Search results

Accuracy regression in DeepSeek-R1-MXFP4 (KV FP8) on MI35x after commit a56b520 in SGLang

Status: Open.

#2656 In ROCm/aiter;

· bingxche opened on Apr 8, 2026

[Issue]: aiter.ops.triton.fused_add_rmsnorm_pad produces different result with the unfused version.

Status: Open.

#2614 In ROCm/aiter;

· charlifu opened on Apr 3, 2026

[BUG] fused_moe moe_sorting_dispatch_policy wrong type

Status: Open.

#2576 In ROCm/aiter;

· tpopp opened on Apr 1, 2026

ZeroDivisionError in deepgemm_fp8_paged_mqa_logits_stage1 with 8 attention heads (GLM-5 TP=8)

Status: Open.

#2563 In ROCm/aiter;

· ChuanLi1101 opened on Apr 1, 2026

gemm_afp4wfp4 Triton kernel produces inf/garbage on gfx1250

Status: Open.

#2543 In ROCm/aiter;

· sunway513 opened on Mar 31, 2026

[Issue]: AITER logs clang error for unsupported -amdgpu-coerce-illegal-types=1 flag during kernel compilation

Status: Open.

#2524 In ROCm/aiter;

· AndreasKaratzas opened on Mar 30, 2026

gfx1250: asm_gemm_a16w16 crashes with 'get_heuristic_kernel not find kernel for bf16gemm'

Status: Open.

#2519 In ROCm/aiter;

· sunway513 opened on Mar 29, 2026

[Bug] mha_batch_prefill returns all-zeros when KV cache > 8.4M pages (int32 overflow in CK element_space_size)

Status: Open.

#2517 In ROCm/aiter;

· ChangyiYang opened on Mar 28, 2026

Triton RMSNorm 5-10x slower than FlashInfer at batch<=1024 on MI355X

Status: Open.

#2516 In ROCm/aiter;

· sunway513 opened on Mar 28, 2026

GemmTuner fails on 76/92 BF16 shapes on gfx950 — ASM tile constraints and grid overflow

Status: Open.

#2515 In ROCm/aiter;

· sunway513 opened on Mar 28, 2026

Paged attention single-request latency 2x slower than FlashInfer at batch=1

Status: Open.

#2495 In ROCm/aiter;

· sunway513 opened on Mar 27, 2026

Need MLA decode micro-benchmark for MI355X vs B300 operator comparison

Status: Open.

#2493 In ROCm/aiter;

· sunway513 opened on Mar 26, 2026