WebGPU<-->BigInt: Tensorflow.js - Initialization of backend webgpu failed TypeError: Cannot convert a BigInt value to a number

[josephrocca]

Version: Deno 1.39.4

The code below works in Deno with an earlier tfjs-webgpu-backend version (see import_map.json here), but I raised the tfjs version numbers to 4.16.0 as shown below, and that produces the error logs below. I'm guessing that this is because newer versions of tf.js have started using BigInts, and Deno hasn't implemented some WebGPU<-->BigInt stuff.

I've confirmed that the code below works fine on Chrome (Windows).

---

deno.json

{
  "importMap": "import_map.json"
}

import_map.json

{
  "imports": {
    "@tensorflow/tfjs-core": "https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-core@4.16.0/dist/tf-core.fesm.js",
    "@tensorflow/tfjs-backend-webgpu": "https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-backend-webgpu@4.16.0/dist/tf-backend-webgpu.fesm.js"
  }
}

main.js

import * as tf from "@tensorflow/tfjs-core";
import "@tensorflow/tfjs-backend-webgpu";

await tf.ready();
tf.randomGamma([2, 2], 1).print();

Error Logs

deno run --unstable --allow-net=cdn.jsdelivr.net main.js
Initialization of backend webgpu failed
TypeError: Cannot convert a BigInt value to a number
    at toNumber (ext:deno_webidl/00_webidl.js:98:11)
    at Array.<anonymous> (ext:deno_webidl/00_webidl.js:255:13)
    at webidl.converters.GPUSize64 (ext:deno_webgpu/01_webgpu.js:5173:42)
    at ext:deno_webidl/00_webidl.js:904:28
    at Array.GPUDeviceDescriptor (ext:deno_webidl/00_webidl.js:788:32)
    at GPUAdapter.requestDevice (ext:deno_webgpu/01_webgpu.js:407:36)
    at Object.factory (https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-backend-webgpu@4.16.0/dist/tf-backend-webgpu.fesm.js:2138:38)
    at async Engine.ready (https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-core@4.16.0/dist/tf-core.fesm.js:4248:29)
    at async file:///home/joe/Downloads/denotfjs/main.js:4:1
error: Uncaught (in promise) Error: Could not initialize any backends, all backend initializations failed.
        throw new Error(`Could not initialize any backends, all backend initializations ` +
              ^
    at Engine.ready (https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-core@4.16.0/dist/tf-core.fesm.js:4254:15)
    at async file:///home/joe/Downloads/denotfjs/main.js:4:1

[Mutefish0]

run into the same problem

[petamoriken]

The spec defines GPUSize64 as unsigned long long, which uses ConvertToInt, so it appears that bigint is not allowed. Is this a spec bug?

[mmastrac]

@josephrocca Could you print the output of this code from your repl in both Chrome and deno?

const adapter = await navigator.gpu.requestAdapter();
console.log(adapter.limits);

[josephrocca]

@mmastrac Sorry for delay - was waiting to get access to the Windows machine again (I have linux, which doesn't support webgpu on chrome). Here are the logs:

Chrome:

maxTextureDimension1D: 16384
maxTextureDimension2D: 16384
maxTextureDimension3D: 2048
maxTextureArrayLayers: 2048
maxBindGroups: 4
maxBindGroupsPlusVertexBuffers: 24
maxBindingsPerBindGroup: 1000
maxDynamicUniformBuffersPerPipelineLayout: 10
maxDynamicStorageBuffersPerPipelineLayout: 8
maxSampledTexturesPerShaderStage: 16
maxSamplersPerShaderStage: 16
maxStorageBuffersPerShaderStage: 10
maxStorageTexturesPerShaderStage: 8
maxUniformBuffersPerShaderStage: 12
maxUniformBufferBindingSize: 65536
maxStorageBufferBindingSize: 2147483644
minUniformBufferOffsetAlignment: 256
minStorageBufferOffsetAlignment: 256
maxVertexBuffers: 8
maxBufferSize: 2147483648
maxVertexAttributes: 16
maxVertexBufferArrayStride: 2048
maxInterStageShaderComponents: 112
maxInterStageShaderVariables: 28
maxColorAttachments: 8
maxColorAttachmentBytesPerSample: 128
maxComputeWorkgroupStorageSize: 32768
maxComputeInvocationsPerWorkgroup: 1024
maxComputeWorkgroupSizeX: 1024
maxComputeWorkgroupSizeY: 1024
maxComputeWorkgroupSizeZ: 64
maxComputeWorkgroupsPerDimension: 65535

Deno:

  maxTextureDimension1D: 16384,
  maxTextureDimension2D: 16384,
  maxTextureDimension3D: 2048,
  maxTextureArrayLayers: 2048,
  maxBindGroups: 8,
  maxBindingsPerBindGroup: 65535,
  maxBufferSize: 2147483647,
  maxDynamicUniformBuffersPerPipelineLayout: 12,
  maxDynamicStorageBuffersPerPipelineLayout: 0,
  maxSampledTexturesPerShaderStage: 16,
  maxSamplersPerShaderStage: 16,
  maxStorageBuffersPerShaderStage: 0,
  maxStorageTexturesPerShaderStage: 0,
  maxUniformBuffersPerShaderStage: 12,
  maxUniformBufferBindingSize: 65536,
  maxStorageBufferBindingSize: 0,
  minUniformBufferOffsetAlignment: 256,
  minStorageBufferOffsetAlignment: 256,
  maxVertexBuffers: 16,
  maxVertexAttributes: 16,
  maxVertexBufferArrayStride: 4294967295,
  maxInterStageShaderComponents: 124,
  maxComputeWorkgroupStorageSize: 0,
  maxComputeInvocationsPerWorkgroup: 0,
  maxComputeWorkgroupSizeX: 0,
  maxComputeWorkgroupSizeY: 0,
  maxComputeWorkgroupSizeZ: 0,
  maxComputeWorkgroupsPerDimension: 0

Edge:

maxTextureDimension1D: 16384
maxTextureDimension2D: 16384
maxTextureDimension3D: 2048
maxTextureArrayLayers: 2048
maxBindGroups: 4
maxBindGroupsPlusVertexBuffers: 24
maxBindingsPerBindGroup: 1000
maxDynamicUniformBuffersPerPipelineLayout: 10
maxDynamicStorageBuffersPerPipelineLayout: 8
maxSampledTexturesPerShaderStage: 16
maxSamplersPerShaderStage: 16
maxStorageBuffersPerShaderStage: 10
maxStorageTexturesPerShaderStage: 8
maxUniformBuffersPerShaderStage: 12
maxUniformBufferBindingSize: 65536
maxStorageBufferBindingSize: 2147483644
minUniformBufferOffsetAlignment: 256
minStorageBufferOffsetAlignment: 256
maxVertexBuffers: 8
maxBufferSize: 2147483648
maxVertexAttributes: 16
maxVertexBufferArrayStride: 2048
maxInterStageShaderComponents: 112
maxInterStageShaderVariables: 28
maxColorAttachments: 8
maxColorAttachmentBytesPerSample: 128
maxComputeWorkgroupStorageSize: 32768
maxComputeInvocationsPerWorkgroup: 1024
maxComputeWorkgroupSizeX: 1024
maxComputeWorkgroupSizeY: 1024
maxComputeWorkgroupSizeZ: 64
maxComputeWorkgroupsPerDimension: 65535

[mattvr]

Update: This error is now fixed in the latest version of tfjs

---

Tensorflow.js now seems to fail in a different way. GPUAdapter.requestAdapterInfo is called by tfjs, but this method is deprecated and not implemented by Deno.

Repro

I ran the original example, but with latest Deno & npm specifiers npm:@tensorflow/tfjs-core and npm:@tensorflow/tfjs-backend-webgpu.

(Note: this reproduces with Deno v2.0.0-rc.9 or 1.46.3 and latest tfjs using either npm: or jsdelivr)

$ deno run mod.ts
Initialization of backend webgpu failed
TypeError: adapter.requestAdapterInfo is not a function
    at [...]/deno/npm/registry.npmjs.org/@tensorflow/tfjs-backend-webgpu/4.21.0/dist/tf-backend-webgpu.node.js:2147:50
    at step ([...]/deno/npm/registry.npmjs.org/@tensorflow/tfjs-backend-webgpu/4.21.0/dist/tf-backend-webgpu.node.js:144:27)
    at Object.next ([...]/deno/npm/registry.npmjs.org/@tensorflow/tfjs-backend-webgpu/4.21.0/dist/tf-backend-webgpu.node.js:93:53)
    at fulfilled ([...]deno/npm/registry.npmjs.org/@tensorflow/tfjs-backend-webgpu/4.21.0/dist/tf-backend-webgpu.node.js:74:28)
error: Uncaught (in promise) Error: Could not initialize any backends, all backend initializations failed.
    at Engine.<anonymous> ([...]/deno/npm/registry.npmjs.org/@tensorflow/tfjs-core/4.21.0/dist/tf-core.node.js:4301:35)
    at step ([...]/deno/npm/registry.npmjs.org/@tensorflow/tfjs-core/4.21.0/dist/tf-core.node.js:138:27)

[fs-eire]

The following code failed with error Cannot convert a BigInt value to a number on 2 of my Windows11 machines.

adapter.requestDevice({
  requiredLimits : {
    maxBufferSize: adapter.limits.maxBufferSize
  }
});

value of adapter.limits.maxBufferSize is 18446744073709551615n.