You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit ae70bce
Browse filesBrowse the repository at this point in the historyBrowse files
row.adapter=awaitp.evaluate(async()=>{consta=awaitnavigator.gpu.requestAdapter();if(!a)throwError('No actual adapter');return{vendor:a.info.vendor,architecture:a.info.architecture,isFallbackAdapter:a.info.isFallbackAdapter};});
8
+
if(variant==='swiftshader'&&row.adapter.architecture!=='swiftshader')throwError('Not actual SwiftShader');
9
+
if(variant==='default'&&row.adapter.vendor!=='apple')throwError('Not actual Apple adapter');
10
+
awaitp.evaluate(asyncmodelId=>{const{WebWorkerMLCEngine,prebuiltAppConfig}=awaitimport('/assets/lib-CSLy4uO1.js');window.__record=prebuiltAppConfig.model_list.find(x=>x.model_id===modelId);if(!window.__record)throwError('Model record missing');window.__engine=newWebWorkerMLCEngine(newWorker('/assets/webllm.worker-CMk6FPDh.js',{type:'module'}),{appConfig:prebuiltAppConfig});window.__loadError=null;window.__engine.reload(modelId,{context_window_size:2048}).then(()=>window.__loaded=true).catch(e=>window.__loadError=String(e));},modelId);
Primary Chromium source documents Metal as the default Apple WebGPU backend and SwiftShader as a Vulkan software adapter enabled with --enable-unsafe-webgpu and --use-webgpu-adapter=swiftshader: https://chromium.googlesource.com/chromium/src/+/main/gpu/command_buffer/service/webgpu_decoder_impl.cc . Actual preflight identified Apple metal-3 and Google swiftshader, both created devices.
4
+
5
+
Run default then SwiftShader, same Chromium version, same persistent cache profile, shipped WebWorkerMLCEngine and emitted worker, same Qwen2.5-3B-Instruct-q4f32_1-MLC model record and chat options. Freeze two prior minimal ISO-date-copy requests (23 then 22) including original messages/schema/temperature0/top_p.8/max512. No prompt tuning, CPU fallback or output substitution. This shorter fixed sequence differs from the original interleaved eight-source experiment, so reproduction and backend causality remain conditional on observed results. Require actual adapter identity and actual WebLLM engine completion. Record failures and close both contexts; bounded model initialization 600 seconds and per-request 300 seconds.
6
+
7
+
No native document edits are made. Date copying is diagnostic only, not summary/translation/rewrite acceptance, physical device coverage or a proposed prompt repair. Preserve model/lib record and raw outputs/timing/errors. Software GPU may be slow or unable to load; retain that outcome without silently selecting another runtime. No model promotion follows.
constr={scope:'Actual WebGPU adapter preflight only; no inference or device-matrix certification',source:'https://chromium.googlesource.com/chromium/src/+/main/gpu/command_buffer/service/webgpu_decoder_impl.cc',runs:[]};
0 commit comments