ROCm/ATOM

View on GitHub

AiTer Optimized Model

Stars: 184
Forks: 149
Open beginner issues: 0
Indexed issues: 90
Avg merge: 2d 8h
Merged PRs (30d): 189

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

90 open issues indexed

[Perf Regression] 9 config(s) regressed @ d0708937 Open

performance regression

Difficulty 4/5 3-5 days Newbie friendliness 25/100

ROCm/ATOM#2323 · Opened Sep 21, 2026
[CI] Approving a PR restarts the CI pipeline Open

ROCm/ATOM#2322 · 1 assignee · Opened Sep 21, 2026
[Perf Regression] 3 config(s) regressed @ a8ddd6be Open

performance regression

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

ROCm/ATOM#2317 · Opened Sep 20, 2026
GLM-5.3-Flash does not work with vllm serve on rocm/atom-dev:vllm-latest Open

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

ROCm/ATOM#2316 · Opened Sep 20, 2026
[TAG] ATOM: Automatically update tags bi‑weekly Open

ROCm/ATOM#2159 · 1 assignee · Opened Sep 8, 2026
[Feature]: vllm atom plugin to support deepseek v4 flash 0731 Open

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

ROCm/ATOM#2136 · Opened Sep 3, 2026
[Feature]: Profiler: option to set a profiler window Open

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

ROCm/ATOM#2081 · Opened Aug 28, 2026
Add configurable priority-aware prefix-cache eviction Open

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

ROCm/ATOM#2079 · Opened Aug 28, 2026
[Issue]: GPT-OSS-120B fails on the ATOM SGLang stack (MXFP4 `triton_kernel` MoE) Open

ROCm/ATOM#2065 · 1 assignee · Opened Aug 27, 2026
[Issue]: ATOM: TP-DSA decode hang in the plugin forward-relay Open

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

ROCm/ATOM#2062 · Opened Aug 27, 2026
[Issue]: DeepSeek V4 TP-MoE does not work Open

ROCm/ATOM#2022 · 2 comments · 1 assignee · Opened Aug 25, 2026
[Issue]: Mori EP-MoE not CUDA-graph-capturable under TP-attention Open

Difficulty 4/5 3-5 days Newbie friendliness 42/100

ROCm/ATOM#2021 · Opened Aug 25, 2026
[Feature]: Multimodal support in ATOM Open

ROCm/ATOM#2014 · 5 comments · 1 assignee · Opened Aug 25, 2026
[Issue]: Open

Difficulty 4/5 3-5 days Newbie friendliness 54/100

ROCm/ATOM#1972 · Opened Aug 20, 2026
[Issue]: Kimi K3 ATOM+SGLang has ~309KB per token for KV while SGLang with no atom plugin uses only 26KB Open

ROCm/ATOM#1959 · 1 assignee · Opened Aug 19, 2026
[Issue]: Open

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

ROCm/ATOM#1957 · Opened Aug 19, 2026
[Issue]: GPT-OSS-120B fails on the ATOM SGLang stack (MXFP4 `triton_kernel` MoE) Open

Difficulty 4/5 3-5 days Newbie friendliness 64/100

ROCm/ATOM#1945 · 2 comments · Opened Aug 18, 2026
[Feature]: FP8 block-quant MoE rejects TP shards narrower than block_n (Qwen3.6-35B-A3B) Open

ROCm/ATOM#1944 · 1 assignee · Opened Aug 18, 2026
[Issue]: ATOM build holds **~16 GB/GPU more resident memory than SGLang without ATOM for the identical model** Open

Difficulty 5/5 Over a week Newbie friendliness 30/100

ROCm/ATOM#1935 · Opened Aug 17, 2026
[Feature]: DeepSeek V4 on ATOM Open

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

ROCm/ATOM#1934 · Opened Aug 17, 2026
[Feature]: Enable DP-attention + EP (mori) for Gated-DeltaNet MoE models on ATOM — pick up SGLang #31794 Open

ROCm/ATOM#1929 · 1 comment · 1 assignee · Opened Aug 17, 2026
[Documentation]: ATOM expansion should be consistent across documents Open

documentation

Difficulty 2/5 1-3 hours Newbie friendliness 55/100

ROCm/ATOM#1928 · Opened Aug 17, 2026
[Feature]: Add Kimi‑K3 support to the ATOM SGLang plugin (enable ATOM‑SGLang serving/benchmarking of Kimi‑K3) Open

ROCm/ATOM#1927 · 4 comments · 1 assignee · Opened Aug 17, 2026
Qwen3.5 cannot start through the SGLang plugin on main (two independent failures) Open

Difficulty 3/5 1-2 days Newbie friendliness 58/100

ROCm/ATOM#1918 · 1 comment · Opened Aug 17, 2026
How can I run the Kimi K3 model on multiple AMD MI308X servers ？ Open

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

ROCm/ATOM#1854 · 1 comment · Opened Aug 10, 2026
How can I install and embed ATOM inside an arbitrary vanilla vLLM container ？ Open

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

ROCm/ATOM#1847 · 9 comments · Opened Aug 10, 2026
[Issue]: Mooncake picks a link-local RoCE GID on ionic NICs, requiring a manual MC_GID_INDEX override Open

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

ROCm/ATOM#1748 · Opened Jul 30, 2026
[Feature]: Structure Output support Open

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

ROCm/ATOM#1696 · 1 reaction · Opened Jul 24, 2026
[Bug][DSV4 MTP/NEXTN] unified_kv_rope AttributeError during draft cudagraph capture (V4-Pro MTP3 / V4-Flash accuracy red on all PRs) Open

ROCm/ATOM#1661 · 2 comments · 2 assignees · Opened Jul 21, 2026
[Issue]: Kimi-K2.5-MXFP4-AttnFP8 fails to load with AssertionError: assert ATOM_USE_TRITON_GEMM or ATOM_USE_TRITON_MOE when strictly following official Kimi-K2.5.md guide Open

ROCm/ATOM#1582 · 8 comments · 1 assignee · Opened Jul 13, 2026
[DeepSeek-V4] index_topk=1024 sparse indexer can't recall verbatim lines from ~18k+ prompts (breaks agentic cwd/env recall) Open

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

ROCm/ATOM#1507 · 1 comment · Opened Jul 8, 2026
[Bug] DeepSeek-V4 TBO decode + CUDAGraph: hipBLASLt heuristic during capture aborts with HIP 900 ('operation not permitted when stream is capturing') Open

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

ROCm/ATOM#1506 · 1 comment · Opened Jul 7, 2026
Optimize the ATOM CI Execution Time Open

frameworks-devops

ROCm/ATOM#1497 · 3 comments · 1 assignee · Opened Jul 7, 2026
Docs: how to set up and run single-node 4P4D (prefill/decode disaggregation) with Mooncake (XGMI/TCP) Open

documentation

Difficulty 3/5 1-2 days Newbie friendliness 58/100

ROCm/ATOM#1483 · 1 comment · Opened Jul 6, 2026
[Issue]: Concurrent TP8's from QPX partitions fail Open

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

ROCm/ATOM#1462 · 4 comments · Opened Jul 3, 2026
MiniMax-M3-MXFP4: image requests rejected — processor has no chat template (multimodal eval blocked) Open

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

ROCm/ATOM#1407 · Opened Jun 30, 2026
[Issue]: DeepSeek-V4-Flash Memory access fault under 64/128 concurrency with 3000+ requests Open

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

ROCm/ATOM#1400 · Opened Jun 29, 2026
rocm/atom-dev:latest crashes on MiniMax-M3 MXFP4: AIter API mismatch in fused_qknorm_idxrqknorm Open

ROCm/ATOM#1347 · 2 comments · 1 assignee · Opened Jun 24, 2026
[BUG] Performance degradation and garbled text output when benchmarking DeepSeek-V4-Pro with ATOM backend on MI355X Open

Difficulty 4/5 3-5 days Newbie friendliness 42/100

ROCm/ATOM#1302 · 1 comment · Opened Jun 21, 2026
[Issue]: Running Minimax 2.5/2.7 using tip of tree (ATOM/AITER/vLLM) or latest nightly version with TP=8 on results in error: Invalid qk hidden dim layout Invalid qk hidden dim layout for qknorm_allreduce_fusion_kernel_2stage kernel: Open

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

ROCm/ATOM#1223 · Opened Jun 15, 2026
[Issue]: Using basic instructions to run simple example results in crash on gfx950 Open

Difficulty 4/5 3-5 days Newbie friendliness 38/100

ROCm/ATOM#1207 · 2 comments · Opened Jun 13, 2026
Real World ATOM Kimi K2.7-Code Production Inferencing Enablement Request Open

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

ROCm/ATOM#1188 · 1 comment · 2 reactions · Opened Jun 12, 2026
[Parity with CUDA vLLM, CUDA SGLang]: native ATOM llm-d Kubernetes Distributed Inferencing Orchestrator for Production LLM Inferencing [Functional Enablement + Guides + Nightly CI] Open

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

ROCm/ATOM#1187 · Opened Jun 12, 2026
[Parity with CUDA vLLM]: ATOM Nightly CI and/or per PR CI Function Calling Eval Open

Difficulty 4/5 3-5 days Newbie friendliness 38/100

ROCm/ATOM#1186 · Opened Jun 12, 2026
[Parity with CUDA SGLang] FP8 & FP4 AMD Multi-Node PD Disagg Nightly Regression Testing With AMD first class Pollara NICs with multiple prefill workers + multiple decode workers Open

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

ROCm/ATOM#1168 · 2 comments · 1 reaction · Opened Jun 10, 2026
[RFC]: Enable DeepSeek V4 in vLLM Plugin Open

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

ROCm/ATOM#1130 · Opened Jun 8, 2026
ATOM Development Roadmap (2026 Q3) Open

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

ROCm/ATOM#989 · 5 comments · Opened May 31, 2026
ATOM Development Roadmap (2026 Q2) Open

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

ROCm/ATOM#988 · Opened May 31, 2026
[Perf Regression] 47 config(s) regressed @ 24e4367b Open

performance regression

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

ROCm/ATOM#922 · Opened May 26, 2026
[Issue]: Running DeepSeek V3.2 on MI308x*8 with Atom-vLLM only 200-300K TPM in single-machine performance Open

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

ROCm/ATOM#896 · Opened May 24, 2026

ROCm/ATOM

Get new issues in your inbox