All pull requests

Pull requests

• perf(cake_dsa_indexer): longest-segment-first unit tickets for multi-segment calls (sm_100a / sm_103a)

  run-ci
  #7953·yyihuang opened Oct 30, 2026Collaborator·

  ·
  2

• fix(cake_lm_head_loss): check a caller's valid-row count on every path

  run-ci
  #7952·yyihuang opened Oct 30, 2026Collaborator·

  ·
  2

• perf(cake_kda_train): make the KCP context one leaf of FSDP2 input walks

  op: linear attentionrun-ci
  #7951·yyihuang opened Oct 30, 2026Collaborator·

  ·
  2

• docs(cake_rmsnorm): the training kernels beside an all-SM kernel, with a stall regression test

  op: miscrun-ci
  #7950·yyihuang opened Oct 30, 2026Collaborator·

  ·
  5

• feat(cake_vsa_sm90): hoist Sage-FP8 GQA pin vf loads and route long ragged pair plans to the shared-transposer queue

  run-ci
  #7949·yyihuang opened Oct 30, 2026Collaborator·

  ·
  2

• perf(cake_kimi_k3_latent_moe): faster Kimi-K3 LatentMoE TP8 prefill front and tail

  run-ci
  #7948·yyihuang opened Oct 30, 2026Collaborator·

  ·
  2

• Refresh SM100 CuTe DSL FMHA kernels and remove PDL specialization

  op: attention
  #7947·xrq-phys opened Oct 30, 2026Contributor·3 of 12 tasks·

  ·
  1

• chore: add Quantization section to CODEOWNERS

  #7946·kahyunnam opened Oct 30, 2026Member·2 of 3 tasks·

  ·
  3

• feat(moe_ep): support MXFP4 W4A4 on SM107

  #7944·akaashrp opened Oct 30, 2026Collaborator·3 of 12 tasks·

  ·
  3

• extend-A8W4-optimizations-to-A4W4

  run-ci
  #7913·yanqinz2 opened Oct 29, 2026Collaborator·3 of 12 tasks·

  ·
  16

• feat(moe_ep): select Rubin MegaMoE variants with autotune v2

  run-ci
  #7911·akaashrp opened Oct 29, 2026Collaborator·5 of 12 tasks·

  ·
  7

• perf(moe_ep): skip masked slots in SM107 quantized combine

  run-ci
  #7910·akaashrp opened Oct 29, 2026Collaborator·3 of 12 tasks·

  ·
  8

• ci(docker): split the CI image pip install into torch, deps and cuDNN layers

  run-ci
  #7899·dierksen opened Oct 29, 2026Collaborator·4 of 6 tasks·

  ·
  7

• ci: share release/nightly build pipeline via a reusable workflow

  run-ci
  #7898·dierksen opened Oct 29, 2026Collaborator·4 of 6 tasks·

  ·
  2

• ci: overlap PR AOT provider builds within one job

  run-ci
  #7897·dierksen opened Oct 29, 2026Collaborator·1 of 12 tasks·

  ·
  6

• DO NOT MERGE: Token Test

  run-ci
  #7892·alex-ameen opened Oct 29, 2026Contributor·

  ·
  2

• feat(api_logging): record autotuning and CUDA-graph capture state per call

  #7888·Vinnie6167 opened Oct 29, 2026Contributor·12 tasks·

  ·
  3

• fix(gdn): fp32 MTP pool scatter against real serving pools

  op: linear attentionrun-ci
  #7882·brandonfzhang opened Oct 29, 2026Collaborator·5 of 12 tasks·

  ·
  13

• fix(page): reject dtype mismatches in append_paged_kv_cache and append_paged_mla_kv_cache

  op: attention
  #7875·Baiiiiiiiiii opened Oct 29, 2026·5 of 6 tasks·

  ·
  1

  1

• fix(moe_ep): launch CuTeDSL reference helpers on the current CUDA stream

  #7864·Heisenberg-Yin opened Oct 29, 2026·1 of 12 tasks·

  ·
  2

• fix(sparse): restrict CUDA-core decode to single-row BSR blocks

  #7859·taking-lying-flat opened Oct 29, 2026·12 tasks·

  ·
  1

• fix(attention): avoid double ALiBi soft-cap scaling

  op: attention
  #7857·taking-lying-flat opened Oct 29, 2026·1 of 12 tasks·

  ·
  4

• Bump CCCL to 67064a8b (±0.0 normalization fix + perf recoup fo…

  #7853·elstehle opened Oct 29, 2026Contributor·5 tasks done·

  ·
  5

• feat(gemm): quantize masked grouped GEMM output to NVFP4 in the epilogue

  op: gemm
  #7851·garry569 opened Oct 29, 2026·3 tasks done·

  ·
  2

• feat(gemm): fuse gated activation into masked grouped GEMM epilogue

  op: gemm
  #7850·garry569 opened Oct 29, 2026·3 tasks done·

  ·
  2