Issues

• Replace the remaining hand-set checkAllclose tolerances once alignment and the measured error are known

  ROCm/aiter#6596 · Filter by author KHassanQazi opened on Oct 15, 2026

• [Feature]: [kimik3][vllm] enable router gemm aiter kernels for fp32 output in vllm

  ROCm/aiter#6591 · Filter by author nsusanto opened on Oct 15, 2026
  

• [Feature]: [Kimik3][vllm] get router gemm fp32 output enabled in VLLM

  ROCm/aiter#6590 · Filter by author nsusanto opened on Oct 15, 2026
  

• RDNA3/RDNA4 (gfx1100, gfx1201) Triton-only test report: fused KDA decode works; two install/config issues

  ROCm/aiter#6574 · Filter by author boxwrench opened on Oct 15, 2026

• [Issue]: a8w4 fused_moe writes NaN at small token counts on gfx950 (uninitialized padded scratch)

  ROCm/aiter#6565 · Filter by author PerryZhang01 opened on Oct 15, 2026
  2

• FlyDSL FP4 paged MQA-logits prefill (flydsl_pa_mqa_logits_fp4_prefill) faults at address 0 with zero query rows (gfx950)

  ROCm/aiter#6562 · Filter by author ChangLi811 opened on Oct 15, 2026

• [Issue]: FlyDSL MoE stage-1 fused FP8 output uses a different MXFP8 scale rule than fused_dynamic_mx_quant_moe_sort

  ROCm/aiter#6548 · Filter by author josh-schell-amd opened on Oct 14, 2026

• mha_batch_prefill reads past the paged KV cache at head_dim 64/96 (GPU memory fault)

  ROCm/aiter#6518 · Filter by author demandal25 opened on Oct 14, 2026

• MLA prefill fails when query blocks exceed the HIP grid-y limit

  Triton/Gluon
  Bug

  ROCm/aiter#6489 · Filter by author azaidy opened on Oct 13, 2026

• [Feature request] fmha v3 asm forward for head_dim 256 on gfx942 (MI300X/MI325X): Qwen3.5/3.6 prefill runs the CK-tile fallback at ~30 % of peak

  ROCm/aiter#6474 · Filter by author kendrick-tinyfish opened on Oct 13, 2026

• fused_moe JIT (preshuffle_off) CK GEMM dispatch failure on gfx950 (MI355) for fp8 (non-block) quant at intermediate_size=384 — distinct from #5766

  ROCm/aiter#6452 · Filter by author divakar-amd opened on Oct 12, 2026

• [mha_v4] Feature request: valid-length masking for the KV tail in sorted-sparse (and dense) calls (gfx950)

  ROCm/aiter#6415 · Filter by author manproai opened on Oct 12, 2026

• [mha_v4] Suspected correctness issue: MXFP6 / MXFP4 V recipes are wrong when the KV length is not a multiple of 128 (gfx950)

  ROCm/aiter#6414 · Filter by author manproai opened on Oct 12, 2026

• [mha_v4] Suspected correctness issue: sorted-sparse output differs from dense for the MXFP8 recipe with an all-true block mask (gfx950)

  ROCm/aiter#6413 · Filter by author manproai opened on Oct 12, 2026

• [gfx1250] BF16 GEMM: Triton gemm_a16w16 leaves the rows of C past byte 2^31 unwritten

  ROCm/aiter#6402 · Filter by author jhinpan opened on Oct 11, 2026
  1

• [gfx1250] tuned_grouped_fmoe.csv: the A8W4 rows for 7168 x 768 with 384 experts are the same for every token count and slower than the defaults at T = 4096

  ROCm/aiter#6346 · Filter by author jhinpan opened on Oct 9, 2026

• [gfx1250] BF16 GEMM: shapes without a tuned entry go to Triton, which is slower than AITER's own FlyDSL kernel

  ROCm/aiter#6345 · Filter by author jhinpan opened on Oct 9, 2026

• [gfx1250] RMSNorm returns wrong values: row tails are not written when N % 8 != 0, and wide rows sometimes get a wrong scale

  ROCm/aiter#6344 · Filter by author jhinpan opened on Oct 9, 2026

• [Bug] _as_int crashes on float strings in tuned_grouped_fmoe.csv

  bug
  ROCm/aiter#6341 · Filter by author srinivamd opened on Oct 9, 2026

• [Issue]: Triton paged_attention_decode V2 double-counts and drops tokens when KV_BLK_SZ does not divide 1024

  ROCm/aiter#6320 · Filter by author truong-v opened on Oct 9, 2026

• [Issue]: Triton paged_attention_decode returns wrong output for non-power-of-two KV block sizes

  ROCm/aiter#6319 · Filter by author truong-v opened on Oct 9, 2026

• [Issue]: topk_gating kernels never select the last num_experts % vec_size experts (latent)

  ROCm/aiter#6317 · Filter by author truong-v opened on Oct 9, 2026

• [Issue]: sampling radix_topk_one_block_kernel ignores rowStarts in the radix passes and the renorm sum

  ROCm/aiter#6316 · Filter by author truong-v opened on Oct 9, 2026

• [TRITON] Merging 2d/3d Unified Attention kernels for easy maintenance

  Triton/Gluon
  ROCm/aiter#6282 · Filter by author cagrikymk opened on Oct 8, 2026
  

• [TRITON] Handling speculative tokens natively in Unified Attention

  ROCm/aiter#6281 · Filter by author cagrikymk opened on Oct 8, 2026