v0.1.6-rc0

ATOM v0.1.6 release candidate 0. Paired with AITER v0.1.17-rc0.

Wheels (3): rocm 7.0/7.1/7.2 (pure-python, py3-none-any)

atom-0.1.6rc0+rocm7.{0,1,2}-py3-none-any.whl — versioned per-rocm to pair with the matching aiter v0.1.17-rc0 wheel.

Base

main @ f9017b5. Includes the fastapi<0.137 cap (avoids prometheus-instrumentator serve-startup crash) and the MiniMax qknorm/backends fixes already on main.

Validation (paired with aiter v0.1.17-rc0, 5-model GSM8K 3-shot, MI355X)

DSR1 0.9507 / GLM-5 0.9371 / Kimi 0.9325 / MiniMax-M2.5 x3 mean 0.934 / Qwen3-235B x3 mean 0.875 — 9/9 PASS.

Note: GLM-5 requires aiter v0.1.17 #4064 (gluon mqa fix); serve GLM-5/large models with modest client concurrency to avoid first-request warmup timeouts.

v0.1.5

ATOM v0.1.5 — paired with AITER v0.1.16.

Paired container

rocm/atom-dev:atom0.1.5-aiter0.1.16 (ROCm 7.2.4 base, gfx942;gfx950, aiter @ 0730b33f)

Key fixes (vs v0.1.4)

• MiniMax-M2.5 torch.compile: restore the qkv.shape[0] <= 256 shape guard on the fused qknorm-allreduce path (#1299). Without it the fused custom collective runs unconditionally at TP>1 inside the compiled attention forward, forcing a top-level graph break that re-enters VllmBackend and hard-fails engine init (VllmBackend can only be called once).
• Compile-cache frozen-path guard: backends.py now skips any non-existent traced path (<frozen posixpath>, etc.) instead of only <string>/<frozen os> (#1299).
• fastapi cap >=0.115,<0.137 to avoid prometheus-fastapi-instrumentator crash on serve startup (fastapi >=0.137 is now on PyPI) (#1221).

Validation

mi355-gpu-15, GSM8K 3-shot flexible-extract on the released container:

• MiniMax-M2.5 0.9325, DeepSeek-R1-0528 0.9484 (smoke)
• Full gate (pilot, same fixes): DSR1 0.952 / GLM5 0.947 / Kimi 0.940 / MiniMax 3x mean 0.929 / Qwen3 3x mean 0.871

v0.1.4

ATOM v0.1.4

First ATOM release in the bi-weekly paired-release cadence with AITER (see cadence proposal). Also the first ATOM release that uses a release-branch workflow (prior versions tagged directly off main).

Same commit as v0.1.4-rc0 (26f23a0b) — zero delta after 3-day RC soak with no partner issues filed against the RC.

Paired AITER version

ATOM v0.1.4 pairs with AITER v0.1.15. The two releases share commit history and were jointly validated 5/5 PASS on GSM8K (mi355-gpu-15).

Commits since v0.1.3

26f23a0  Revert "Qwen3.5-35B-A3B-FP8: GDN decode lossy fast path + fused MRoPE QK (#838)"
f51d7be  Revert "Remove qkv 256 tok limitation (#999)"
e3c97b9  Remove qkv 256 tok limitation (#999)             ← reverted
4c4ae4f  fix(spec_decode): support DP attention with MTP in Deepseek V4 (#1001)
dfb8eda  Qwen3.5-35B-A3B-FP8: GDN decode lossy fast path + fused MRoPE QK (#838) ← reverted
6260cc1  Debug 'no such file or directory benchmark_matrix.json' (#994)
11be15e  [atom-vllm benchmark] refine model case name (#995)
9f9e97b  Debug name 'AWS_P0' is not defined (#991)

Highlights (net effective changes from v0.1.3)

• fix(spec_decode): DP attention with MTP support in DeepSeek-V4 (#1001)
• CI/benchmark fixes: AWS_P0 env, model case names, benchmark_matrix.json (#991 / #994 / #995)

Cherry-pick exclusions (intentional)

Two PRs reverted from v0.1.4 during joint validation:

• #999 (minimax qkv 256 tok limit): caused MiniMax-M2.5 server crash in joint test with AITER v0.1.15 (VllmBackend can only be called once). Will revisit in v0.1.5.
• #838 (Qwen3.5 GDN + fused MRoPE QK): reverted from v0.1.4 cycle to keep Kimi-K2.5 accuracy unaffected. Will revisit in v0.1.5.

Joint validation (paired with AITER v0.1.15)

GSM8K 3-shot flexible-extract, MI355X (gfx950), ATOM stack on 26f23a0b + AITER wheel 0.1.15rc0+rocm7.1.manylinux.2.28-cp312:

Model	Score	Threshold	Result
DeepSeek-R1-0528 (TP=8, fp8 KV)	0.9537	0.94	PASS
MiniMax-M2.5 (TP=2, fp8 KV)	0.9318	0.92	PASS
Qwen3-235B-A22B-FP8 (TP=8, fp8 KV)	0.8681	0.87	PASS (borderline)
GLM-5-FP8 (TP=8, fp8 KV)	0.9409	0.93	PASS
Kimi-K2.5-MXFP4 (TP=4, fp8 KV)	0.9219	0.92	PASS

5/5 PASS by upstream-canonical thresholds.

Container

Production: rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.4 (will be published after Nightly Docker Release workflow completes, ETA tonight)

Paired-pilot dev pin (immediate use): rocm/atom-dev:atom0.1.4-aiter0.1.15

Both ship the paired AITER v0.1.15 wheel pre-installed with matching triton 3.6.0 + flydsl 0.1.9.dev599.

Source

git clone --branch v0.1.4 https://github.com/ROCm/ATOM

Known issues

• Qwen3 borderline accuracy (0.8681 vs threshold 0.87): single-question swing within GSM8K 3-shot noise band. Tracking for v0.1.5.
• Kimi accuracy lower than ATOM HEAD upstream (0.9219 vs upstream-current 0.9393): root cause not fully bisected; possibly residual from #1001 DSv4 MTP change interaction with Kimi MoE path. Above threshold but below historical baseline. Will continue investigation in v0.1.5.

Feedback

• ATOM issues: https://github.com/ROCm/ATOM/issues — tag v0.1.4
• AITER issues (paired): https://github.com/ROCm/aiter/issues — tag v0.1.15
• Direct: peng.sun@amd.com, lingpeng.jin@amd.com

ATOM v0.1.4-rc0

ATOM v0.1.4-rc0

First release candidate for ATOM v0.1.4. Cut from release/v0.1.4 at 26f23a0b. This is the first ATOM release that uses a release branch workflow (prior versions tagged directly off main) and the first ATOM release paired with an AITER release in the bi-weekly cadence pilot (see ATOM bi-weekly release cadence proposal).

Final promotion (v0.1.4) planned for 2026-06-08 (~2 days RC soak).

Paired AITER version

ATOM v0.1.4-rc0 is validated to pair with AITER v0.1.15-rc0 (ABI-fixed wheels uploaded 2026-06-05).

Commits since v0.1.3

26f23a0  Revert "Qwen3.5-35B-A3B-FP8: GDN decode lossy fast path + fused MRoPE QK (#838)"
f51d7be  Revert "Remove qkv 256 tok limitation (#999)"
e3c97b9  Remove qkv 256 tok limitation (#999)             ← reverted
4c4ae4f  fix(spec_decode): support DP attention with MTP in Deepseek V4 (#1001)
dfb8eda  Qwen3.5-35B-A3B-FP8: GDN decode lossy fast path + fused MRoPE QK (#838) ← reverted
6260cc1  Debug 'no such file or directory benchmark_matrix.json' (#994)
11be15e  [atom-vllm benchmark] refine model case name (#995)
9f9e97b  Debug name 'AWS_P0' is not defined (#991)

Highlights (net effective changes from v0.1.3)

• fix(spec_decode): DP attention with MTP support in DeepSeek-V4 (#1001)
• CI/benchmark fixes: AWS_P0 env, model case names, benchmark_matrix.json (#991 / #994 / #995)

Cherry-pick exclusions (intentional)

Two PRs that landed on release/v0.1.4 were reverted during joint validation with AITER v0.1.15:

• #999 (minimax qkv 256 tok limit): caused MiniMax-M2.5 server crash in joint test with v0.1.15 (VllmBackend can only be called once). Reverted; will revisit in v0.1.5.
• #838 (Qwen3.5 GDN + fused MRoPE QK): kept in tree for Qwen3.5-35B-A3B-FP8 but reverted from v0.1.4 cycle to keep Kimi-K2.5 accuracy unaffected by the MRoPE path change. Will revisit in v0.1.5.

Joint validation (paired with AITER v0.1.15-rc0)

GSM8K 3-shot flexible-extract, MI355X (gfx950), ATOM stack on 26f23a0b + AITER wheel 0.1.15rc0+rocm7.1.manylinux.2.28-cp312:

Model	Score	Threshold	Result
DeepSeek-R1-0528 (TP=8, fp8 KV)	0.9537	0.94	PASS
MiniMax-M2.5 (TP=2, fp8 KV)	0.9318	0.92	PASS
Qwen3-235B-A22B-FP8 (TP=8, fp8 KV)	0.8681	0.87	PASS (borderline)
GLM-5-FP8 (TP=8, fp8 KV)	0.9409	0.93	PASS
Kimi-K2.5-MXFP4 (TP=4, fp8 KV)	0.9219	0.92	PASS

5/5 PASS by upstream-canonical thresholds.

Install / use

Container build paired with this RC will be pushed to rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.4-rc0 after Lingpeng cuts it. Until then, use the source tag:

git clone --branch v0.1.4-rc0 https://github.com/ROCm/ATOM

Pair with AITER:

pip install   --extra-index-url https://rocm.frameworks-devreleases.amd.com/whl-staging/gfx942-gfx950/   https://github.com/ROCm/aiter/releases/download/v0.1.15-rc0/amd_aiter-0.1.15rc0+rocm7.1.manylinux.2.28-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl

Known issues

• Qwen3 borderline accuracy (0.8681 vs threshold 0.87): single-question swing within GSM8K 3-shot noise band. Tracking for v0.1.5.
• Kimi accuracy lower than ATOM HEAD upstream (0.9219 vs upstream-current 0.9393 on SGLang side): root cause not fully bisected; possibly residual from #1001 DSv4 MTP change interaction with Kimi MoE path. Above threshold but below historical baseline. Will continue investigation in v0.1.5.

Next steps

• 2 days RC soak with paired AITER v0.1.15-rc0
• 2026-06-08: promote v0.1.4-rc0 → v0.1.4 (final, Latest) if no partner issues raised
• v0.1.5 / AITER v0.1.16 next cycle: revisit #999 + #838 fixes, investigate Kimi delta

Feedback

• ATOM issues: https://github.com/ROCm/ATOM/issues — tag v0.1.4-rc0
• AITER issues (paired): https://github.com/ROCm/aiter/issues — tag v0.1.15-rc0
• Direct: peng.sun@amd.com, lingpeng.jin@amd.com

— Peng + Lingpeng

---

Container image

Paired all-inclusive image is being pushed (~50GB) to:

rocm/atom-dev:atom0.1.4-rc0-aiter0.1.15-rc0

Contents:

• ATOM source at v0.1.4-rc0 (commit 26f23a0b)
• AITER wheel: amd_aiter-0.1.15rc0+rocm7.1.manylinux.2.28-cp312
• triton 3.6.0
• flydsl 0.1.9.dev599 (from AMD nightlies mirror)
• Base: rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.post

Usage:

docker pull rocm/atom-dev:atom0.1.4-rc0-aiter0.1.15-rc0
docker run -dt --privileged --device=/dev/kfd --device=/dev/dri   --shm-size=64g --ipc=host   -v /path/to/models:/data   --name atom_v014_rc0   rocm/atom-dev:atom0.1.4-rc0-aiter0.1.15-rc0 sleep infinity

No partner-side flydsl/triton install needed — both baked into the image.

v0.1.3

What's Changed

• CI: Verify ATOM tests on MI35x runners by @gyohuangxin in #444
• [Dashboard] Add column sorting and show Total Throughput in Data tab by @ChuanLi1101 in #447
• CI: reduce HF token exposure in atom-test logs by @gyohuangxin in #453
• Update GLM-5.md to refer to atom-dev docker container by @dwiddows in #454
• [plugin] refine full OOT validation & OOT benchmark by @zejunchen-zejun in #388
• Fix: support transformers 4.57.6 and 5.2.0 for gpt-oss by @PerryZhang01 in #419
• [plugin][oot] Add Kimi-K2.5 support by @gbyu-amd in #401
• [FIX] wrapper fused_qk_rmsnorm to ensure correct dispatch by @gbyu-amd in #456
• [plugin][triton version] align triton version with ATOM native by @zejunchen-zejun in #458
• fix deepseek tp 4 mtp3 mla metadata error by @junhaha666 in #460
• [dashboard] Fix column wrapping and polish UI by @carlushuang in #461
• Fix GLM-5-FP8 Indexer.weights_proj GEMM crash via exclude name remapping by @thpereir in #451
• [OOT Plugin] Fix qwen3.5 fp8 functionality and accuracy issue by @ganyi1996ppo in #448
• [OOT][Recipe] Update qwen3next recipe for performance measturement by @ganyi1996ppo in #466
• [Qwen3Next][Perf] add fused chunk split kernel for qkvzba and qkvz,ba case by @ganyi1996ppo in #457
• add GLM5 and Kimi2.5 to CI per PR by @valarLip in #470
• [Bugfix] remap_layer_name with kw parameter passing by @ganyi1996ppo in #469
• [Enhancement] load quark gpt_oss 120b wmxfp4 afp8 by @haoyangli0109 in #445
• Revert "add kernel comparison dashboard (ATOM OOT vs vLLM v0.18)" by @valarLip in #471
• CI: skip model download when cache exists by @gyohuangxin in #474
• Support enable_thinking False in v1/chat/completions mode by @ZhangLirong-amd in #472
• [Performance] Relaxed mtp by @haoyangli0109 in #411
• CI: improve atom test image and model cache handling by @gyohuangxin in #476
• [feat] Make ATOM work with SGLang out-of-tree by @zhuyuhua-v in #355
• fix: fix tp8 accuracy for gpt-oss by @PerryZhang01 in #449
• [plugin][OOT CI] refine OOT CI/dashboard/OOT docker release by @zejunchen-zejun in #459
• [plugin][OOT benchmark] set the final job for uploading data to dashboard by @zejunchen-zejun in #482
• Remove watermark overlay from dashboard by @functionstackx in #481
• fix(eagle): skip attn_metadata update for non-16-head models by @valarLip in #484
• refactor(dashboard): redesign based on data visualization best practices by @valarLip in #492
• CI: enable AMD CI monitor workflow in ATOM by @gyohuangxin in #504
• fix(dashboard): unify trends chart click to showPopover by @valarLip in #498
• [plugin][fix] fix kimi k2.5 weight loading by @gbyu-amd in #496
• CI: add lock-protected model downloads to ATOM tests by @gyohuangxin in #509
• [plugin][recipe] update kimi recipes by @gbyu-amd in #513
• Update README with installation instructions by @asleepzzz in #505
• support dual stream in prepare decode by @ZhangLirong-amd in #499
• Fix GLM5-MXFP4 loading error by @thpereir in #493
• [plugin][script] update env var for oot benchmark/test by @gbyu-amd in #516
• [plugin][upgrade vLLM] upgrade vLLM to 0.19.0 commit 2a6994 by @zejunchen-zejun in #483
• docs: Add model run guide and update GPU support table by @sunway513 in #520
• [dashboard] support light mode by @carlushuang in #529
• feat(ci): add SGLang image release and validation workflows by @zhuyuhua-v in #510
• fix: remove rmsnorm allreduce fusion in gpt-oss by @PerryZhang01 in #526
• Use SKILL to automatically enable Llama workloads for ATOM vLLM Plugin by @wuhuikx in #467
• feat(benchmark): add Llama-3.1-70B-Instruct-FP8 to nightly benchmark by @sunway513 in #536
• Add rocm-trace-lite (RTL) for GPU kernel profiling by @sunway513 in #535
• CI: inject AMD_HF_TOKEN on predownload runner by @gyohuangxin in #542
• add GLM-5.1-FP8 into accuracy ci by @valarLip in #519
• feat: replace triton fused_rms_fp8_group_quant with HIP kernel by @valarLip in #507
• [server] Refactor OpenAI server with tool calling, reasoning, and debug logging by @carlushuang in #489
• CI: prefer S3 aiter wheel before artifact fallback by @gyohuangxin in #538
• fix glm mtp weight by @jiayyu in #531
• feat: improve profiler robustness and enable stream parallelism for decode metadata by @valarLip in #547
• [plugin][dashboard] use nightly date tagged docker by @zejunchen-zejun in #503
• [Qwen3.5] add gemm fusion for qwen3.5 qkvzba bf16 case by @ganyi1996ppo in #543
• [UI] Scope ATOM watermark to Performance tab only by @sunway513 in #524
• docs: add Hermes Agent setup guide by @carlushuang in #551
• CI: Keep scheduled main runs from blocking push-triggered validation. by @gyohuangxin in #556
• fix: allow MiniMax-M2.5 loading in TP1 on MI355X by @benenzhu in #558
• [Qwem3.5] atom native support for qwen3.5 by @ganyi1996ppo in #517
• ci: fix benchmark workflow bugs and add GLM-5.1-MXFP4 by @valarLip in #562
• [recipe] ds r1 fp4 mtp3 model change by @seungrokj in #563
• ci: add MiniMax-M2.5-MXFP4 to nightly benchmark and accuracy test by @valarLip in #564
• [Qwen3Next/Qwen3.5] fuse gated_rmsnorm_quant by @ganyi1996ppo in #421
• [Feat][Plugin] Enable Sparse MLA and GLM-5 for vLLM-ATOM by @kliuae-amd in #399
• try to print server log by @jiayyu in #550
• [BugFix] enable deepseek r1 fp4 by @ZLkanyo009 in #527
• fix: disable gradient tracking on all nn.Parameter for inference by @valarLip in #574
• fix(docker): remove pinned torchvision/torchaudio wheels for vllm and sglang by @zhuyuhua-v in #575
• [Bugfix] Remove custom config class since transformers 5.2.0 already support qwen3.5 and fix bf16 loading issue by @ganyi1996ppo in #570
• fix mamba blocks ref count by @jiayyu in #581
• fix(dashboard): Fix SGLang benchmark workflow and integrate into dashboard by @zhuyuhua-v in #548
• integrate flydsl gdr decode by @ganyi1996ppo in #568
• [atom-vllm][atom-sglang][CI] build CI image on GPU machine instead of a build-only machine by @zejunchen-zejun in #561
• fix: fix thresholds for nightly model by @PerryZhang01 in #584
• [nightly][vllm] add GLM-5.1-FP8 to vLLM nightly coverage by @whx-sjtu in #569
• Suppport TBO in ATOM by @ZhangLirong-amd in #515
• [ATOM Test] fix file issue when finish lmeval acc test by @zejunchen-zejun in #585
• mxfp4 support for qwen3.5 by @ganyi1996ppo in #576
• [Feat] remove flatten in atom sglang mla like atom vllm mla by @ZLkanyo009 in #525
• fix kv cache fp8 issue by @ganyi1996ppo in #588
• fix: optimize attention metadata by @PerryZhang01 in #571
• fix: Add dashboard_model for SGLang deepseek TP4 by @zhuyuhua-v in #595
• Add Qwen3.5 FP4 to vLLM-ATOM nightly accuracy check and benchmark by @wuhuikx in #593
• fix: prefix benchmark artifacts to avoid downloading multi-GB traces by @valarLip in #596
• Update mori version and make tbo in nightly by @ZhangLirong-amd in #590
• [dashboard] show docker info in dashboard accuracy chart by @zejunchen-zejun in #587
• ci: add Qwen3.5-397B-A17B-MXFP4 to per-PR accuracy test by @valarLip in https://github.com/ROCm/ATOM/pu...

v0.1.2

What's Changed

• move from private repo to ROCm by @valarLip in #1
• Update logo by @carlushuang in #2
• update the link to the repo by @sunway513 in #3
• fix the example code cmd by @sunway513 in #5
• Update README.md and pyproject.toml by @andyluo7 in #4
• update readme for deepseek by @valarLip in #6
• support gpt oss by @junhaha666 in #7
• gpt_oss update: add fused_qk_rope_reshape_and_cache by @junhaha666 in #9
• Fix server startup message to show after model is loaded by @indianspeedster in #10
• fix gpt_oss accuracy drop by @junhaha666 in #12
• deepseek fp4 by @junhaha666 in #8
• gpt_oss: fix moe pad && use uniified attention 3d for full attention decode by @junhaha666 in #15
• move preprocess into threadpool avoid serial process by @HaonanWang98 in #19
• [perf] add qknorm_quant fusion for DS by @gbyu-amd in #18
• engine max_model_len : default is set to hf_config.max_position_embed… by @junhaha666 in #21
• [perf] add qknorm_quant and ar_rmsnorm fusion for DS by @gbyu-amd in #17
• reduce data for ScheduledBatch by @valarLip in #23
• fix port by @valarLip in #24
• Mla cache udpate by @junhaha666 in #20
• Add license and copyright headers by @ppalaniappan-amd in #30
• add ATOM_PROFILER by @amd-ruitang3 in #33
• use aiter hip fused_qk_rope_concat_and_cache_mla by @junhaha666 in #31
• add qwen3 moe model support by @gbyu-amd in #22
• support mtp stage 1: support draft model load by @jiayyu in #39
• update benchmark to inferencemax version by @HaonanWang98 in #38
• update server by @valarLip in #41
• refactor prepare_kv_indices by @valarLip in #43
• CI: Initial ATOM CI by @gyohuangxin in #40
• limit max_split_per_batch to 16 by @valarLip in #47
• support block size convert by @junhaha666 in #51
• fix num_kvcache_blocks error by @junhaha666 in #52
• Refactor arg_utils.py by @HaonanWang98 in #53
• Adapt lm-eval chat completion request. by @HaonanWang98 in #58
• CI: Add Dockerfile and nightly docker release pipeline by @gyohuangxin in #46
• remove global dict for request id in stream mode by @HaonanWang98 in #60
• CI: Add timeout for Nightly docker release by @gyohuangxin in #62
• CI: Add ROCm 7.2 preview nightly image by @gyohuangxin in #66
• update readme by @valarLip in #67
• update readme by @valarLip in #71
• CI: Add deepseek in ATOM tests by @gyohuangxin in #64
• Making BMM use fp4 weights by @omuhamma in #57
• llfp4 weight scale shuffle fix by @amirumoAMD in #74
• ds3.2: add one param for top_k_per_row_prefill ops by @PerryZhang01 in #77
• support aiter.gemm_a4w4 api changes by @junhaha666 in #70
• remove timeout for inter token latency by @HaonanWang98 in #79
• Enable INT4 QR for LLFP4 by @amirumoAMD in #76
• update utiliy by @valarLip in #81
• update_server by @valarLip in #82
• Update Dockerfile by @valarLip in #83
• Update Dockerfile by @valarLip in #87
• Graph: add param check for cuda graph capture by @PerryZhang01 in #85
• CI: Temporarily split gfx942 and gfx950 in nigthly docker release by @gyohuangxin in #89
• CI: Temporarily split gfx942 and gfx950 in nigthly docker release pushing by @gyohuangxin in #90
• CI: Increase timeout when building nightly docker image by @gyohuangxin in #91
• CI: Update dockerfile to use PREBUILD_KERNELS=1 by @gyohuangxin in #92
• remove async eng by @HaonanWang98 in #86
• Perf: save perfermance info in beautiful format by @PerryZhang01 in #80
• CI: Update base image to rocm/pytorch:latest in ATOM tests by @gyohuangxin in #88
• CI: Fix issues in nightly build pipeline by @gyohuangxin in #93
• CI: skip tests when building gfx942 nigthly docker image by @gyohuangxin in #96
• MLA: update aiter mqa kernel by @PerryZhang01 in #95
• CI: Fix node issues and use pre-download to accelerate tests by @gyohuangxin in #94
• clear schedule redundant variables by @inkcherry in #100
• Gpt oss triton moe by @junhaha666 in #98
• CI: Fix output issues and add gsm8k accuracy tests in CI by @gyohuangxin in #73
• CI: Fix issues by @gyohuangxin in #102
• CI: Add gpt-oss model by @gyohuangxin in #103
• feat: support pa_decode_gluon and refactor attention ops by @PerryZhang01 in #42
• CI: Fix CI issues by @gyohuangxin in #104
• PA: add ATOM_GPT_OSS_MODEL env for prefill attention by @PerryZhang01 in #105
• CI: Add MAX_JOBS when building the nightly image by @gyohuangxin in #106
• Update docker-release.yaml by @gyohuangxin in #107
• [Perf][Qwen3] Enable qknorm_rope_cache_quant fusion by @gbyu-amd in #65
• [fix] fix assert for Qwen3 by @gbyu-amd in #108
• DeepSeek v3.2: add sparse prefill mla and fix indexer rope by @junhaha666 in #109
• [CI] add Qwen3-235B-A22B-Instruct-2507-FP8 to CI by @gbyu-amd in #110
• Update Dockerfile to put aiter/atom under dir /app by @valarLip in #112
• adapt for opitimized ps_gluon_pa by @Bernard-Liu in #117
• fuse rmsnorm + quant for llama fp8 by @scxiao in #56
• code cleanup by @valarLip in #120
• fix deepseek accuracy when ENABLE_DS_QKNORM_QUANT_FUSION=1 by @junhaha666 in #121
• Update Dockerfile to install latest RCCL by @valarLip in #123
• Update atom_test.sh by @valarLip in #122
• CI: Enhance the docker release pipeline by @gyohuangxin in #125
• llfp4 fail temporary workaround by @amirumoAMD in #75
• CI: Fix the docker relase pipeline by @gyohuangxin in #131
• Fix torch 2.9 rlock error in torch compile by @ZhangLirong-amd in #114
• CI: Fix CI issues by @gyohuangxin in #135
• adapt for upstream gluon pa by @Bernard-Liu in #137
• [fix] fix gluon pa with bf16 kv by @gbyu-amd in #124
• CI: Speed up CI by using a nightly image instead of rebuilding each time by @gyohuangxin in #136
• [recipe] Add qwen3 235b recipe by @gbyu-amd in #111
• Fix defer output for conc>max_num_seqs by @valarLip in #134
• CI: Collect Accuracy tests summary by @gyohuangxin in #132
• [Triton] DS FP4/FP8 Triton fusion and GEMM optimization by @k50112113 in #119
• Fix DP issues in benchmark and support Mori in Moe by @ZhangLirong-amd in #72
• re-enable ATOM_ENABLE_DS_QKNORM_QUANT_FUSION regardless of ATOM_USE_T… by @k50112113 in #139
• fuse rmsnorm/quant and act_mul/quant for mxfp4 llama70B by @scxiao in #129
• use ck mha instead of triton unified_attention for sink and window by @junhaha666 in #118
• Fix attention mha logic error by @ZhangLirong-amd in #141
• CI: Add gpt-oss-120b 2 GPUs test by @gyohuangxin in #143
• shuffle_weights_update by @valarLip in #144
• Add the external facing doc draft for review by @ChuanLi1101 in #99
• CI: Re-enable dual-arch builds in the Docker nightly releases by @gyohuangxin...