Issues

• [Perf Regression] 25 config(s) regressed @ 95a7147f

  performanceregression
  ROCm/ATOM#3360 · Filter by author github-actions opened on Nov 14, 2026

• [Perf Regression] 20 config(s) regressed @ 5b744ee3

  performanceregression
  ROCm/ATOM#3322 · Filter by author github-actions opened on Nov 13, 2026

• MiniMax-M3 skip-index layers: one fused writer for every KV dtype

  ROCm/ATOM#3263 · Filter by author ganyi1996ppo opened on Nov 12, 2026

• Per-model fp4 KV support hook instead of _check_fp4_kv in shared config.py

  ROCm/ATOM#3262 · Filter by author ganyi1996ppo opened on Nov 12, 2026

• MhaKvPool: take only the KV layout (drop kv_dtype); the layout reports the transfer unit

  ROCm/ATOM#3261 · Filter by author ganyi1996ppo opened on Nov 12, 2026

• MiniMax-M3.1 MXFP4 KV: route sparse PA decode through run_pa_decode and drop the segment plan loop

  ROCm/ATOM#3260 · Filter by author ganyi1996ppo opened on Nov 12, 2026

• [RFC] Pluggable KV pool backend: MORI-UMBP owns the L1 (HBM) pool so DP ranks share it

  ROCm/ATOM#3259 · Filter by author TianDi101 opened on Nov 12, 2026
  1

• [Perf Regression] 17 config(s) regressed @ 47e0c707

  performanceregression
  ROCm/ATOM#3239 · Filter by author github-actions opened on Nov 12, 2026

• [RFC] Decouple LMCache MP mode from ATOM

  ROCm/ATOM#3231 · Filter by author maobaolong opened on Nov 12, 2026
  4

  

• offload: two known gaps from #3027 (vLLM plugin media keys unsalted; Kimi-K3 chunk-aligned resend gets no tier hit)

  ROCm/ATOM#3227 · Filter by author ZLkanyo009 opened on Nov 12, 2026

• --kv_cache_dtype fp4 not supported on public main (required by DSV4.1-Flash DEP4 recipe)

  bug
  ROCm/ATOM#3222 · Filter by author srinivamd opened on Nov 11, 2026
  1

  

• vLLM plugin refactor: design

  ROCm/ATOM#3190 · Filter by author PerryZhang01 opened on Nov 11, 2026

• [Perf Regression] 3 config(s) regressed @ 724f7bf4

  performanceregression
  ROCm/ATOM#3175 · Filter by author github-actions opened on Nov 10, 2026

• [ROCm Product Quality Control]: Weekly Kimi Vendor Verifier (KVV) accuracy regression testing on AMD ATOM

  ROCm/ATOM#3149 · Filter by author functionstackx opened on Nov 10, 2026

• [Bug] Qwen3.8-Flash-Next (qwen4_exp) image prompts crash on a prefix-cache hit inside or past an image

  ROCm/ATOM#3123 · Filter by author syang-genai opened on Nov 9, 2026

• GLM-5.3-Flash with bf16 KV: aiter MLA decode misreads KV stored past 4 GiB of a layer's cache

  ROCm/ATOM#3120 · Filter by author PerryZhang01 opened on Nov 9, 2026

• [Perf Regression] 7 config(s) regressed @ 6d22f7db

  performanceregression
  ROCm/ATOM#3096 · Filter by author github-actions opened on Nov 8, 2026

• [Issue]: [DeepSeek-V4] TP ranks select MoE experts from their own replicated router logits and can disagree on identical inputs (gfx950)

  ROCm/ATOM#3095 · Filter by author yelkhamra opened on Nov 8, 2026

• [Perf Regression] 13 config(s) regressed @ 6d22f7db

  performanceregression
  ROCm/ATOM#3094 · Filter by author github-actions opened on Nov 7, 2026

• [Perf Regression] 11 config(s) regressed @ 6d22f7db

  performanceregression
  ROCm/ATOM#3089 · Filter by author github-actions opened on Nov 7, 2026

• [Issue]: Scheduled regression reruns skip labelled benchmark variants and can suppress the [Perf Regression] issue

  ROCm/ATOM#3078 · Filter by author yelkhamra opened on Nov 6, 2026

• [Perf Regression] 23 config(s) regressed @ 64412f10

  performanceregression
  ROCm/ATOM#3047 · Filter by author github-actions opened on Nov 6, 2026

• [Perf Regression] 13 config(s) regressed @ 083891e9

  performanceregression
  ROCm/ATOM#3037 · Filter by author github-actions opened on Nov 5, 2026

• mono: resident_grid_refusals compares the raw kv_connector spelling, so connector aliases and multi silently turn mono decode off

  ROCm/ATOM#3032 · Filter by author ZLkanyo009 opened on Nov 5, 2026

• Plugin mode: atom_config.load_dummy is always None, so a framework's dummy load is invisible to ATOM

  ROCm/ATOM#3028 · Filter by author ZLkanyo009 opened on Nov 5, 2026