Server

Workflow runs for Server

• cuda: opt-in MoE epilogue fusions (GGML_CUDA_MOE_EP_FUSION)

  completed successfully:
  Server#568:Pull request#423openedbybri-prism
  cuda/moe-epilogue-fusion

  Oct 31, 202612m 6s
  More actions for workflow run 568

• build: fix the SYCL and Windows CUDA 13.3 release builds after the upstream merge

  completed successfully:
  Server#567:Pull request#422openedbybri-prism
  fix/sycl-glu-reorder-dup

  Oct 31, 202612m 24s
  More actions for workflow run 567

• build: fix RPC, Vulkan and SYCL builds after the upstream merge

  completed successfully:
  Server#566:Pull request#421openedbybri-prism
  fix/merge-release-build

  Oct 30, 202614m 11s
  More actions for workflow run 566

• cuda: bit-serial Q1_0 dot and four rows per CTA for one-column mat-vec

  completed successfully:
  Server#565:Pull request#420openedbybri-prism
  cuda/q1_0-bit-serial-mmvq

  Oct 30, 202616m 10s
  More actions for workflow run 565

• cuda: keep the extended tall mat-vec types off fused gate+up launches

  completed successfully:
  Server#564:Pull request#419openedbybri-prism
  cuda/mmvq-tall-skip-fused-gate

  Oct 30, 202613m 16s
  More actions for workflow run 564

• fix(chat): accept </parameter> without the surrounding newlines

  requires action with the application:
  Server#563:Pull request#232synchronizebydongtran16092006
  dongtran16092006:fix/parameter-close-newlines

  Oct 30, 2026Action required
  More actions for workflow run 563

• Merge upstream master (d854d98ba) into prism

  completed successfully:
  Server#562:Pull request#352synchronizebybri-prism
  merge/upstream-20261011

  Oct 29, 202612m 53s
  More actions for workflow run 562

• cuda: 5120-wide fused norm + Hadamard with PDL, and two-row PQ2_0 decode tiles on Hopper

  completed successfully:
  Server#561:Pull request#417synchronizebybri-prism
  cuda/pq2-decode-tune

  Oct 29, 202614m 58s
  More actions for workflow run 561

• cuda: use the float4 RMS norm path by default, with a pinned rounding order

  completed successfully:
  Server#560:Pull request#418synchronizebybri-prism
  cuda/rms-norm-vec-default

  Oct 29, 202612m 27s
  More actions for workflow run 560

• Merge upstream master (d854d98ba) into prism

  completed successfully:
  Server#559:Pull request#352synchronizebybri-prism
  merge/upstream-20261011

  Oct 29, 202612m 9s
  More actions for workflow run 559

• Merge upstream master (d854d98ba) into prism

  completed successfully:
  Server#558:Pull request#352synchronizebybri-prism
  merge/upstream-20261011

  Oct 29, 202613m 51s
  More actions for workflow run 558

• Merge upstream master (d854d98ba) into prism

  completed successfully:
  Server#557:Pull request#352synchronizebybri-prism
  merge/upstream-20261011

  Oct 29, 202614m 7s
  More actions for workflow run 557

• cuda: use the float4 RMS norm path by default, with a pinned rounding order

  completed successfully:
  Server#556:Pull request#418openedbybri-prism
  cuda/rms-norm-vec-default

  Oct 29, 202615m 19s
  More actions for workflow run 556

• cuda: 5120-wide fused norm + Hadamard with PDL, and two-row PQ2_0 decode tiles on Hopper

  completed successfully:
  Server#555:Pull request#417openedbybri-prism
  cuda/pq2-decode-tune

  Oct 29, 202612m 54s
  More actions for workflow run 555

• Merge upstream master (d854d98ba) into prism

  completed successfully:
  Server#554:Pull request#352synchronizebybri-prism
  merge/upstream-20261011

  Oct 29, 202614m 25s
  More actions for workflow run 554

• ggml-backend : skip empty inputs in the batched sched uploads (fix Metal GGML_ASSERT(buf_src) since #393)

  completed successfully:
  Server#553:Pull request#416openedbybri-prism
  fix/sched-skip-empty-uploads

  Oct 29, 202620m 5s
  More actions for workflow run 553

• cuda: warp-per-row PTQ1_0 mat-vec for MUL_MAT_ID when K <= 4096

  completed successfully:
  Server#552:Pull request#415openedbybri-prism
  cuda/ptq1-moe-warprow

  Oct 29, 202617m 2s
  More actions for workflow run 552

• cuda: fuse the Gemma 4 MoE layer tail (GGML_CUDA_TAIL_FUSION, off by default)

  completed successfully:
  Server#551:Pull request#414openedbybri-prism
  gemma/tail-fusion

  Oct 28, 202619m 59s
  More actions for workflow run 551

• cuda: tall one-column mat-vec for Q8_0/Q4_0/Q2_0/Q4_K/Q6_K on Hopper, and allow folded Gemma 4 weights

  completed successfully:
  Server#550:Pull request#413openedbybri-prism
  gemma/tall-ext

  Oct 28, 202614m 27s
  More actions for workflow run 550

• llama/cuda: tiered KV cache (--kv-vram-cells): VRAM head, pinned-host tail in one VMM range

  requires action with the application:
  Server#549:Pull request#319synchronizebyprofessorpalmer
  professorpalmer:claude/tiered-kv

  Oct 27, 2026Action required
  More actions for workflow run 549

• speculative: --spec-draft-window (MTP drafting at every depth)

  requires action with the application:
  Server#548:Pull request#320synchronizebyprofessorpalmer
  professorpalmer:claude/spec-draft-window

  Oct 27, 2026Action required
  More actions for workflow run 548

• cuda: GGML_CUDA_OP_TIMING=1 per-node GPU time breakdown

  requires action with the application:
  Server#547:Pull request#323synchronizebyprofessorpalmer
  professorpalmer:claude/op-timing

  Oct 27, 2026-1s
  More actions for workflow run 547

• server: --checkpoint-every-nt N, context checkpoints inside a long message (off by default)

  requires action with the application:
  Server#546:Pull request#338synchronizebyprofessorpalmer
  professorpalmer:checkpoint-every-nt

  Oct 27, 2026Action required
  More actions for workflow run 546

• tests: MUL_MAT_ID cases for the one-column mat-vec store (small K, several expert rows per block)

  requires action with the application:
  Server#545:Pull request#295synchronizebyprofessorpalmer
  professorpalmer:cuda-mmvq-mmid-row-bound

  Oct 27, 2026Action required
  More actions for workflow run 545

• cuda: gate (SwiGLU) fused PTQ1_0 mat-vec for 2-4 columns

  requires action with the application:
  Server#544:Pull request#310synchronizebysb32445
  sb32445:pr/ptq1-gate-up-fuse-mc

  Oct 27, 2026-1s
  More actions for workflow run 544