictMode>
    <DitLab />
  </StrictMode>,
)