All pull requests

Pull requests

• [Bugfix] DSV4: publish decode block tables at the padded running_bs

  ci:atom
  #2488·yhl-amd opened Oct 7, 2026Contributor·2 of 3 tasks·

  ·
  1

• [gfx1250] Add ATOM_MLA_ABSORB_A16W8: BF16-activation x FP8-weight MLA absorb

  #2486·karverma-amd opened Oct 7, 2026·

  ·
  1

• [ATOM-vLLM] Add support for GLM-5.3-Flash with LMCache offload

  #2481·kliuae opened Oct 6, 2026Contributor·1 task·

  ·
  1

  

• fix(rollout): stop long RL runs falling back to eager after ~40 sleep/wake cycles

  #2477·sudhu2k opened Oct 6, 2026·

  ·
  3

  

• [Perf] MoRI v2 fused transport: mask DP pad rows (ATOM_MEGA_MASK_PAD_ROWS)

  ci:atom
  #2474·JiaoliangYu opened Oct 5, 2026Contributor·

  ·
  1

  

• fix(logging): isolate atom logger and support ATOM_LOG_LEVEL

  #2470·amd-ruitang3 opened Oct 4, 2026Contributor·

  ·
  2

  

• feat(scheduler): reserve KV for projected decode growth

  #2469·amd-ruitang3 opened Oct 4, 2026Contributor·

  ·
  1

  

• feat(observability): add decode-step engine stats

  #2471·amd-ruitang3 opened Oct 4, 2026Contributor·

  ·
  1

  

• fix(eplb): wire DeepSeek-R1 MoE layer ids into EPLB

  #2467·JiaoliangYu opened Oct 4, 2026Contributor·

  ·
  1

  

• fix(kimi-k3): preserve gfx1250 MXFP4 scale layouts

  #2466·XiaobingSuper opened Oct 4, 2026Contributor·

  ·
  2

  

• feat(offload): mooncake_store — prefill prefix cache in a Mooncake Store without LMCache (GLM-5.2 cpp4/dcp4)

  #2464·Jasen2201 opened Oct 3, 2026Contributor·3/3·

  ·
  13

• feat(offload): LMCache MP servers per NUMA node and a Mooncake Store L2 pooled across nodes (GLM-5.2 cpp4/dcp4)

  #2459·Jasen2201 opened Oct 3, 2026Contributor·12 of 18 tasks·

  ·
  

• [Perf][Kimi-K3] Optimize prefill gated RMSNorm pipeline

  #2458·XiaobingSuper opened Oct 3, 2026Contributor·

  ·
  16

  

• [AMD][DSR1] feat(dp): backpressure (lazy) DP dispatch to avoid forced-eager tail

  #2456·karverma-amd opened Oct 2, 2026·

  ·
  1

  

• build(docker): build Mooncake with Mooncake Store (WITH_STORE=ON)

  #2455·Jasen2201 opened Oct 2, 2026Contributor·2 of 4 tasks·

  ·
  5

  

• [1/3 entrypoint refactor] [feat(mesh)]: unify multi-API ingress for HTTP and ext-proc

  #2453·Yuechguo opened Oct 2, 2026Contributor·

  ·
  30

  

• fix(minimax-m3): order mono stores before their done flags; keep pad rows out of routing

  #2452·akii96 opened Oct 2, 2026·

  ·
  1

• fix(glm5-next): MTP verify always writes sparse KV indices; kpool tail uses the history ring layout

  #2451·Jasen2201 opened Oct 2, 2026Contributor·1 task·

  ·
  1

  3

  

• perf(k3-prefill): warm prefill GEMM buckets at startup; stop advertising evicted KDA state images

  #2449·Jasen2201 opened Oct 2, 2026Contributor·2/2·1 task·

  ·

• feat(offload/mp): make engine_driven transfers usable, and match the in-process tier

  #2448·PerryZhang01 opened Oct 2, 2026Contributor·

  ·
  

• fix(kimi-k3): add compatibility fallback for ptpc SiTU quant

  #2447·XiaobingSuper opened Oct 2, 2026Contributor·

  ·
  4

• feat(plugin/mp): run the K3 offload connector on ATOM's MP LMCache tier

  #2443·PerryZhang01 opened Oct 1, 2026Contributor·

  ·
  2

  

• docs(offload): clarify LMCache MP lookup timeout scope

  #2442·kvnloo opened Oct 1, 2026Contributor·

  ·
  1

  

• add more MP LMCache instructions hosting Kimi-K3 from vllm ATOM plugin

  #2439·linsun12 opened Sep 30, 2026Contributor·

  ·
  

• [DCP]sparse prefill use QREP, removing AllGather Q and copy

  #2437·ZhiweiYan-96 opened Sep 30, 2026Contributor·

  ·
  2