All pull requests

Pull requests

• FMHA prefill: give the cached-KV mainloop branch the new-KV prefetch schedule

  #526·ckvermaAI opened Oct 1, 2026·

  ·
  2

• perf(norm): fast path for contiguous fused_add_rmsnorm

  run-ci
  #524·rahulvijayaraghavan opened Sep 30, 2026·

  ·
  7

• upstream cri specific changes/optimizations

  run-ci
  #522·mkumargarg opened Sep 29, 2026Collaborator·

  ·

• added support for d_qk=576 in mla_sparse_decode for xe20, xe35 device

  run-ci
  #521·pralay-das opened Sep 29, 2026Collaborator·

  ·
  5

• Upgrade to PyTorch 2.14 and oneAPI 2026.1

  #520·arathi-hlab opened Sep 29, 2026Collaborator·

  ·

• To support bmm_fp8/bf16/fp16 kernel

  #519·mkumargarg opened Sep 28, 2026Collaborator·

  ·
  8

• Switch the build and CI environment from conda to uv

  #514·htzo opened Sep 24, 2026Collaborator·

  ·
  3

• [quant] add round_mode parameter in per_token_group_quant_8bit & per_token_group_quant_mxfp4

  #513·MeowMiaoJ opened Sep 24, 2026Contributor·

  ·

• [ci] Add workflow to test a prebuilt wheel on a given ref and Python version

  #511·arathi-hlab opened Sep 23, 2026Collaborator·

  ·
  2

• Fix moe jit function signature issue

  run-ci
  #510·gaopengff opened Sep 23, 2026Collaborator·

  ·
  1

• [XPU] Add the DeepSeek V4 expand_prefill_causally attention metadata op

  #505·unrahul opened Sep 21, 2026·

  ·
  1

• [XPU] Fuse QK into the fp8 paged MQA logits triton kernel

  #504·unrahul opened Sep 21, 2026·

  ·
  1

• Add fused scale-residual + LayerNorm + scale-shift SYCL kernel for XPU

  run-ci
  #497·ckvermaAI opened Sep 18, 2026·

  ·
  18

• perf(moe): fuse the W4A16 MoE prepare path at M==1 and memoize the bf16->fp32 bias promotion

  #496·jmunetong opened Sep 17, 2026Collaborator·7 tasks done·

  ·
  7

• Align on Python 3.12 as the single supported interpreter

  #493·arathi-hlab opened Sep 16, 2026Collaborator·

  ·

• ci(xpu): add a CRI (Xe3P) build gate to PR CI

  #492·arathi-hlab opened Sep 16, 2026Collaborator·2 tasks done·

  ·
  1

• Xpu kernel upgrade to 1.0 for sglang-v0.5.19 rebase + Mamba-Hicache fix needed new new kernel

  #488·XiangKaren opened Sep 13, 2026·

  ·

• [MoE] Adopt Column-Major Layout B for Xe2 Grouped GEMM

  #484·CaoE opened Sep 10, 2026Collaborator·

  ·

• [MoE] Optimize small-input preparation

  #483·CaoE opened Sep 10, 2026Collaborator·

  ·

• [MoE] Optimize MoE input preparation and adopt Column-Major Layout B for Grouped GEMM

  #482·CaoE opened Sep 9, 2026Collaborator·

  ·

• [Inkling] Add relative projection kernel

  run-ci
  #468·yuankuns opened Sep 2, 2026Collaborator·

  ·
  2

• CI: Update baseline.json (from PR #457)

  #463·MingxuZh opened Sep 1, 2026Collaborator·

  ·
  2

• CI: Update baseline.json (from PR #459)

  #460·MingxuZh opened Aug 31, 2026Collaborator·

  ·
  1

• attention kvcache-mxfp4

  run-ci
  #458·sunjiweiswift opened Aug 31, 2026Collaborator·

  ·
  1

• [CI] fix bench_jit_per_token_group_quant_8bit sglang import cascade

  run-ci
  #453·arathi-hlab opened Aug 27, 2026Collaborator·

  ·