All pull requests

Pull requests

• fix(umbp): backend-policy SSD instances keep the UMBP_SSD_* settings

  #967·TianDi101 opened Nov 10, 2026Collaborator·

  ·

• feat(ep): GPU-free dispatch/combine simulation (MORI_EP_SIM_ENABLE)

  #966·TianDi101 opened Nov 10, 2026Collaborator·

  ·

• fix(ep): 1250x dispatch scale rows; bench_ep bus bandwidth; remove EP v1's gfx125x and cco code

  #965·zhangfei829 opened Nov 10, 2026Collaborator·

  ·

• feat(ep/v2): bf16 push combine on gfx1250, chosen per call

  #964·zhangfei829 opened Nov 10, 2026Collaborator·

  ·

• gemm_ar: run the triggered all-reduce's reducer beside the GEMM

  #962·BKP opened Nov 9, 2026·6/6·4 of 5 tasks·

  ·

• gemm_ar: triggered GEMM + two-shot all-reduce op

  #961·BKP opened Nov 9, 2026·5/6·4 of 5 tasks·

  ·

• gemm_ar: gate table and gate-epilogue view

  #960·BKP opened Nov 9, 2026·4/6·3 of 4 tasks·

  ·

• feat(cco): Python bindings for trigger queues

  #959·BKP opened Nov 9, 2026·3/6·4 of 5 tasks·

  ·

• feat(cco): host-armed SDMA trigger queues

  #957·BKP opened Nov 9, 2026·2/6·5 of 6 tasks·

  ·

• feat(anvil): private SDMA queues with host access to the ring

  #956·BKP opened Nov 9, 2026·1/6·3 of 4 tasks·

  ·

• feat(ep): sync KIWI EP with upstream tuning, proxy affinity and fixes

  #950·charlifu opened Nov 9, 2026Collaborator·

  ·

• feat(umbp): bound standalone lookups with lookup_timeout_ms

  #949·charxwu opened Nov 9, 2026·6 tasks done·

  ·
  

• ci: run only the GPU lanes a PR's changes need; move the CCO unit test to PIT-G17

  #948·jhchouuu opened Nov 9, 2026Collaborator·

  ·

• fix(cco): import a torch.symm_mem buffer whose data starts inside the allocation

  #946·jhchouuu opened Nov 9, 2026Collaborator·

  ·
  1

• fix(mlx5): resolve the next-hop MAC for RoCE DEVX QPs

  #945·YeSho-cpp opened Nov 9, 2026Contributor·

  ·
  1

  

• fix(sdma): pick one even engine per peer on an SR-IOV VF

  #944·YeSho-cpp opened Nov 9, 2026Contributor·

  ·
  1

  

• fix(cco): register fabric window MRs with iova=0 for gfx1250

  #942·kawhil-amd opened Nov 9, 2026Contributor·2 of 4 tasks·

  ·
  1

• fix(umbp): re-register standalone memory after the server restarts

  #941·charxwu opened Nov 9, 2026·

  ·
  

• fix(topology): reuse the SMI library already loaded in the process

  #940·jhchouuu opened Nov 9, 2026Collaborator·

  ·

• chore(cco): warn when the mori-cco running env lacks CONFIG_PCI_P2PDMA

  #939·kawhil-amd opened Nov 9, 2026Contributor·

  ·
  1

• fix(io): bounded WaitFor polls a status that only completes when polled

  #938·amirakb89 opened Nov 9, 2026Collaborator·

  ·
  1

  2

• perf(ep/v1ll): lower MI355X EP32 small-batch latency for hidden 3584, top-16

  #937·jhchouuu opened Nov 9, 2026Collaborator·

  ·

• fix(io/sim): skip unknown catalog fields with a warning; MORI_IO_SIM_CATALOG_STRICT=1 rejects them

  #934·ZhaiFeiyue opened Nov 7, 2026Collaborator·

  ·
  1

• feat(io/sim): post_on_host -- per-op / per-segment cost on the calling thread, overlapped with the wire

  #933·ZhaiFeiyue opened Nov 7, 2026Collaborator·

  ·
  1

  1

• poc(cco): MegaMoE GEMM1 fused with an SDMA dispatch (gfx950)

  #930·kawhil-amd opened Nov 6, 2026Contributor·

  ·
  1