Update HRX binaries to FLM 1.0.5 - #714

Open

Abhishek-Varma wants to merge 22 commits into

ROCm:mainfrom

Abhishek-Varma:update_hrx_105

Commits on Sep 16, 2026

chore(hrx): regenerate Linux engines against HRX 2026.09.02

Abhishek-Varma
and
cursoragent
committed
8215b89
chore(hrx): regenerate Windows engine dll/lib against HRX 2026.09.02

Abhishek-Varma
and
cursoragent
committed
d385f67
fix(hrx): share buffer ownership so HRX engines match IRON ABI

Abhishek-Varma
and
cursoragent
committed
0015971
fix(hy-mt2): emit Model RAW Output so qualify sees responses

Abhishek-Varma
and
cursoragent
committed
3d62d7f
fix(hy-mt2): gate Model RAW Output on FLM_LOG_RAW_OUTPUT

Abhishek-Varma
and
cursoragent
committed
0c46d5d
chore(hrx): refresh Windows engine dll/lib (HRX 2026.09.02) after rebase onto main

Abhishek-Varma
and
cursoragent
committed
a154765
chore(hrx): rebuild Linux engines from IRON fix/hrx-prefill-use-flm-rt

Abhishek-Varma
and
cursoragent
committed
afcb12b
chore(hrx): refresh Windows engines against flm-hrx-amdxdna-v2026.09.09

Abhishek-Varma
and
cursoragent
committed
51897f4
chore(hrx): rebuild Linux engines against flm-hrx-amdxdna-v2026.09.09

Abhishek-Varma
and
cursoragent
committed
50fc0cc
fix(hrx): reference-count executables to free NPU contexts on model unload
Abhishek-Varma
committed
8f09d6b
chore(hrx): rebuild Linux engines from IRON reference-count executables fix

Abhishek-Varma
and
cursoragent
committed
ba9d31f
chore(hrx): bump public package to v2026.09.14 and refresh Windows engines

Abhishek-Varma
and
cursoragent
committed
a961c62
chore(hrx): rebuild Linux engines against flm-hrx-amdxdna-v2026.09.14

Abhishek-Varma
and
cursoragent
committed
1754cf6
chore(hrx): update release pin to v2026.09.15

jtuyls
authored and
Abhishek-Varma
committed
b2cb660
ci: re-enable HRX portable build

jtuyls
authored and
Abhishek-Varma
committed
6c357b3
ci: avoid future mtimes in HRX release extraction

jtuyls
authored and
Abhishek-Varma
committed
74c40a9
ci: gate optional flash engines by backend libs

jtuyls
authored and
Abhishek-Varma
committed
f6c3793
feat(hrx): regenerate engine libs against HRX v2026.09.15 and enable flash engines unconditionally

Abhishek-Varma
and
cursoragent
committed
0de4b14
fix(hrx): ship optimize_qwen3 base engine libs so qwen3vl_flash MHA attention works

Abhishek-Varma
and
cursoragent
committed
cffe4a6
refactor(hrx): remove FLM_HAS_*_FLASH gates; flash engines always compiled in

Abhishek-Varma
and
cursoragent
committed
60810c9
fix(hrx): regenerate all engine libs from optimize_qwen3 (28c78fc5) against matching hrx_cpp/npu_utils API

Abhishek-Varma
and
cursoragent
committed
2457b2e
fix(hrx): ship the working flash-compatible base+flash engine libs

Abhishek-Varma
and
cursoragent
committed
f6e8d2c

Navigation Menu

Update HRX binaries to FLM 1.0.5 - #714

Update HRX binaries to FLM 1.0.5#714
Abhishek-Varma wants to merge 22 commits into
ROCm:mainfrom
Abhishek-Varma:update_hrx_105

Commits on Sep 16, 2026

chore(hrx): regenerate Linux engines against HRX 2026.09.02

chore(hrx): regenerate Windows engine dll/lib against HRX 2026.09.02

fix(hrx): share buffer ownership so HRX engines match IRON ABI

fix(hy-mt2): emit Model RAW Output so qualify sees responses

fix(hy-mt2): gate Model RAW Output on FLM_LOG_RAW_OUTPUT

chore(hrx): refresh Windows engine dll/lib (HRX 2026.09.02) after rebase onto main

chore(hrx): rebuild Linux engines from IRON fix/hrx-prefill-use-flm-rt

chore(hrx): refresh Windows engines against flm-hrx-amdxdna-v2026.09.09

chore(hrx): rebuild Linux engines against flm-hrx-amdxdna-v2026.09.09

fix(hrx): reference-count executables to free NPU contexts on model unload

chore(hrx): rebuild Linux engines from IRON reference-count executables fix

chore(hrx): bump public package to v2026.09.14 and refresh Windows engines

chore(hrx): rebuild Linux engines against flm-hrx-amdxdna-v2026.09.14

chore(hrx): update release pin to v2026.09.15

ci: re-enable HRX portable build

ci: avoid future mtimes in HRX release extraction

ci: gate optional flash engines by backend libs

feat(hrx): regenerate engine libs against HRX v2026.09.15 and enable flash engines unconditionally

fix(hrx): ship optimize_qwen3 base engine libs so qwen3vl_flash MHA attention works

refactor(hrx): remove FLM_HAS_*_FLASH gates; flash engines always compiled in

fix(hrx): regenerate all engine libs from optimize_qwen3 (28c78fc5) against matching hrx_cpp/npu_utils API

fix(hrx): ship the working flash-compatible base+flash engine libs