App and mobile checks

Workflow runs for App and mobile checks

• cuda qk_rows: head dims to 512, a weightless V norm and MLX's rounded rows

  currently running:
  App and mobile checks#750:Pull request#1071openedbyjschmied
  jschmied:cuda-qk-gemma

  Oct 19, 2026In progress
  More actions for workflow run 750

• cuda embed_rows: MLX affine tables at 2 to 8 bits and a bf16 scale

  currently running:
  App and mobile checks#749:Pull request#1070openedbyjschmied
  jschmied:cuda-embed-gemma

  Oct 19, 2026In progress
  More actions for workflow run 749

• cuda act_rows: GeGLU and SwiGLU in MLX's bf16 steps

  queued:
  App and mobile checks#748:Pull request#1068synchronizebyjschmied
  jschmied:cuda-act-gelu

  Oct 19, 2026Queued
  More actions for workflow run 748

• cuda norms: bf16 residual tails with a branch norm, a layer scalar and MLX's rounded rows

  completed successfully:
  App and mobile checks#747:Pull request#1056synchronizebyjschmied
  jschmied:cuda-norms-gemma

  Oct 19, 202629m 14s
  More actions for workflow run 747

• GLM-5.3-Flash (Metal): a tensor-tile kernel for the MTP draft head's eh_proj, on by default — ~5% prefill improvement

  completed successfully:
  App and mobile checks#746:Pull request#1066synchronizebyEugeneClaw
  EugeneClaw:eh-nax-ehproj-tensor-tile

  Oct 19, 202634m 27s
  More actions for workflow run 746

• cuda act_rows: GeGLU and SwiGLU in MLX's bf16 steps

  completed successfully:
  App and mobile checks#745:Pull request#1068synchronizebyjschmied
  jschmied:cuda-act-gelu

  Oct 19, 202636m 44s
  More actions for workflow run 745

• Flash Next CUDA: drafted windows of several streams verified in one pass

  completed successfully:
  App and mobile checks#744:Pull request#1069openedbyMirrdhyn
  hygurlabs:flashnext-cuda-shared-verify

  Oct 19, 20268m 47s
  More actions for workflow run 744

• GLM-5.3-Flash (Metal): a tensor-tile kernel for the MTP draft head's eh_proj, on by default — ~5% prefill improvement

  completed successfully:
  App and mobile checks#743:Pull request#1066synchronizebyEugeneClaw
  EugeneClaw:eh-nax-ehproj-tensor-tile

  Oct 19, 202610m 3s
  More actions for workflow run 743

• GLM-5.3-Flash (Metal): a tensor-tile kernel for the MTP draft head's eh_proj, on by default — ~5% prefill improvement

  completed successfully:
  App and mobile checks#742:Pull request#1066synchronizebyEugeneClaw
  EugeneClaw:eh-nax-ehproj-tensor-tile

  Oct 19, 20269m 35s
  More actions for workflow run 742

• GLM-5.3-Flash (Metal): a tensor-tile kernel for the MTP draft head's eh_proj, on by default — ~5% prefill improvement

  completed successfully:
  App and mobile checks#741:Pull request#1066synchronizebyEugeneClaw
  EugeneClaw:eh-nax-ehproj-tensor-tile

  Oct 19, 20268m 15s
  More actions for workflow run 741

• Flash Next CUDA: structured output and required tool calls

  completed successfully:
  App and mobile checks#740:Pull request#1062synchronizebyMirrdhyn
  hygurlabs:flashnext-cuda-structures

  Oct 19, 20269m 0s
  More actions for workflow run 740

• app tests: keep the test certificate's serial a minimal DER integer

  completed successfully:
  App and mobile checks#739:Commitd1018b3pushedbyashhart
  main

  Oct 19, 20268m 41s
  More actions for workflow run 739

• GLM-5.3-Flash (Metal): a tensor-tile kernel for the MTP draft head's eh_proj, on by default — ~5% prefill improvement

  completed successfully:
  App and mobile checks#738:Pull request#1066synchronizebyEugeneClaw
  EugeneClaw:eh-nax-ehproj-tensor-tile

  Oct 19, 20269m 48s
  More actions for workflow run 738

• GLM-5.3-Flash (Metal): a tensor-tile kernel for the MTP draft head's eh_proj, on by default — ~5% prefill improvement

  cancelled:
  App and mobile checks#737:Pull request#1066synchronizebyEugeneClaw
  EugeneClaw:eh-nax-ehproj-tensor-tile

  Oct 19, 202631s
  More actions for workflow run 737

• App and mobile checks

  completed successfully:
  App and mobile checks#736:byashhart
  v1.1.1

  Oct 19, 20266m 11s
  More actions for workflow run 736

• Merge 1.1.1 into main

  failed:
  App and mobile checks#735:Commit47ea1bbpushedbyashhart
  main

  Oct 19, 20267m 45s
  More actions for workflow run 735

• GLM-5.3-Flash (Metal): a tensor-tile kernel for the MTP draft head's eh_proj, on by default — ~5% prefill improvement

  completed successfully:
  App and mobile checks#734:Pull request#1066openedbyEugeneClaw
  EugeneClaw:eh-nax-ehproj-tensor-tile

  Oct 19, 20268m 47s
  More actions for workflow run 734

• --thinking-reserve: max_tokens limits the answer when thinking is on

  completed successfully:
  App and mobile checks#733:Pull request#1064openedbyBobClawblaw
  BobClawblaw:thinking-reserve

  Oct 19, 20268m 45s
  More actions for workflow run 733

• Benchmark: ROCm receipts name their GPU, CPU and checkpoint the way the board checks them

  completed successfully:
  App and mobile checks#732:Pull request#1043synchronizebyThinkOffApp
  ThinkOffApp:benchmark-rocm-host-identity

  Oct 19, 202610m 28s
  More actions for workflow run 732

• Withdrawn

  failed:
  App and mobile checks#731:Pull request#1063openedbyphilip-pentatonic
  Pentatonic-Ltd:see-5083-tensorfold-prefix-cache-retention-keep

  Oct 19, 20267m 44s
  More actions for workflow run 731

• HIP: refuse --drafter until the engine runs DFlash2

  completed successfully:
  App and mobile checks#730:Pull request#1044synchronizebyThinkOffApp
  ThinkOffApp:hip-refuses-drafter

  Oct 19, 20269m 40s
  More actions for workflow run 730

• Benchmark: keep each attempt's draft counts beside the receipt

  failed:
  App and mobile checks#729:Pull request#1045synchronizebyThinkOffApp
  ThinkOffApp:benchmark-drafter-counts

  Oct 19, 20268m 22s
  More actions for workflow run 729

• Flash Next CUDA: structured output and required tool calls

  failed:
  App and mobile checks#728:Pull request#1062openedbyMirrdhyn
  hygurlabs:flashnext-cuda-structures

  Oct 19, 20269m 17s
  More actions for workflow run 728

• qwen3_5 CUDA: a captured bucket's padding rows no longer go through the routed experts (35B, every multi-row forward)

  completed successfully:
  App and mobile checks#727:Pull request#1047synchronizebyjschmied
  jschmied:q35-moe-live-rows

  Oct 19, 20267m 2s
  More actions for workflow run 727

• DeepSeek-V4.1-1017: experimental two-Mac MoE partition with SSD Engram

  completed successfully:
  App and mobile checks#726:Pull request#1061openedbycwschroeder
  cwschroeder:feat/ds1017-tp2-integration

  Oct 19, 202610m 30s
  More actions for workflow run 726