fix(ENG-EXPERT-STREAM): the CUDA blocker is a missing `vt::MatmulBTAlphaBeta`, and the Gemma-4 device-expert upload now refuses before it happens (#1126, #1205, #1197, #1218) - #1200

Merged

localai-bot merged 15 commits into

mainfrom

row/ENG-EXPERT-STREAM-1126

Aug 18, 2026

Commits on Aug 18, 2026

record(ENG-EXPERT-STREAM): the CUDA `DeviceMemoryInfo` override is not a one-line port, and #1126's own step 3 would be a defect (#1126, #1197)
mudler
committed
0acf2ca
fix(ENG-EXPERT-STREAM): the CUDA blocker is a missing `vt::MatmulBTAlphaBeta`, not a slower third path, and the arm now refuses by name (#1126, #1205)
mudler
committed
5031bd1
Merge origin/main into row/ENG-EXPERT-STREAM-1126
mudler
committed
45ff6be
record(ENG-EXPERT-STREAM): name the anchor re-audit and point the #1205 row at the right numbered point (#1205)
mudler
committed
7beada1
fix(ENG-EXPERT-STREAM): refuse the Gemma-4 device-expert upload before it happens, and stop telling a ROCm caller the kernel is missing (#1126, #1205, #1197, #1218)
mudler
committed
85a9c6e
Merge origin/main into row/ENG-EXPERT-STREAM-1126
mudler
committed
8743783
docs(ENG-EXPERT-STREAM): say why the guard test may re-register a backend after main, and name gemma4.cpp in full (#1205)
mudler
committed
ee2df78
Merge origin/main into row/ENG-EXPERT-STREAM-1126
mudler
committed
a16cd54
record(ENG-EXPERT-STREAM): re-derive the rocm_backend.hip anchor the merge moved, and check anchors semantically rather than only for range (#1205)
mudler
committed
e85eee1
fix(ENG-EXPERT-STREAM): put the merged #1126 index row back, byte-for-byte (#1205)
mudler
committed
2188c88
Merge origin/main into row/ENG-EXPERT-STREAM-1126
mudler
committed
385bd7b
fix(ENG-EXPERT-STREAM): drop the #1193 row a union merge duplicated (#1205)
mudler
committed
581942d
Merge origin/main into row/ENG-EXPERT-STREAM-1126
mudler
committed
c4d9cb2
Merge origin/main into row/ENG-EXPERT-STREAM-1126
mudler
committed
f1134c7
Merge origin/main into row/ENG-EXPERT-STREAM-1126
mudler
committed
00aab95

Navigation Menu

fix(ENG-EXPERT-STREAM): the CUDA blocker is a missing `vt::MatmulBTAlphaBeta`, and the Gemma-4 device-expert upload now refuses before it happens (#1126, #1205, #1197, #1218) - #1200

fix(ENG-EXPERT-STREAM): the CUDA blocker is a missing `vt::MatmulBTAlphaBeta`, and the Gemma-4 device-expert upload now refuses before it happens (#1126, #1205, #1197, #1218)#1200
localai-bot merged 15 commits into
mainfrom
row/ENG-EXPERT-STREAM-1126

Commits on Aug 18, 2026

record(ENG-EXPERT-STREAM): the CUDA `DeviceMemoryInfo` override is not a one-line port, and #1126's own step 3 would be a defect (#1126, #1197)

fix(ENG-EXPERT-STREAM): the CUDA blocker is a missing `vt::MatmulBTAlphaBeta`, not a slower third path, and the arm now refuses by name (#1126, #1205)

Merge origin/main into row/ENG-EXPERT-STREAM-1126

record(ENG-EXPERT-STREAM): name the anchor re-audit and point the #1205 row at the right numbered point (#1205)

fix(ENG-EXPERT-STREAM): refuse the Gemma-4 device-expert upload before it happens, and stop telling a ROCm caller the kernel is missing (#1126, #1205, #1197, #1218)

Merge origin/main into row/ENG-EXPERT-STREAM-1126

docs(ENG-EXPERT-STREAM): say why the guard test may re-register a backend after main, and name gemma4.cpp in full (#1205)

Merge origin/main into row/ENG-EXPERT-STREAM-1126

record(ENG-EXPERT-STREAM): re-derive the rocm_backend.hip anchor the merge moved, and check anchors semantically rather than only for range (#1205)

fix(ENG-EXPERT-STREAM): put the merged #1126 index row back, byte-for-byte (#1205)

Merge origin/main into row/ENG-EXPERT-STREAM-1126

fix(ENG-EXPERT-STREAM): drop the #1193 row a union merge duplicated (#1205)

Merge origin/main into row/ENG-EXPERT-STREAM-1126

Merge origin/main into row/ENG-EXPERT-STREAM-1126

Merge origin/main into row/ENG-EXPERT-STREAM-1126