All workflows

Workflow runs

• GLM on CUDA: the prompt history fetch under the latent split (TF_GLM_KV_SPLIT, #1431 part 3; stacked on #2935)

  completed successfully:
  Zig format#7778:Pull request#2937synchronizebywojo
  wojo:up-kvsplit-p3

  Nov 14, 202618s
  More actions for workflow run 7778

• GLM on CUDA: the prompt history fetch under the latent split (TF_GLM_KV_SPLIT, #1431 part 3; stacked on #2935)

  currently running:
  No GPU#7788:Pull request#2937synchronizebywojo
  wojo:up-kvsplit-p3

  Nov 14, 2026In progress
  More actions for workflow run 7788

• GLM on CUDA: owner-only latent rows and owner-computed attention (TF_GLM_KV_SPLIT, #1431 part 2; stacked on #2415)

  completed successfully:
  Zig format#7777:Pull request#2935synchronizebywojo
  wojo:up-kvsplit-p2

  Nov 14, 202619s
  More actions for workflow run 7777

• GLM on CUDA: owner-only latent rows and owner-computed attention (TF_GLM_KV_SPLIT, #1431 part 2; stacked on #2415)

  currently running:
  No GPU#7787:Pull request#2935synchronizebywojo
  wojo:up-kvsplit-p2

  Nov 14, 2026In progress
  More actions for workflow run 7787

• GLM on CUDA: kept states, carries and prompt twins under the latent split (TF_GLM_KV_SPLIT, #1431 part 4; stacked on #2937)

  currently running:
  No GPU#7786:Pull request#2938synchronizebywojo
  wojo:up-kvsplit-p4

  Nov 14, 2026In progress
  More actions for workflow run 7786

• GLM on CUDA: kept states, carries and prompt twins under the latent split (TF_GLM_KV_SPLIT, #1431 part 4; stacked on #2937)

  completed successfully:
  Zig format#7776:Pull request#2938synchronizebywojo
  wojo:up-kvsplit-p4

  Nov 14, 202621s
  More actions for workflow run 7776

• glm metal: use shared events for paired transfers

  requires action with the application:
  No GPU#7785:Pull request#3293synchronizebyjcscocca
  jcscocca:codex/glm-shared-event-3158

  Nov 14, 2026Action required
  More actions for workflow run 7785

• glm metal: use shared events for paired transfers

  requires action with the application:
  Zig format#7775:Pull request#3293synchronizebyjcscocca
  jcscocca:codex/glm-shared-event-3158

  Nov 14, 2026Action required
  More actions for workflow run 7775

• cuda: a resumed prompt encodes only the images its kept state does not hold (Gemma 4, Qwen3.5, Flash Next on one rank)

  cancelled:
  No GPU#7784:Pull request#3164synchronizebyashhart
  lead/cuda-encoder-skip-cached

  Nov 14, 202612m 59s
  More actions for workflow run 7784

• cuda: a resumed prompt encodes only the images its kept state does not hold (Gemma 4, Qwen3.5, Flash Next on one rank)

  completed successfully:
  Zig format#7774:Pull request#3164synchronizebyashhart
  lead/cuda-encoder-skip-cached

  Nov 14, 20261m 26s
  More actions for workflow run 7774

• spark25 cuda: gate and up load as one stacked weight, and every CUDA chip decodes them in one launch

  cancelled:
  No GPU#7783:Pull request#3130synchronizebyashhart
  lead/spark25-gate-up-stack

  Nov 14, 202613m 16s
  More actions for workflow run 7783

• spark25 cuda: gate and up load as one stacked weight, and every CUDA chip decodes them in one launch

  completed successfully:
  Zig format#7773:Pull request#3130synchronizebyashhart
  lead/spark25-gate-up-stack

  Nov 14, 20266m 36s
  More actions for workflow run 7773

• Vulkan: a dispatch records a barrier only when it conflicts with an earlier one, 2% to 20% less run time on an M1 and an M1 Max with the same logits (stacked on #2017)

  completed successfully:
  Zig format#7772:Pull request#2038synchronizebyjoshuaswarren
  joshuaswarren:vulkan-barrier-hazards

  Nov 14, 202610m 31s
  More actions for workflow run 7772

• Vulkan: a dispatch records a barrier only when it conflicts with an earlier one, 2% to 20% less run time on an M1 and an M1 Max with the same logits (stacked on #2017)

  currently running:
  No GPU#7782:Pull request#2038synchronizebyjoshuaswarren
  joshuaswarren:vulkan-barrier-hazards

  Nov 14, 2026In progress
  More actions for workflow run 7782

• glm metal: use shared events for paired transfers

  requires action with the application:
  No GPU#7781:Pull request#3293synchronizebyjcscocca
  jcscocca:codex/glm-shared-event-3158

  Nov 14, 2026Action required
  More actions for workflow run 7781

• glm metal: use shared events for paired transfers

  requires action with the application:
  Zig format#7771:Pull request#3293synchronizebyjcscocca
  jcscocca:codex/glm-shared-event-3158

  Nov 14, 2026Action required
  More actions for workflow run 7771

• hip: ModelOpt NVFP4 and FP8 checkpoints keep their codes and decode in flight

  failed:
  Zig format#7770:Pull request#2490synchronizebyElderOrb
  ElderOrb:hip-nvfp4

  Nov 14, 202611m 54s
  More actions for workflow run 7770

• hip: ModelOpt NVFP4 and FP8 checkpoints keep their codes and decode in flight

  completed successfully:
  No GPU#7780:Pull request#2490synchronizebyElderOrb
  ElderOrb:hip-nvfp4

  Nov 14, 202657m 15s
  More actions for workflow run 7780

• dsv41 metal: the DSpark drafter on an FP8 head, a concurrent propose and resident buffers

  completed successfully:
  Zig format#7769:Pull request#3296openedbyashhart
  lead/dsv41-dspark-head

  Nov 14, 20267m 46s
  More actions for workflow run 7769

• dsv41 metal: the DSpark drafter on an FP8 head, a concurrent propose and resident buffers

  currently running:
  No GPU#7779:Pull request#3296openedbyashhart
  lead/dsv41-dspark-head

  Nov 14, 2026In progress
  More actions for workflow run 7779

• Flash Next on CUDA: warm turns, so the next turn resumes past the reply

  cancelled:
  No GPU#7778:Pull request#3219synchronizebyashhart
  lead/flashnext-cuda-warm-turns

  Nov 14, 20267m 12s
  More actions for workflow run 7778

• Flash Next on CUDA: warm turns, so the next turn resumes past the reply

  cancelled:
  Zig format#7768:Pull request#3219synchronizebyashhart
  lead/flashnext-cuda-warm-turns

  Nov 14, 20267m 11s
  More actions for workflow run 7768

• logit_bias: token biases and -100 bans on each row's draw for Qwen CUDA, Qwen3.5-2B Metal and Nemotron Parse HIP

  cancelled:
  No GPU#7777:Pull request#3146synchronizebyashhart
  lead/logit-bias

  Nov 14, 20267m 24s
  More actions for workflow run 7777

• logit_bias: token biases and -100 bans on each row's draw for Qwen CUDA, Qwen3.5-2B Metal and Nemotron Parse HIP

  cancelled:
  Zig format#7767:Pull request#3146synchronizebyashhart
  lead/logit-bias

  Nov 14, 20267m 23s
  More actions for workflow run 7767

• dory cuda: decode rows run q, k, v and g as one grouped NVFP4 launch, and gate and up as another

  cancelled:
  No GPU#7776:Pull request#3131synchronizebyashhart
  lead/dory-nvfp4-grouped

  Nov 14, 20267m 36s
  More actions for workflow run 7776