ROCm/aiter

AI Tensor Engine for ROCm

Stars: 565
Forks: 585
Open beginner issues: 2
Indexed issues: 331
Avg merge: 3d 4h
Merged PRs (30d): 366

Dominant language: Python
License: MIT
Last GitHub push: Sep 20, 2026
Latest indexed: Sep 20, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed

Beginner friendly Everything

Open All

331 open issues indexed

Consolidate gfx942 unified-attention tuning for Gemma-4 (prefill #5650 + decode #5664/#5675) Open

ROCm/aiter#5696 · 1 assignee · Opened Sep 20, 2026
gemma-4-12B on gfx1201: ROCM_AITER_UNIFIED_ATTN is 4.65x faster at prefill than TRITON_ATTN (A/B on rocm/vllm:latest + aiter main) Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ROCm/aiter#5683 · 4 comments · Opened Sep 19, 2026
gfx942: non-persistent MLA decode returns wrong output for bf16 nhead16 / qseqlen 8 (MTP verify) Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

ROCm/aiter#5679 · Opened Sep 18, 2026
gfx942: sliding-window decode forced to 2D (no KV-split) — poor low-batch occupancy Open

ROCm/aiter#5674 · 1 comment · 1 assignee · Opened Sep 18, 2026
gfx942: head_size>=512 decode over-segments the KV split at low/mid batch (NUM_SEGMENTS ~2x optimum) Open

ROCm/aiter#5673 · 1 assignee · Opened Sep 18, 2026
bench_unified_attention.py: throughput/bandwidth metrics ignore sliding_window (inflated, above-roofline numbers) Open

ROCm/aiter#5662 · 1 assignee · Opened Sep 18, 2026
bench_unified_attention.py: fp8 KV -test reference is mis-scaled (false correctness failures) Open

ROCm/aiter#5661 · 1 assignee · Opened Sep 18, 2026
dsv4_bf16_tuned_gemm.csv contains only DeepSeek-V3 shapes; every DeepSeek-V4.1 dense GEMM lookup misses and falls back to torch Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

ROCm/aiter#5646 · Opened Sep 17, 2026
[Issue]: Flash Attention -Triton / MI35X (1GPU) CI fails during building docker image Open

ROCm/aiter#5611 · 3 comments · 1 assignee · Opened Sep 17, 2026
[ROCm][AITER]: get_2stage_cfgs: topk -= int(is_ep) incorrectly strips Fused Shared Expert column, producing wrong kernel config and gibberish output Open

Difficulty 3/5 1-2 days Newbie friendliness 68/100

ROCm/aiter#5595 · 1 comment · Opened Sep 16, 2026
Add fused MXFP4 dequantize-transpose-H16 requantization Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ROCm/aiter#5553 · Opened Sep 15, 2026
Add fused dual-layout H16 MXFP4 quantization Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ROCm/aiter#5552 · Opened Sep 15, 2026
Add 32x32 block-scaled MXFP4 quantization Open

Difficulty 4/5 3-5 days Newbie friendliness 68/100

ROCm/aiter#5547 · Opened Sep 15, 2026
Add fused SiLU-and-multiply backward kernel Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

ROCm/aiter#5541 · Opened Sep 15, 2026
Add packed FP4 logical transpose Open

Difficulty 5/5 Over a week Newbie friendliness 48/100

ROCm/aiter#5537 · Opened Sep 15, 2026
Add gfx950 stochastic MXFP4 quantization Open

Difficulty 4/5 3-5 days Newbie friendliness 50/100

ROCm/aiter#5530 · Opened Sep 15, 2026
BF16 causal prefill prefix differs across sequence lengths on MI300X Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ROCm/aiter#5515 · 2 comments · Opened Sep 14, 2026
[Feature] Package gfx942 BF16 sparse-decode donor for a supported vLLM adapter Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

ROCm/aiter#5503 · Opened Sep 14, 2026
[Bug] gfx942 CK SiLU MoE drops positive swiglu_limit in BF16/block-FP8 paths Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ROCm/aiter#5502 · 1 comment · Opened Sep 14, 2026
gfx950 paged MQA logits: next_n spends the tile on program count, leaving the kernel's own M-efficiency curve unused (M=32 vs 1.53x at M=128) Open

Difficulty 5/5 Over a week Newbie friendliness 45/100

ROCm/aiter#5477 · 1 comment · Opened Sep 12, 2026
Crash for DeepSeek-V4-Flash TP+EP: fused_moe tuned-config lookup keys on local (per-rank) expert count under EP, never matches CSVs tuned for global expert count Open

Difficulty 4/5 3-5 days Newbie friendliness 66/100

ROCm/aiter#5469 · Opened Sep 11, 2026
[MoE][BF16] CK 2stages rejects unaligned `inter_dim`, forcing callers to pad and pay the extra GEMM (gfx942, gfx950) Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

ROCm/aiter#5444 · Opened Sep 11, 2026
[Issue]: DeepSeek R1 MXFP4 core dumps due to compiler hazard after #5061 (0200ada1a) Open

bug Triton/Gluon

ROCm/aiter#5418 · 2 comments · 1 assignee · Opened Sep 10, 2026
gfx950: no FP16 counterpart to the BF16 preshuffled-B ASM GEMM Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ROCm/aiter#5417 · Opened Sep 10, 2026
CK 2-stage fused MoE requires callers to pad and zero-fill non-128-aligned intermediate sizes; intermediate_pad is ignored on this path Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

ROCm/aiter#5413 · 1 reaction · Opened Sep 10, 2026
[Issue]: arch_info fails to import with no GPU present, and blames JAX for it Open

Difficulty 4/5 3-5 days Newbie friendliness 72/100

ROCm/aiter#5407 · Opened Sep 10, 2026
[Config] bf16_tuned_gemm.csv ships no gfx942 rows - every bf16 GEMM falls back to torch on MI300X Open

Difficulty 3/5 1-2 days Newbie friendliness 68/100

ROCm/aiter#5380 · 1 comment · Opened Sep 9, 2026
[tune] a4w4_blockscale tuner has no --libtype axis, so CK Tile's MX GEMM can never compete Open

Difficulty 4/5 3-5 days Newbie friendliness 58/100

ROCm/aiter#5356 · Opened Sep 8, 2026
`get_num_xcds()` is hard-coded to 8 everywhere Open

Difficulty 3/5 1-2 days Newbie friendliness 72/100

ROCm/aiter#5344 · Opened Sep 8, 2026
[MI355X][MegaMoEV2] GLM-5.2 EP8 tuning handoff and C14 performance gap Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ROCm/aiter#5322 · Opened Sep 8, 2026
[Bug] AITER_FLYDSL_FORCE_REDUCE: comment states the inverse default, and the switch is inert on the tuned-config path Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

ROCm/aiter#5299 · Opened Sep 7, 2026
deepgemm_fp8_paged_mqa_logits crashes with SIGFPE for ChunkK=64 on gfx950 Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

ROCm/aiter#5298 · Opened Sep 6, 2026
[Perf] pa_decode_sparse: BLOCK_K=64 is suboptimal above the CU count — up to +43 % on gfx950 with a load-dependent choice Open

Difficulty 4/5 3-5 days Newbie friendliness 68/100

ROCm/aiter#5293 · 1 comment · Opened Sep 5, 2026
mhc_post silently computes wrong results when post_layer_mix / comb_res_mix are non-contiguous Open

Difficulty 4/5 3-5 days Newbie friendliness 68/100

ROCm/aiter#5292 · Opened Sep 5, 2026
[Issue]: gemm_a8w8_blockscale (Triton) returns NaN when K/NUM_KSPLIT is not a multiple of BLOCK_SIZE_K (gfx1201 DEFAULT config, K=12800 -> vLLM garbage on R9700) Open

Difficulty 3/5 1-2 days Newbie friendliness 74/100

ROCm/aiter#5286 · Opened Sep 5, 2026
[Perf] a8w8_blockscale tuned table for gfx950 appears to be generated without --splitK: re-tuning gains 13-61% on low-M shapes Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ROCm/aiter#5277 · Opened Sep 4, 2026
Untuned-shape logging exists only for bf16 (AITER_TUNE_GEMM); fp8/blockscale/a4w4 have none Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

ROCm/aiter#5267 · Opened Sep 3, 2026
JIT baton dead-holder check is ambiguous under --network host (pid + hostname collide across containers) Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ROCm/aiter#5266 · Opened Sep 3, 2026
One infeasible tune candidate fails every kernel in the module (ninja aborts the whole build) Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

ROCm/aiter#5265 · Opened Sep 3, 2026
CK Tile tune livelocks: the JIT build counts against the per-task --timeout Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

ROCm/aiter#5264 · Opened Sep 3, 2026
GEMM tuners: an interrupted run loses everything (results are only written at the end) Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

ROCm/aiter#5263 · 1 comment · Opened Sep 3, 2026
fused_moe_2stages passes num_rows to per_tensor_quant_hip, which asserts it is None Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ROCm/aiter#5256 · 1 comment · Opened Sep 3, 2026
[Perf] 0.1.21 slower than 0.1.19 for GLM-5.3 MXFP4 on gfx950 (−18 % prefill, −10 % throughput); FlyDSL MoE tuning? Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ROCm/aiter#5243 · 1 comment · Opened Sep 3, 2026
MXFP4 MoE: CK-Tile stage-2 mis-indexes E8M0 scales when inter_dim % 256 != 0, returning silently wrong results Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

ROCm/aiter#5233 · Opened Sep 3, 2026
[Issue]: gfx1201 (R9700) — Triton unified_attention overflows RDNA's LDS budget, and the ROCm 10.0 vLLM container ships no gfx1201 code objects Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ROCm/aiter#5229 · 12 comments · Opened Sep 3, 2026
[Bug]: MLA get_meta_param returns an lru_cache'd device tensor, corrupting the split indptr under CUDA graph replay Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

ROCm/aiter#5193 · Opened Sep 2, 2026
[Bug]: Memory access fault due to top_k_renorm_probs Open

Difficulty 4/5 3-5 days Newbie friendliness 35/100

ROCm/aiter#5192 · 1 comment · Opened Sep 1, 2026
update_config_files auto-resolves duplicate tuned-GEMM entries, then raises — unrecoverable in containers Open

ROCm/aiter#5184 · 1 assignee · Opened Sep 1, 2026
[gfx950] Request: port the fmha_fwd_with_sink asm family (currently gfx1250-gated) — d64+sinks prefill has no asm path on MI350X/MI355X Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ROCm/aiter#5161 · Opened Sep 1, 2026
[gfx950] CK mha_batch_prefill: no group-mode paged instances at head_dim 64 (blocks gpt-oss-style GQA prefill entirely) Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ROCm/aiter#5160 · Opened Sep 1, 2026

A short digest of beginner-friendly GitHub issues.