[tawei-amd]

Problem

de3b12a9 ("feat: update hrx lib") rebuilt every Linux HRX engine on Ubuntu 24.04 + GCC 13.3, so the vendored libs need glibc 2.32–2.38 and GLIBCXX 3.4.32 (also log2@GLIBC_2.29). The SDK's manylinux_2_28 FLM job (glibc 2.28) fails linking flm against them. #822 fixed libwhisper_npu.so; this PR does the other 27 libs.

Change

Replaces the 27 src/lib/hrx/lib*.so rebuilt in the manylinux_2_28 environment, from FastFlowLM_IRON main (2d9079194, which includes the af74ea790 aiebu link fix) and the pinned HRX package flm-hrx-amdxdna-v2026.10.13:

• gcc-toolset-13 from ghcr.io/rocm/therock_build_manylinux_x86_64 against its glibc 2.28 sysroot; newer libstdc++ symbols come from libstdc++_nonshared.a (the same mechanism as FLM_DLL/build_manylinux.sh).
• .dll/.lib and libwhisper_npu.so are untouched.

	before (main)	after
max glibc	2.32–2.38	2.22
max GLIBCXX	3.4.32	3.4.21
max CXXABI	1.3.x	1.3.11

Verification

• FLM_DLL/check_abi.sh (FastFlowLM_IRON put note in bench #74/docs update v0.9.0 #76; ceiling glibc 2.28 / GLIBCXX 3.4.25 / CXXABI 1.3.11): 28/28 libs pass.
• No lib contains the aiebu Header version banner.
• Built flm (HRX) with these libs and ran on the NPU: Gemma4-E2B and Gemma4-12B answer correctly (these use the gemma4e / gemma4_12b engines plus q4nx, dequant, gemm, mha, lm_head). Other engines are built the same way but were not run here (no models locally).
• Not run: the SDK's manylinux_2_28 FLM link itself; please re-run that job on this branch.

Notes

• Engine Makefiles use -march=native (and -mavx512*), as in IRON CI; I built on a Zen 5 (Ryzen AI 9 HX 370) host.
• The RPATH of the libs contains the build-host HRX package path, as the existing vendored libs contain their CI runner paths; $ORIGIN entries come first.
• XRT libs (glibc 2.38) are not covered; they need an XRT dev tree in a 2.28 image.

🤖 Generated with Claude Code