All issues

[ASM] a16w16 split-K: the entry-zero counter is unsafe under CUDA graph capture; opus-style fp32 partials would remove the class

#4920

· JohnQinAMD opened

on Aug 21, 2026

[Bug][gfx942] 1-stage fused all-reduce + RMSNorm (custom_fused_ar_rms, use_1stage=True) is silently wrong on the first cold HIP-graph replay

#4888

· indivats opened

on Aug 20, 2026

[Issue]: Silent wrong results on gfx942 TP8 with GLM-5.2-FP8: degenerate output, fixed by VLLM_ROCM_USE_AITER_CUSTOM_AR=0

#4855

· jin-amd opened

on Aug 19, 2026

FlyDSL cache-name collision risk: compile_moe_reduction() residual-pointer signature change keeps the pre-change cache name when use_residual=False

#4840

· vvagaytsev opened

on Aug 18, 2026

[Feature] gfx942 (CDNA3/MI300X) support — or fail-fast at selection time — for `mla_gluon` MLA decode and the `fp8_ds_mla` quant-insert path

#4831

· skimanwhite opened

on Aug 18, 2026

[Bug] `top_k_top_p_sampling_from_probs` segfaults (host-side) during vLLM sampler warm-up on MI300X (gfx942), aiter main @ e8f438a

#4830

· skimanwhite opened

on Aug 18, 2026

[Bug] MoeSortingMultiPhaseKernel_P0_v1: missing lower bound on signed expert id causes GPU memory fault under EP (one-line fix)

#4827

· xudonlyu opened

on Aug 18, 2026

sampling.cuh may be missing hipcub includes — the ROCm 10 gate from #4754 may be unnecessary

#4819

· zufayu opened

on Aug 18, 2026

[Bug] Interrupted JIT codegen can leave and persistently reuse an incomplete CK MHA cache

#4799

· Ruye-aa opened

on Aug 17, 2026

AITER fused_moe: GPU memory permission fault on gfx942 at FP8 e4m3fnuz per-1x128 block-scaled MoE shapes M ∈ {304, 336}

#4783

· vipsharm opened

on Aug 16, 2026

[Bug] GroupNorm does not follow FP32 autocast policy

#4780

· akshatvishu opened

on Aug 15, 2026

[Bug] JIT-built module_fmha_fwd_bf16_opus rejects tensors from prebuilt module_aiter_core: pybind11 internals mismatch (v11 wheel vs v12 runtime JIT)

#4770

· hongxiayang opened

on Aug 14, 2026