Issues

• [Perf Regression] 8 config(s) regressed @ aa4e7a04

  performanceregression
  ROCm/ATOM#2476 · Filter by author github-actions opened on Oct 6, 2026

• [Perf Regression] 10 config(s) regressed @ 868cb029

  performanceregression
  ROCm/ATOM#2473 · Filter by author github-actions opened on Oct 5, 2026

• [Perf Regression] 9 config(s) regressed @ 922b3519

  performanceregression
  ROCm/ATOM#2457 · Filter by author github-actions opened on Oct 3, 2026

• [Perf Regression] 12 config(s) regressed @ 922b3519

  performanceregression
  ROCm/ATOM#2440 · Filter by author github-actions opened on Oct 1, 2026

• GLM-5.3-Flash MTP: verify step never writes sparse KV indices below index_topk (or under CUDA graphs); DSA layers read the previous request's selection

  ROCm/ATOM#2424 · Filter by author capy-ai-staging opened on Sep 29, 2026
  1

• vLLM plugin pin (2cf0a6915, Aug 23) predates GLM-5.3-Flash support; no published atom-dev image can serve it

  bugenhancement
  ROCm/ATOM#2360 · Filter by author npoulad1 opened on Sep 22, 2026
  1

• moriio registers as a reads_block_regions backend but never reads them: DSA models decode garbage past the dense window, and its completion test frees in-flight TransferStatus objects

  ROCm/ATOM#2354 · Filter by author IvanShan177 opened on Sep 22, 2026
  2

• [CI] Approving a PR restarts the CI pipeline

  ROCm/ATOM#2322 · Filter by author junhaha666 opened on Sep 21, 2026
  

• GLM-5.3-Flash does not work with vllm serve on rocm/atom-dev:vllm-latest

  ROCm/ATOM#2316 · Filter by author sajandhy opened on Sep 20, 2026

• [TAG] ATOM: Automatically update tags bi‑weekly

  ROCm/ATOM#2159 · Filter by author amd-ruitang3 opened on Sep 8, 2026
  

• [Feature]: vllm atom plugin to support deepseek v4 flash 0731

  ROCm/ATOM#2136 · Filter by author peizhang56 opened on Sep 3, 2026

• [Feature]: Profiler: option to set a profiler window

  ROCm/ATOM#2081 · Filter by author devalshahamd opened on Aug 28, 2026

• [Issue]: GPT-OSS-120B fails on the ATOM SGLang stack (MXFP4 triton_kernel MoE)

  ROCm/ATOM#2065 · Filter by author ibeattyo opened on Aug 27, 2026
  

• [Issue]: ATOM: TP-DSA decode hang in the plugin forward-relay

  ROCm/ATOM#2062 · Filter by author ibeattyo opened on Aug 27, 2026

• [Issue]: DeepSeek V4 TP-MoE does not work

  ROCm/ATOM#2022 · Filter by author ibeattyo opened on Aug 25, 2026
  2

  

• [Issue]: Mori EP-MoE not CUDA-graph-capturable under TP-attention

  ROCm/ATOM#2021 · Filter by author ibeattyo opened on Aug 25, 2026

• [Feature]: Multimodal support in ATOM

  ROCm/ATOM#2014 · Filter by author cubezhang opened on Aug 25, 2026
  5

  

• [Issue]:

  ROCm/ATOM#1972 · Filter by author ibeattyo opened on Aug 20, 2026

• [Issue]: Kimi K3 ATOM+SGLang has ~309KB per token for KV while SGLang with no atom plugin uses only 26KB

  ROCm/ATOM#1959 · Filter by author ibeattyo opened on Aug 19, 2026
  

• [Issue]:

  ROCm/ATOM#1957 · Filter by author ibeattyo opened on Aug 19, 2026

• [Issue]: GPT-OSS-120B fails on the ATOM SGLang stack (MXFP4 triton_kernel MoE)

  ROCm/ATOM#1945 · Filter by author ibeattyo opened on Aug 18, 2026
  2

• [Feature]: FP8 block-quant MoE rejects TP shards narrower than block_n (Qwen3.6-35B-A3B)

  ROCm/ATOM#1944 · Filter by author ibeattyo opened on Aug 18, 2026
  

• [Issue]: ATOM build holds **~16 GB/GPU more resident memory than SGLang without ATOM for the identical model**

  ROCm/ATOM#1935 · Filter by author ibeattyo opened on Aug 17, 2026

• [Feature]: DeepSeek V4 on ATOM

  ROCm/ATOM#1934 · Filter by author ibeattyo opened on Aug 17, 2026

• [Feature]: Enable DP-attention + EP (mori) for Gated-DeltaNet MoE models on ATOM — pick up SGLang #31794

  ROCm/ATOM#1929 · Filter by author ibeattyo opened on Aug 17, 2026
  1