AITER v0.1.12.post2 release summary (2026-04-23)

[sunway513]

AITER v0.1.12.post2 — Release Summary

Personal-view summary of the v0.1.12.post2 release shipped 2026-04-23.

TL;DR

Released v0.1.12.post2 from release/v0.1.12 after merging two production blockers and rebuilding all 6 wheels under a manylinux_2_28 + torch==2.10.0 ABI contract.

• Tag: v0.1.12.post2 → commit 28a7b6a5c85f4f34d266c8bbb9af6045b93f28fd (immutable)
• Release: https://github.com/ROCm/aiter/releases/tag/v0.1.12.post2 (currently --prerelease, awaiting downstream confirmation before flipping to --latest)
• Wheels: 6 total (ROCm 7.0 / 7.1 / 7.2 × Python 3.10 / 3.12), all built in CI from the tagged commit
• Validated: end-to-end DSR1 serving in vllm/vllm-openai-rocm:v0.19.1 on both MI300X (gfx942) and MI355X (gfx950)

Fixes contained

#	Issue	Symptom	Fix
1	ROCm#2843	ImportError: cannot import name 'flash_attn_varlen_func' and c10_hip_check_implementation undefined symbol when loading aiter inside vllm/vllm-openai-rocm:v0.19.1	PR ROCm#2846 — removed C10_HIP_KERNEL_LAUNCH_CHECK() macro from gated_rmsnorm_quant_kernels.cu (signature drift between PyTorch versions)
2	ROCm#2864 / vllm#39485	RuntimeError: This GEMM is not supported! running DSR1 on MI300X via vLLM serving	Cherry-pick of ROCm#2645 (GEMM dispatch) onto release branch + hand-port of SynchronizedCache template (originally from ROCm#2221) into csrc/include/aiter_hip_common.h
3	(no issue)	GLIBCXX_3.4.32 symbol mismatch and c10::cuda::getCurrentCUDAStream(signed char) ABI drift between AITER wheel and vllm/vllm-openai-rocm:v0.19.1	Workflow rewrite to build on pytorch/manylinux2_28-builder:rocm7.X (AlmaLinux 8 + devtoolset-13, glibc 2.28) and pin torch==2.10.0+rocm7.1 for all three ROCm builds

Build / release workflow improvements

• PR ci(release): add manylinux2_28+ROCm builder path with auditwheel gate (fixes #2843 long-term) ROCm/aiter#2851 — added use_manylinux workflow input (merged earlier, used for trial wheels)
• PR ci(release): add torch_pin + torch_index_url inputs for ABI-targeted release builds ROCm/aiter#2875 — added torch_pin + torch_index_url workflow inputs (merged into main, future releases skip the manual side-branch hack)
• PR [release/v0.1.12] Cherry-pick #2645 (multi-arch CK GEMM dispatch) + SynchronizedCache backport for v0.1.12.post2 ROCm/aiter#2874 — final cherry-pick + ruff F821/F841 cleanups onto release/v0.1.12

The workflow now produces wheels named:

amd_aiter-0.1.12.post2+rocm7.X.manylinux.2.28.torch210-cp3YY-cp3YY-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl

Validation matrix

Stage	Tool	Result
AITER tests	aiter-test CI	PASS
SGLang downstream	sglang-downstream-test CI	PASS
Black + Ruff	checks CI	PASS
ATOM DeepSeek-R1 (MI325)	atom-test CI	PASS
ATOM gpt-oss-120b (MI35X)	atom-test CI	non-blocker (ATOM destroy_dist_env import not present on release/v0.1.12)
6-wheel build matrix	aiter-release CI	PASS (3 dispatches × 2 py = 6/6)
DSR1 E2E in vllm v0.19.1 (MI300X)	manual repro on f08u19	PASS
DSR1 E2E in vllm v0.19.1 (MI355X)	manual repro on f08u19	PASS
Symbol floor	auditwheel + objdump GLIBCXX/GLIBC ceiling check	PASS (≤ 3.4.29 / ≤ 2.34)

Lessons / follow-ups

1. Tag immutability rule held. Per Release tags are moving ROCm/aiter#2691, never moved v0.1.12.post2. Fix iteration happened on release/v0.1.12 branch, tag created only after CI confirmed.
2. Manylinux is mandatory. Ubuntu-22-built wheels carry GLIBCXX_3.4.32 from gcc 13, which breaks dlopen inside vllm-rocm containers (Ubuntu 22 base, libstdc++ tops at 3.4.30). All future production wheels must use pytorch/manylinux2_28-builder or equivalent.
3. PyTorch ABI pin is mandatory across ROCm versions. torch 2.10 vs 2.11 disagree on c10::cuda::getCurrentCUDAStream namespace masquerade signature. Wheels must declare and pin a torch version explicitly to avoid silent linker drift.
4. PEP 440 dev-segment digits-only. First trial run failed because 0.1.12.post2.dev2645torch210+rocm7.2.manylinux_2_28 is invalid; corrected to 0.1.12.post2.dev2645+rocm7.2.manylinux.2.28.torch210.
5. Cross-arch validation needed. First MI355X test silently passed because the renamed /tmp/torch210-wheel.whl was rejected by pip; required strict PEP 427 filename. Confirmed both MI300X and MI355X work from the same wheel after the filename fix.
6. vLLM worker spawn lacks an AITER-status log line. Several hours of debug burned because the worker silently fell back to non-AITER paths. Open vLLM upstream issue is on the follow-up list.
7. CI Atom gpt-oss test references newer API (destroy_dist_env) not present in release branches. ATOM dashboard model assumes head; non-blocker per the established rule "ATOM never blocks AITER release."

Outstanding

• Greg / Richard final confirmation of the tagged wheel (drop-in replacement for the dev2645 staging wheel they validated; same source, clean version stamp)
• Flip --prerelease=false --latest after confirmation
• Reply on Trial wheel: v0.1.12.post2 candidate (ROCm 7.2.2 + py3.12) available for smoke test ROCm/aiter#2843, 0.1.12 fails on DeepSeek R1 on MI300 ROCm/aiter#2864, vllm#39485 with the released wheel link
• Teams announcement in vLLM Atom Meetup chat
• File downstream version-bump PRs to vLLM and SGLang
• File CDN multi-segment 403 issue with Satya
• Cleanup throwaway pensun/post2-* side branches
• Open vLLM upstream issue: emit AITER-status log line at worker init