Issues

• [MLA] gfx942 persistent a8w8 decode returns output / kv_scale at q_len 3 and 4

  ROCm/aiter#7393 · Filter by author krakhit opened on Nov 3, 2026

• [Issue]: gfx950 Opus MoE sort auto dispatch keeps E=384 small-token decode on the slower one-shot kernel

  ROCm/aiter#7358 · Filter by author yelkhamra opened on Nov 2, 2026

• [Issue]: AITER_AR_1STAGE_MAX_KB is compared as bytes, not KiB, in the fused all-reduce one-stage limit

  ROCm/aiter#7356 · Filter by author yelkhamra opened on Nov 2, 2026

• [Issue]: FlyDSL MoE sorting shares one cached multiphase workspace across calls, so concurrently replayed graphs corrupt each other's sort

  ROCm/aiter#7354 · Filter by author yelkhamra opened on Nov 2, 2026

• [Issue]: FlyDSL MoE sorting p23_kernel reuses cumsum LDS without a barrier (unmasked read/reuse race)

  ROCm/aiter#7352 · Filter by author yelkhamra opened on Nov 2, 2026

• [Issue]: FlyDSL fused_compress_attn corrupts the first compressed entry of each sequence (negative-history state-cache iterations)

  ROCm/aiter#7350 · Filter by author yelkhamra opened on Nov 2, 2026

• [Perf] Gemma 4 26B-A4B BF16 on MI350P: CK MoE stage-2 kernel at ~30% of peak, stalled on memory waits (MoE 1.14x H200 prefill, 1.48x decode)

  ROCm/aiter#7333 · Filter by author amd-sriram opened on Nov 2, 2026
  1

• [Perf][gfx950][MXFP4] gemm_a4w4 selects a slower path for narrow-N dense GEMM shapes

  ROCm/aiter#7287 · Filter by author yifehuan opened on Nov 2, 2026

• [Issue]: Triton MHA prefill attn_fwd (flash_attn_triton_amd) computes q/k/v/out row offsets in int32, which overflow for tensors above 2^31 elements

  ROCm/aiter#7278 · Filter by author Ragua1 opened on Nov 1, 2026

• RFC: Versioned Kernel Provider Interface for AITER

  ROCm/aiter#7240 · Filter by author zufayu opened on Oct 30, 2026
  

• FlyDSL a8w4 MoE stage 1 t32x128x256_w{2,3,4}_gui_fp8 is nondeterministic: rows 16-31 of a 32-row sorted block get corrupted 4x16 patches (shipped DeepSeek-V4 tuned rows 256/512/1024)

  ROCm/aiter#7158 · Filter by author vivienfanghuagood opened on Oct 29, 2026

• RFC: decouple backend names from public op definitions (start with MoE)

  ROCm/aiter#7150 · Filter by author zufayu opened on Oct 29, 2026

• paged_attention_ragged: out-of-context workgroups dominate decode time under SGLang's 1024-partition grid (small and mixed-length batches)

  ROCm/aiter#7051 · Filter by author Arist12 opened on Oct 27, 2026

• retune GEMM-AFP8WFP8_A32_W1X32 on 950

  Triton/Gluon
  ROCm/aiter#7046 · Filter by author Boss2002n opened on Oct 27, 2026

• Triton nightly status

  triton-nightly
  ROCm/aiter#7037 · Filter by author github-actions opened on Oct 27, 2026
  4

  

• MXMOE coupled tuner (--mxfp4-flydsl): restore hardening removed from #6464

  ROCm/aiter#7025 · Filter by author Bernard-Liu opened on Oct 27, 2026
  

• [gfx950][MLA] FP8-Q persistent decode mla_a8w8_qh16_qseqlen{2,4} returns wrong, nondeterministic output

  ROCm/aiter#7007 · Filter by author yuychang opened on Oct 27, 2026

• [Bug] gfx950 fmha_v3_bwd hd128 a32 psskddv kernel returns wrong dq/dk/dv when seqlen_q == 1

  ROCm/aiter#6964 · Filter by author jeffdaily opened on Oct 26, 2026

• [Issue]: flash_attn_triton_amd prefill: RDNA uses one config for every dtype and head dim; on gfx1101 head dim 64 (bf16/fp16) runs 1.16–1.32x faster with another

  ROCm/aiter#6962 · Filter by author Ragua1 opened on Oct 26, 2026

• CustomAllreduce: fused_ar_rms reads the wrong peer memory when the eager input pool doesn't start its allocation (IPC offset is always 0)

  ROCm/aiter#6918 · Filter by author merceod opened on Oct 25, 2026

• Add gfx950 FlyDSL 4-wave INT8 bpreshuffle GEMM

  ROCm/aiter#6827 · Filter by author ksikiric opened on Oct 22, 2026

• Fixes from running agentic RL rollouts with SGLang on MI350X / MI355X

  ROCm/aiter#6818 · Filter by author Arist12 opened on Oct 22, 2026
  1

• Update model_shapes.json to reflect new and latest models' shapes/kernels

  Triton/Gluon
  ROCm/aiter#6781 · Filter by author Boss2002n opened on Oct 21, 2026
  

• Tuning should use the benchmarks from benchmark folder

  ROCm/aiter#6777 · Filter by author Boss2002n opened on Oct 21, 2026

• remove nesting from config files

  Triton/Gluon
  ROCm/aiter#6776 · Filter by author Boss2002n opened on Oct 21, 2026