Issues

• [Bug] moe_gemm_a16w4 gfx950 Gluon kernel fails to compile when indices are upcast to int64 (buffer_load_to_shared needs int32 offsets)

  ROCm/aiter#6072 · Filter by author zhanglx13 opened on Oct 2, 2026
  

• [vLLM] DSR1 FP4 Optimizations

  ROCm/aiter#6062 · Filter by author farlukas opened on Oct 1, 2026
  

• FlyDSL PA decode kernel

  ROCm/aiter#6059 · Filter by author ahmed-bsod opened on Oct 1, 2026
  

• FlyDSL MLA bf16 kernel

  ROCm/aiter#6058 · Filter by author ahmed-bsod opened on Oct 1, 2026
  

• MXFP8 GEMM for DSV4.1

  ROCm/aiter#6057 · Filter by author ahmed-bsod opened on Oct 1, 2026
  

• [Feature]: Improve top_k_per_row_decode performance at <64k ctx, topk=512

  ROCm/aiter#6053 · Filter by author simondanielsson opened on Oct 1, 2026

• DSV4.1 Flash Optimization

  Task

  ROCm/aiter#6023 · Filter by author ahmed-bsod opened on Sep 30, 2026
  

• Create Triton kernel backup for paged_mxfp4_mqa_logits operator

  ROCm/aiter#6021 · Filter by author brunomazzottiamd opened on Sep 30, 2026

• [Feature]: Create a Perf CI

  Triton/Gluon
  Feature

  ROCm/aiter#6014 · Filter by author Boss2002n opened on Sep 30, 2026
  

• [Kimik3] Accumulate the latent-MoE up-projection in aiter a16w16 atomic

  ROCm/aiter#6012 · Filter by author nsusanto opened on Sep 30, 2026
  1

  

• [Kimik3] Fuse Kimi-K3 MLA chunked-context gather on AITER

  ROCm/aiter#6011 · Filter by author nsusanto opened on Sep 30, 2026
  

• [KimiK3] Fuse shared-expert down-proj into the latent-MoE tail GEMM

  ROCm/aiter#6010 · Filter by author nsusanto opened on Sep 30, 2026
  

• move gluon kernels into proper Dirs

  Triton/Gluon
  Task

  ROCm/aiter#6008 · Filter by author Boss2002n opened on Sep 30, 2026
  

• [Issue]: gfx950 FlyDSL BF16 GEMM (gemm_a16w16) reduces split-K / slice-K partial sums in BF16: ~6% of outputs deviate ≥2 ULP from an FP32-accumulation reference at small M

  ROCm/aiter#5976 · Filter by author irvineoy opened on Sep 30, 2026

• [Issue]: AITER_TRITON_ONLY leaves aiter.dtypes unbound, silently disabling aiter.ops.triton.quant

  ROCm/aiter#5959 · Filter by author Ragua1 opened on Sep 29, 2026

• [Bug] gfx950 OPUS persistent BF16 kid 300 output depends on partial-K row padding

  ROCm/aiter#5954 · Filter by author harsh-nod opened on Sep 29, 2026

• gdr_decode_packed_bf16 accepts NaN fixed scale and returns NaN output

  ROCm/aiter#5951 · Filter by author harsh-nod opened on Sep 29, 2026

• [Feature]: Tune fp8 unified attention for Gemma4 prefill shapes [gfx942]

  ROCm/aiter#5947 · Filter by author simondanielsson opened on Sep 29, 2026
  1

• [Task]: Tune DSr1 GEMMs on gfx1250

  Triton/Gluon
  ROCm/aiter#5946 · Filter by author Boss2002n opened on Sep 29, 2026
  

• [Issue]: [Bugfix][ROCm] Force Triton MXFP4 for 96-row ASM scale tiles

  ROCm/aiter#5931 · Filter by author abrahamzewoudie opened on Sep 29, 2026

• [HIP] Duplicate hip_bfloat16 template specialization in quant_mxfp6_gemm.cu

  ROCm/aiter#5930 · Filter by author shantipriya-amd opened on Sep 29, 2026
  

• [AITER] K3 GEMM tuning

  ROCm/aiter#5923 · Filter by author omuhamma opened on Sep 29, 2026
  

• [Feature]: Centralized MXFP8 (E4M3 + E8M0 1x32) dense GEMM support

  ROCm/aiter#5916 · Filter by author sshlyapn opened on Sep 28, 2026

• [Feature] Optionally emit fp8-quantized Q from fused_qk_norm_rope_cache_pts_quant_shuffle

  ROCm/aiter#5912 · Filter by author mpashkovskii opened on Sep 28, 2026
  1

  

• [Bug] [fa]gfx950 hd128 causal bwd kernels write dK = 0 under GPU sharing (missing s_barrier)

  ROCm/aiter#5894 · Filter by author Johnsonms opened on Sep 27, 2026
  1