openvino: merge refactor PR - isolate op support, manage buffers, document API - #284

Draft

zhaixuejun1993 with Copilot wants to merge 135 commits into

dev_backend_openvinofrom

copilot/xuejunrefactor-ggml-openvino

Commits on Jul 27, 2026

OpenVINO backend: 1) enable gpt-oss moe on OV bk; 2) enable mxfp4 support

zhaixuejun1993
authored and
wine99
committed
1f88919
OpenVINO backend: disable TOPK_MOE op test

zhaixuejun1993
authored and
wine99
committed
4707c44
OpenVINO Backend: Add op FILL support

zhaixuejun1993
authored and
wine99
committed
bbe855b
OpenVINO backend: enable set rows with multi dims

zhaixuejun1993
authored and
wine99
committed
5c3b962
fix the name missmatch in setrow + view

zhaixuejun1993
authored and
wine99
committed
4b68df1
OpenVINO backend: enable op GGML_UNARY_OP_SIGMOID

zhaixuejun1993
authored and
wine99
committed
36d0bc3
OpenVINO Backend: enable SQR & SQRT

zhaixuejun1993
authored and
wine99
committed
ef525ec
OpenVINO backend: 1) ensure unique node names for OpenVINO; 2) add org_src to recorde the src ggml tensor for OpenVINO dynamic shape infer

zhaixuejun1993
authored and
wine99
committed
301882b
OpenVINO backend: enable fallback for openVINO to CPU backend

zhaixuejun1993
authored and
wine99
committed
e197045
OpenVINO backend: fix accurace issue in gemma3n arch test

zhaixuejun1993
authored and
wine99
committed
bec8a61
fix mpt failed case

zhaixuejun1993
authored and
wine99
committed
29d14d8
OpenVINO backend: clean nodeinfo

zhaixuejun1993
authored and
wine99
committed
c6d80e9
OpenVINO Backend: enable zero-size copy for view

zhaixuejun1993
authored and
wine99
committed
32d2b86
add concat ssm_conv in compute_dynamic_dim
wine99
committed
bf460cb
OpenVINO backend: disable EXP with FP32, which failed in op test. Root reason: the backend test initializes unary op inputs over a wide range, [-150, 150]. For FP32, exp(x) overflows around x ~= 88…

zhaixuejun1993
authored and
wine99
committed
745a95b
OpenVINO backend: fix CPY op test failed issue

zhaixuejun1993
authored and
wine99
committed
5f77b5f
OpenVINO backend: fix GATED_DELTA_NET op test failed issue

zhaixuejun1993
authored and
wine99
committed
eba9c7c
handle in-place op, handle qwen35 dynamic clearing of cache in cgraph
wine99
committed
05a92b8
handle qwen35 dynamic clearing of cache correctly
wine99
committed
cb09887
Enable qwen35 dense multi seq
wine99
committed
8546cb4
Fix qwen35 9b gqa
wine99
committed
c13deb5
Fix after rebase
wine99
committed
c21bf94
Disable SOLVE_TRI
wine99
committed
487db27
openvino: fix NEOX RoPE accuracy on GPU stateful (mixed-rank Multiply)

cavusmustafa
authored and
wine99
committed
867a993
OpenVINO backend: 1) remove the unique name in llama.cpp; 2) add new ov name in ov bk; 3) fix issue in arch test & op test with latest code update

zhaixuejun1993
authored and
wine99
committed
64c9d62
OpenVINO Backenb: remove changes in llama.cpp

zhaixuejun1993
authored and
wine99
committed
58ab393
Doc change (use x64 Native Tools Command Prompt for VS)

virajwad
authored and
wine99
committed
7591b24
Cleaner sentence

authored and
wine99
committed
8452765
OpenVINO Backend: cache key upgrade includes all src name

zhaixuejun1993
authored and
wine99
committed
88fbdc6
OpenVINO Backend: enable llama arch test on ci

zhaixuejun1993
authored and
wine99
committed
07f5918
OpenVINO Backend: move parameter node creating from decoder into translate

zhaixuejun1993
authored and
wine99
committed
37d76cb
OpenVINO Backend: create extra input ov node move from decoder to translate

zhaixuejun1993
authored and
wine99
committed
110e57a
fix for op regression due to is_model_splitted

suryasidd
authored and
wine99
committed
175f422

Commits on Jul 28, 2026

openvino: fix CPY writeback for recurrent state rollback
wine99
committed
e4085c0

Commits on Jul 30, 2026

qwen35 moe
wine99
committed
7162f3a
optimize MoE expert aggregation with ReduceSum
wine99
committed
820e72c

Commits on Jul 31, 2026

Commits on Aug 1, 2026

Commits on Aug 3, 2026

openvino: fix GPU mul_mat_id op tests
zhaixuejun1993
committed
ec01bd6
Merge pull request #271 from zhaixuejun1993/xuejun/fix-op-test-failed
zhaixuejun1993
authored
fd36389
ggml-openvino: add GGML_OPENVINO_RELEASE_WEIGHTS to reclaim host weight RSS on GPU

Mustafa Cavus
authored and
zhaixuejun1993
committed
fa089b1
ggml-openvino: stream weight requantization to cut the compile-time RSS peak

Mustafa Cavus
authored and
zhaixuejun1993
committed
98698af
ggml-openvino: avoid redundant token_embd requantization at compile

Mustafa Cavus
authored and
zhaixuejun1993
committed
9222c69
ggml-openvino: gate compile-memory optimizations behind GGML_OPENVINO_REDUCE_COMPILE_MEM

Mustafa Cavus
authored and
zhaixuejun1993
committed
7af5940
ggml-openvino: add frontend model cache (GGML_OPENVINO_MODEL_CACHE_DIR)

Mustafa Cavus
authored and
zhaixuejun1993
committed
72090f4
ggml-openvino: harden frontend model cache correctness
zhaixuejun1993
committed
2efe0b5
ggml-openvino: add memory optimization umbrella switch
zhaixuejun1993
committed
edf91ec
ggml-openvino: rename compiled model cache env
zhaixuejun1993
committed
b3c3d8b

Commits on Aug 4, 2026

Commits on Aug 7, 2026

Commits on Aug 10, 2026

ggml-openvino: extract op support policy into dedicated module
zhaixuejun1993
committed
91c707a
ggml-openvino: document full op policy table
zhaixuejun1993
committed
70f308f
ggml-openvino: manage buffer extras with RAII
zhaixuejun1993
committed
bdbca74
ggml-openvino: split weight buffer release helper
zhaixuejun1993
committed
4b16b4e
ggml-openvino: avoid duplicated buffer storage state
zhaixuejun1993
committed
fa364d8
ggml-openvino: drop unused buffer context name
zhaixuejun1993
committed
646a11e
ggml-openvino: split concrete buffer implementation
zhaixuejun1993
committed
befe071
ggml-openvino: document public backend API
zhaixuejun1993
committed
0dc7070
ggml-openvino: store host buffer type names
zhaixuejun1993
committed
2bd4cd4
ggml-openvino: tag buffer type contexts
zhaixuejun1993
committed
635a50b
ggml-openvino: share stateful execution helper
zhaixuejun1993
committed
2784cb5
ggml-openvino: move unused markers before returns
zhaixuejun1993
committed
83f4a30
ggml-openvino: share buffer type initialization
zhaixuejun1993
committed
0f8f384
ggml-openvino: trim stale main includes
zhaixuejun1993
committed
4492719
ggml-openvino: document logical device count
zhaixuejun1993
committed
ce60be9
ggml-openvino: document registry singleton ownership
zhaixuejun1993
committed
aa0f519

Commits on Aug 12, 2026

OpenVINO backend: 1) enable gpt-oss moe on OV bk; 2) enable mxfp4 support

zhaixuejun1993
authored and
wine99
committed
65db590
OpenVINO backend: disable TOPK_MOE op test

zhaixuejun1993
authored and
wine99
committed
44fcf81
OpenVINO Backend: Add op FILL support

zhaixuejun1993
authored and
wine99
committed
c867f51
OpenVINO backend: enable set rows with multi dims

zhaixuejun1993
authored and
wine99
committed
53ce035
fix the name missmatch in setrow + view

zhaixuejun1993
authored and
wine99
committed
c1d775f
OpenVINO backend: enable op GGML_UNARY_OP_SIGMOID

zhaixuejun1993
authored and
wine99
committed
186db4f
OpenVINO Backend: enable SQR & SQRT

zhaixuejun1993
authored and
wine99
committed
68da958
OpenVINO backend: 1) ensure unique node names for OpenVINO; 2) add org_src to recorde the src ggml tensor for OpenVINO dynamic shape infer

zhaixuejun1993
authored and
wine99
committed
7631aaf
OpenVINO backend: enable fallback for openVINO to CPU backend

zhaixuejun1993
authored and
wine99
committed
0529e9e
OpenVINO backend: fix accurace issue in gemma3n arch test

zhaixuejun1993
authored and
wine99
committed
bd532e4
fix mpt failed case

zhaixuejun1993
authored and
wine99
committed
026f1d0
OpenVINO backend: clean nodeinfo

zhaixuejun1993
authored and
wine99
committed
df1fbfd
OpenVINO Backend: enable zero-size copy for view

zhaixuejun1993
authored and
wine99
committed
b201559
add concat ssm_conv in compute_dynamic_dim
wine99
committed
116c048
OpenVINO backend: disable EXP with FP32, which failed in op test. Root reason: the backend test initializes unary op inputs over a wide range, [-150, 150]. For FP32, exp(x) overflows around x ~= 88…

zhaixuejun1993
authored and
wine99
committed
98c9e93
OpenVINO backend: fix CPY op test failed issue

zhaixuejun1993
authored and
wine99
committed
2736fa5
OpenVINO backend: fix GATED_DELTA_NET op test failed issue

zhaixuejun1993
authored and
wine99
committed
63fdf3d
handle in-place op, handle qwen35 dynamic clearing of cache in cgraph
wine99
committed
f223472
handle qwen35 dynamic clearing of cache correctly
wine99
committed
98cbce0
Enable qwen35 dense multi seq
wine99
committed
af70b49
Fix qwen35 9b gqa
wine99
committed
26f6a73
Fix after rebase
wine99
committed
1be15a9
Disable SOLVE_TRI
wine99
committed
3ab8fce
openvino: fix NEOX RoPE accuracy on GPU stateful (mixed-rank Multiply)

cavusmustafa
authored and
wine99
committed
1192ab5
OpenVINO backend: 1) remove the unique name in llama.cpp; 2) add new ov name in ov bk; 3) fix issue in arch test & op test with latest code update

zhaixuejun1993
authored and
wine99
committed
0ab29cb
OpenVINO Backenb: remove changes in llama.cpp

zhaixuejun1993
authored and
wine99
committed
f2088c5
Doc change (use x64 Native Tools Command Prompt for VS)

virajwad
authored and
wine99
committed
430d1db
Cleaner sentence

authored and
wine99
committed
0f0047e
OpenVINO Backend: cache key upgrade includes all src name

zhaixuejun1993
authored and
wine99
committed
7871c83
OpenVINO Backend: enable llama arch test on ci

zhaixuejun1993
authored and
wine99
committed
7da52bc
OpenVINO Backend: move parameter node creating from decoder into translate

zhaixuejun1993
authored and
wine99
committed
a5bd654
OpenVINO Backend: create extra input ov node move from decoder to translate

zhaixuejun1993
authored and
wine99
committed
3d999cb
fix for op regression due to is_model_splitted

suryasidd
authored and
wine99
committed
9dd17d0
openvino: fix CPY writeback for recurrent state rollback
wine99
committed
e2e52ac
qwen35 moe
wine99
committed
e2b116e
optimize MoE expert aggregation with ReduceSum
wine99
committed
0303d27
Skip GET_ROWS inaccurate test
wine99
committed
36feef4
openvino: fallback dynamic MUL_MAT_ID shapes

zhaixuejun1993
authored and
wine99
committed
2d6c8bf
OpenVINO Backend: fix error in arch test model mpt

zhaixuejun1993
authored and
wine99
committed
3885699
fix error caused by cpy in arch test model kimi-linear

zhaixuejun1993
authored and
wine99
committed
518e507
OpenVINO Backend: fix error in arch test model minimax-m3

zhaixuejun1993
authored and
wine99
committed
3485ef3
openvino: fix GPU mul_mat_id op tests

zhaixuejun1993
authored and
wine99
committed
4a0e802
ggml-openvino: add GGML_OPENVINO_RELEASE_WEIGHTS to reclaim host weight RSS on GPU

Mustafa Cavus
authored and
wine99
committed
524f869
ggml-openvino: stream weight requantization to cut the compile-time RSS peak

Mustafa Cavus
authored and
wine99
committed
ef9f854
ggml-openvino: avoid redundant token_embd requantization at compile

Mustafa Cavus
authored and
wine99
committed
868b022
ggml-openvino: gate compile-memory optimizations behind GGML_OPENVINO_REDUCE_COMPILE_MEM

Mustafa Cavus
authored and
wine99
committed
63e06ce
ggml-openvino: add frontend model cache (GGML_OPENVINO_MODEL_CACHE_DIR)

Mustafa Cavus
authored and
wine99
committed
747b4e9
ggml-openvino: harden frontend model cache correctness

zhaixuejun1993
authored and
wine99
committed
4be0fd7
ggml-openvino: add memory optimization umbrella switch

zhaixuejun1993
authored and
wine99
committed
db15f2e
ggml-openvino: rename compiled model cache env

zhaixuejun1993
authored and
wine99
committed
a0108b0
docs: document OpenVINO memory optimization env vars

zhaixuejun1993
authored and
wine99
committed
30905fe
ggml-openvino: fix Qwen3VL crash and deepstack correctness bug
wine99
committed
12f3c78
ggml-openvino: fix IMROPE inp_pos padding for NPU static shapes
wine99
committed
6491854
disable test-llama-archs tests.

ravi9
authored and
wine99
committed
289ddb3
openvino: gate fallback with env var

zhaixuejun1993
authored and
wine99
committed
c66a9c9
Revert changes in test-llama-archs
wine99
committed
00457ce
Apply editor config
wine99
committed
2e0c1be
reject CPY with quantized destination as unsupported
ravi9
authored
ae02f5e
Merge pull request #283 from ravi9/ravi9-patch-1
ravi9
authored
37b164f

Commits on Aug 14, 2026

Merge fork/xuejun-refactor: isolate op support, manage buffers, document API

Copilot
and
zhaixuejun1993
authored
0710a15