[BUG] Find an error initializing the WebGPU device OperationError: D3D12

[0dminnimda]

While running on the candy chrome I got those errors: (firstly the Init error and on the second request the rest)

[System Initalize] Initialize GPU device: WebGPU - intel
[System Initalize] Fetching param cache[71/163]: 1765MB fetched. 43% completed, 656 secs elapsed. It can take a while when we first visit this page to populate the cache. Later refreshes will become faster.
Init error, OperationError: The operation failed for an operation-specific reason
Find an error initializing the WebGPU device OperationError: D3D12 create command queue failed with DXGI_ERROR_DEVICE_REMOVED (0x887A0005) at CheckHRESULTImpl (..\..\third_party\dawn\src\dawn\native\d3d\D3DError.cpp:94) at Initialize (..\..\third_party\dawn\src\dawn\native\d3d12\DeviceD3D12.cpp:84) at Create (..\..\third_party\dawn\src\dawn\native\d3d12\DeviceD3D12.cpp:69)
Init error, Error: Find an error initializing WebGPU: OperationError: D3D12 create command queue failed with DXGI_ERROR_DEVICE_REMOVED (0x887A0005) at CheckHRESULTImpl (..\..\third_party\dawn\src\dawn\native\d3d\D3DError.cpp:94) at Initialize (..\..\third_party\dawn\src\dawn\native\d3d12\DeviceD3D12.cpp:84) at Create (..\..\third_party\dawn\src\dawn\native\d3d12\DeviceD3D12.cpp:69)

console:

ID3D12Device::CreateCommittedResource failed with DXGI_ERROR_DEVICE_REMOVED (0x887A0005)
    at CheckHRESULTImpl (..\..\third_party\dawn\src\dawn\native\d3d\D3DError.cpp:94)
    at CreateCommittedResource (..\..\third_party\dawn\src\dawn\native\d3d12\ResourceAllocatorManagerD3D12.cpp:565)
    at AllocateMemory (..\..\third_party\dawn\src\dawn\native\d3d12\ResourceAllocatorManagerD3D12.cpp:391)
    at Initialize (..\..\third_party\dawn\src\dawn\native\d3d12\BufferD3D12.cpp:158)
    at Create (..\..\third_party\dawn\src\dawn\native\d3d12\BufferD3D12.cpp:105)
    at CreateBuffer (..\..\third_party\dawn\src\dawn\native\Device.cpp:1508)

llm_chat.js:547 undefined
llm_chat.js:421 undefined
llm_chat.js:547 Error: Find an error initializing WebGPU: OperationError: D3D12 create command queue failed with DXGI_ERROR_DEVICE_REMOVED (0x887A0005)
    at CheckHRESULTImpl (..\..\third_party\dawn\src\dawn\native\d3d\D3DError.cpp:94)
    at Initialize (..\..\third_party\dawn\src\dawn\native\d3d12\DeviceD3D12.cpp:84)
    at Create (..\..\third_party\dawn\src\dawn\native\d3d12\DeviceD3D12.cpp:69)

    at #asyncInitTVM (https://mlc.ai/web-llm/dist/llm_chat.js:423:13)
    at async LLMChatInstance.asyncInit (https://mlc.ai/web-llm/dist/llm_chat.js:443:5)
    at async LLMChatInstance.generate (https://mlc.ai/web-llm/dist/llm_chat.js:544:7)
    at async tvmjsGlobalEnv.asyncOnGenerate (https://mlc.ai/web-llm/dist/llm_chat.js:606:3)

OS: Windows 10

[0dminnimda]

It always seems to be stopped at 71/72

[tqchen]

what is your GPU RAM limit, since we are definitely pressure testing GPU a bit and if you do not have more than 6.5G of RAM, it might cause problems

[0dminnimda]

[0dminnimda]

It seems likely that the intel graphics card will fail after about 1765MB when the cap in 2160MB, including the fact that it renders the monitors

[0dminnimda]

Although the candy chrome is using both it seems chrome://gpu:

<Integrated GPU> D3D12 backend - Intel(R) HD Graphics 4600
[WebGPU Status]
Available

...

<Discrete GPU> D3D12 backend - NVIDIA GeForce GT 750M
[WebGPU Status]
Available

[aseok]

facing this error while running on windows 10 x64 chrome Version 112.0.5615.121: Error: Find an error initializing WebGPU: OperationError: Required limit (1073741824) is greater than the supported limit (268435456).

• While validating maxBufferSize

• While validating required limits

  at #asyncInitTVM (llm_chat.js:423:13) at async LLMChatInstance.asyncInit (llm_chat.js:443:5) at async LLMChatInstance.generate (llm_chat.js:544:7) at async tvmjsGlobalEnv.asyncOnGenerate (llm_chat.js:606:3)

[tqchen]

@aseok this is related to #21 , you can try to upgrade your browser to chrome 113(beta)

[tqchen]

@0dminnimda Unfortunately year the card's memory is not enough for this model, since we need ~ 6G , we will work on bringing more optimizations and smaller models

[inspiredlabs]

Here's a quick fix if you're on macOS:

• download Canary.
• install it into /Applications.
• go to zsh, or whatever terminal yo prefer and run: "/Applications/Google Chrome Canary.app/Contents/MacOS/Google Chrome Canary" --enable-dawn-features=disable_robustness, nb: the quotes around the Filename Spaces.
• in the URL bar go to https://mlc.ai/web-llm/#chat-demo`
• it should take ~5 min to download the LLM into your browser cache.
• If you want something That Just Works™️, then try saving this script (vicuna.sh):

/Applications/Google\ Chrome\ Canary.app/Contents/MacOS/Google\ Chrome\ Canary --enable-dawn-features=disable_robustness --new-window "https://mlc.ai/web-llm/#chat-demo"

# don't forget to bless this file from the PWD: `chmod +x vicuna.sh`

• Before you double click on it from your Desktop/Applications etc. Bless it using: chmod +x vicuna.sh.