Search results

[tech-debt] fused_moe.py routing relies on 15+ process-level env vars — consolidate into config/API

Status: Open.

#4226 In ROCm/aiter;

· zufayu opened on Jul 14, 2026

[Feature]: Missing Prefix LM kernel for Multimodal Range Tensors

Status: Open.

#4168 In ROCm/aiter;

· GodRishUniverse opened on Jul 9, 2026

[Issue]: Crash with Qwen3.5 with tool calling enabled

Status: Open.

#4142 In ROCm/aiter;

· jbelloncastro opened on Jul 8, 2026

[Regression] Triton 3.7.0 causes ~4x regression in unified_attention on gfx950

Status: Open.

#4137 In ROCm/aiter;

· mjkvaak-amd opened on Jul 8, 2026

[Bug] int32/i24 offset overflow in FMHA prefill ASM kernel for seqlen >= 8Mi

Status: Open.

#4126 In ROCm/aiter;

· yudigege86 opened on Jul 7, 2026

[Build] aiter fails to compile on ROCm 7.0/7.1 — opus __builtin_amdgcn_tensor_load_to_lds needs ROCm 7.2 clang

Status: Open.

#4077 In ROCm/aiter;

· sunway513 opened on Jul 4, 2026

[Issue]: Grouped MoE build (gfx1250) ignores GPU_ARCHS

Status: Open.

#4071 In ROCm/aiter;

· simondanielsson opened on Jul 3, 2026

[Issue]: module_rmsnorm cold JIT compile is extremely slow (~225s) on gfx950 / ROCm 7.2.2

Status: Open.

#4055 In ROCm/aiter;

· YangFan0918 opened on Jul 2, 2026

[Regression]: Deepseek-R1 fusedMoE configs causing perf regression

Status: Open.

#4050 In ROCm/aiter;

· amnamasood-amd opened on Jul 1, 2026

[gfx950] Fused MoE NaN at M=1 with block_m=16 (ck_moe_stage1, fp8 per_1x128) — clean input → NaN out

Status: Open.

#4032 In ROCm/aiter;

· ntgiang71096 opened on Jul 1, 2026

[Regression] EP MoE tuned FMoE configs not picked up after #3377 (topk -= int(is_ep) breaks tuned-config lookup, e.g. GLM-4.7 tuned for topk=8 looked up as topk=7)

Status: Open.

#3999 In ROCm/aiter;

· mjkvaak-amd opened on Jun 29, 2026

[Issue]: [TRITON] DeepSeekV4 fused routing (moe_routing/routing.py) lacks expert_map

Status: Open.

#3995 In ROCm/aiter;

· yitingw1 opened on Jun 29, 2026