All pull requests

Pull requests

• [fused_prefill] Run all attention on attn_bfp16

  #2200·erwei-xilinx opened Oct 26, 2026Collaborator·

  ·
  3

• [LLMs] Add six affine-Q4 deployments with BFP16 prefill

  #2198·tonyjie opened Oct 26, 2026Contributor·

  ·
  5

• [AIR] airrt-to-npu: split the shim feed weave into segments by job, and keep a hoisted read's dependency on the loop that fills its buffer

  #2195·vule20 opened Oct 25, 2026Contributor·

  ·
  5

• [AIR] airrt-to-npu: keep the cores' stack size in the lock-reset device clone

  #2193·vule20 opened Oct 25, 2026Contributor·

  ·
  1

• [AIR] airrt-to-npu: keep the shim burst weave valid when one herd has far fewer tasks than another

  #2192·vule20 opened Oct 25, 2026Contributor·

  ·
  4

• [gpu] Megakernel decode on AMD GPUs: whole-model persistent launch, chiplet-aware scheduling

  #2011·erwei-xilinx opened Sep 21, 2026Collaborator·

  ·

• [fused_decode] Spell data movement as air.dma_memcpy_nd and derive the far half

  #1977·erwei-xilinx opened Sep 3, 2026Collaborator·

  ·

• [multi-gpu] Phase 7: aircc integration (--multi-gpu flag)

  #1582·erwei-xilinx opened May 3, 2026Collaborator·

  ·

• Fix air-to-std crash for DMA ops with >4 offset/size dimensions

  #1517·erwei-xilinx opened Apr 8, 2026Collaborator·3 tasks done·

  ·
  2

• Reduce Python Code Duplication

  #1476·hunhoffe opened Mar 28, 2026Contributor·

  ·

• Enable control packet reconfiguration tests on ci

  #1275·erwei-xilinx opened Feb 3, 2026Collaborator·

  ·
  1