Skip to content

Pull requests: ROCm/aiter

New pull request New
281 Open 2,473 Closed
281 Open 2,473 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

Optimize dynamic_per_group_scaled_quant via compile-time group_size t…
#3037 opened May 5, 2026 by RichardChamberlain1 Contributor • Draft
1
Add -o flag to bench_gmm.py for CSV output
#3036 opened May 5, 2026 by apicciau Contributor • Draft
1 task done
1
add rope/rotate_activation/fp4_quant_inplace fused kernel for dsv4
#3035 opened May 5, 2026 by junhaha666 Contributor Loading…
1 task
12
[TRITON] Add scattered-pointer Q4_K_M MoE matvec kernel for streaming inference
#3034 opened May 5, 2026 by ssubbotin Loading…
3 tasks done
1
Fix sqrsum store race condition in mhc_pre_gemm_sqrsum_kernel
#3033 opened May 5, 2026 by kkHuang-amd Contributor Loading…
2 of 3 tasks
3
fix(batch_prefill): OOB page table read fix via CK bump + regression tests (AICK-1171)
#3032 opened May 5, 2026 by Jeff-Huang Contributor Loading…
4 tasks done
1
attention.cu: guard out-of-head Q load in mfma16 paged-attention kernel
#3031 opened May 5, 2026 by JohnQinAMD Loading…
1 task
2
[Refactor] Clean up cktile wrappers, lru_cache sizes, get_padded_M, and fused_topk gating
#3030 opened May 4, 2026 by peymanr • Draft
1
[Perf][Kernel] Simplify per_1x32 MXFP4 quant path with token-count switch in fused_moe_2stages
#3029 opened May 4, 2026 by peymanr • Draft
1
[Refactor] Remove FlyDSL kernel dispatch layer and ck_moe_stage1 wrapper from fused_moe
#3028 opened May 4, 2026 by peymanr • Draft
1
[Perf][Kernel] Add gfx950 per_1x128 FP8 blockscale 1-stage ASM fast path in fused_moe_1stage
#3027 opened May 4, 2026 by peymanr • Draft
1
[Perf] Cache moe_sorting output buffers and quant partials to reduce HIP malloc overhead
#3026 opened May 4, 2026 by peymanr • Draft
1 task
1
[Perf] Cache GFX string and CU count at module import time
#3025 opened May 4, 2026 by peymanr • Draft
1 task
1
Fix batched_model_benchmark_shapes returning hidden/intermediate sizes swapped
#3019 opened May 4, 2026 by apicciau Contributor Loading…
1 task done
3
Remove async copy override from Triton test workflow ci:triton-300x ci:triton-355
#3018 opened May 4, 2026 by nidal567 Contributor Loading…
1 task done
@nidal567
1
CI: drop signal artifact, limit downstream on Checks run state
#3017 opened May 4, 2026 by leo-automation Loading…
1
1
[MoE] Cache split-K scratch buffers to avoid per-call hipMalloc
#3016 opened May 4, 2026 by frida-andersson Contributor Loading…
1
test: xfail test_moe_routing on gfx950 for known topk tie-breaking mismatch
#3015 opened May 4, 2026 by sunway513 Collaborator Loading…
3 tasks
5
Add fp8 mla decode kernel for sub_kv=64, sub_qh=8 (gqa_ratio=8, qseql…
#3014 opened May 3, 2026 by JohnNikolay84 Contributor Loading…
1 task
@JohnNikolay84
1
fix(mla): bypass fp8 qseqlen2 kernel precision issue on gfx950
#3013 opened May 3, 2026 by fangche123 Contributor Loading…
1 task
2
ci: replace deprecated zmq with pyzmq in CI scripts
#3007 opened May 3, 2026 by sunway513 Collaborator Loading…
3
Add HipKittens based nhead=32 MLA kernel on MI35x / gfx950
#3003 opened May 1, 2026 by hubertlu-tw Contributor • Draft
8 of 9 tasks
1
ci(nightly): fix wheel/image ABI mismatch + 0-test false-pass (run 25202894144)
#3002 opened May 1, 2026 by sunway513 Collaborator Loading…
1
Remove sorting for fmoe
#3001 opened May 1, 2026 by JohnNikolay84 Contributor • Draft
1 task
@JohnNikolay84
1
Replace QH16 bf16 kernel with a new one that does not use ptr_RP
#2999 opened May 1, 2026 by JohnNikolay84 Contributor Loading…
1 task
@JohnNikolay84
1
ProTip! Exclude everything labeled bug with -label:bug.