All pull requests

Pull requests

• perf(cake_kimi_k3_latent_moe): round 8 -- single-wave aligned K halves for the front prefill trailing wave, evict_first front instance for T <= 512, coalesced stream-K partial slots (SM100a/SM103a)

  #5671·yyihuang opened Sep 29, 2026Collaborator·

  ·
  5

• perf(cake_latent_moe): Kimi-K3 TP12 fused LatentMoE tail round 6: fused K23 up to eight tokens (capacity ladder), fp64-gated numerics

  run-ci
  #5668·yyihuang opened Sep 29, 2026Collaborator·

  ·
  4

• feat: add concat_mla_kv_quant_fp8, a fused MLA context K/V pack with saturating fp8 cast

  #5667·brandonfzhang opened Sep 29, 2026·

  ·
  7

• Add support for persistent TopK kernel and explicitly skipping CuTe-DSL finalize kernel

  #5666·djns99 opened Sep 29, 2026Contributor·

  ·
  18

• feat(kda): kda-for-kda CAKE PTX impl

  op: linear attentionrun-ci
  #5665·Lyken17 opened Sep 29, 2026·

  ·
  16

• perf(moe): conflict-free, warp-group-local write-back in the SM90 grouped FP8 block-scale GEMM

  op: moe
  #5664·charon14 opened Sep 29, 2026·

  ·
  1

• fix(moe): reject unsupported routing modes and per-token scales in every unified runner

  op: moerun-ci
  #5663·feih-nv opened Sep 29, 2026Collaborator·

  ·
  8

• feat(moe_ep): add Rubin SiTU, prequantized weights, and NVFP4 scaling

  run-ci
  #5662·akaashrp opened Sep 28, 2026Collaborator·

  ·
  4

• fix(gdn): keep CP prefill correct when max_seqlen is omitted

  op: linear attention
  #5661·lucifer1004 opened Sep 28, 2026Contributor·

  ·
  1

• feat(cake_nvfp4_attn): opt-in uncompensated (nodelta) NVFP4 varlen attention variant for MiniMax-H3 SM120

  run-ci
  #5660·yyihuang opened Sep 28, 2026Collaborator·

  ·
  11

• fix(cake_kda): repack the BF16 inverse fragment before block composition in the native prefill facade bodies

  op: linear attentionrun-ci
  #5659·yyihuang opened Sep 28, 2026Collaborator·

  ·
  5

• ci: add @yyihuang to every CODEOWNERS rule

  #5654·yyihuang opened Sep 28, 2026Collaborator·

  ·
  1

• fix(comm): synchronize LL clusters before peer shared-memory stores

  op: comm
  #5652·benchislett opened Sep 28, 2026·

  ·
  1

• fix(sampling): reserve RNG offsets for the current call

  op: misc
  #5651·Lonnyyy123 opened Sep 28, 2026·

  ·
  1

  1

• Add native W4A16 prefetch and grouped-tile tactics on SM121

  op: gemm
  #5648·stecasta opened Sep 28, 2026Contributor·

  ·
  1

• perf(cutlass_moe): shared-memory tile expert sort for large multi-block prologues

  op: moe
  #5640·charon14 opened Sep 28, 2026·

  ·
  1

• perf(cake_fused_moe): Kimi-K3 W4A8 MXFP4 SiTU routed MoE (SM100/SM103): mixed 192-row tiles on the MoE-TP shard's dense rows

  op: moerun-ci
  #5639·yyihuang opened Sep 28, 2026Collaborator·

  ·
  20

• feat(comm): add H6144 MNNVL CuTe DSL profiles (TP4/TP8)

  op: comm
  #5635·3xela opened Sep 28, 2026Contributor·

  ·
  1

• perf(cake_gqa): serve long/ragged BF16 paged GQA decode (q_len 1 and packed MTP 3..8) on SM100/SM103 with the on-device load-balanced ForGen programs

  op: attentionrun-ci
  #5634·yyihuang opened Sep 28, 2026Collaborator·

  ·
  9

• fix(bench): label paged-prefill cuDNN KV cache as HND, not NHD

  #5633·AmirF194 opened Sep 28, 2026·

  ·
  1

  1

• fix(comm): support Kimi K3 shapes in MNNVL HT allreduce_fusion

  op: comm
  #5632·syuoni opened Sep 28, 2026Contributor·

  ·
  6

  

• fix(moe): isolate CUTLASS tuning configuration per call

  op: moe
  #5631·ZenAlexa opened Sep 28, 2026Contributor·

  ·
  1

  4

• perf(attention): split the KV range in SM90 FA3 batch prefill when SMs would sit idle

  op: attention
  #5620·SamMausberg opened Sep 27, 2026Contributor·

  ·
  5

• perf(attention): run FP8 KV decode tiles in FP16 without lane shuffles

  op: attention
  #5619·SamMausberg opened Sep 27, 2026Contributor·

  ·
  1

• perf(gemm): split K across a thread block cluster in small-M fp8_blockscale_gemm_sm90

  op: gemm
  #5618·SamMausberg opened Sep 27, 2026Contributor·

  ·
  1