Issues

• [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

• FlyDSL int32 overflow in grouped_gemm_mxfp4 arena_handle on gfx1250 EP4 DSR1

  bug
  ROCm/aiter#7241 · Filter by author srinivamd opened on Oct 31, 2026
  3

  

• 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
  3

  

• 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

• Automate the addition of new model shapes and kernels into model shapes json in benchmarks

  Triton/Gluon
  ROCm/aiter#6775 · Filter by author Boss2002n opened on Oct 21, 2026

• track Upstream vs Internal triton regressions

  Triton/Gluon
  ROCm/aiter#6774 · Filter by author Boss2002n opened on Oct 21, 2026

• [Bug] gfx942: non-persistent bf16 MLA decode still wraps past 4 GiB (m16x4/m32x4) and ignores causal at 64/128 heads

  ROCm/aiter#6733 · Filter by author AndreasKaratzas opened on Oct 20, 2026

• [Bug] gfx942: fp8 MLA decode mla_a8w8_qh64_qseqlen4_gqaratio16_ps does not apply kv_scale to V

  ROCm/aiter#6732 · Filter by author AndreasKaratzas opened on Oct 20, 2026

• [Bug] gfx942: persistent bf16 MLA decode (mla_a16w16_qh16_m16x4_n16x1_coex0_mask1_ps) returns wrong output in some calls, nondeterministically

  ROCm/aiter#6731 · Filter by author AndreasKaratzas opened on Oct 20, 2026

• [Feature]: dynamic_mxfp4_quant(use_sr=True) outside gfx950: Triton emulation of the stochastic E2M1 conversion

  ROCm/aiter#6686 · Filter by author Ragua1 opened on Oct 19, 2026