Qwen35moe perf - #312

Merged

wine99 merged 18 commits into

dev_backend_openvinofrom

Sep 24, 2026

Commits on Sep 16, 2026

ggml-openvino: add detailed inference profiling
wine99
committed
a132adc
ggml-openvino: use remote output tensors by default
wine99
committed
04fa107
ggml-openvino: optimize single-sequence recurrent state
wine99
committed
6a496d8

Commits on Sep 24, 2026

opt1: remove recurrent reset for single sequence, opt2: direct gdn outputs (break parallel sequence)
wine99
committed
8e65685
fix parallel sequences
wine99
committed
cb650d5
ggml-openvino: simplify graph cache key

ynimmaga
authored and
wine99
committed
743f61a
enable stateful for qwen35 single sequence
wine99
committed
8caa572
Fix after rebasing
wine99
committed
2583de0
Add k-requant option q4_asym64
wine99
committed
55776ec
Fix qwen35 llama-bench -p 0
wine99
committed
42da9c0
Simplify RESHAPE translation
wine99
committed
2c21f42
openvino: fuse MoE routing
wine99
committed
a231cc9
openvino: fuse GDN qk normalization
wine99
committed
7e51bba
openvino: enable GPU MoE fusion by default
wine99
committed
87ce31b
ggml-openvino: add cache_only mode to import cached compiled model on disk directly
wine99
committed
6266bd8
openvino : report the device allocation limit to ggml

lslusarczyk
authored and
wine99
committed
3cae90e
Merge branch 'dev_backend_openvino' into qwen35moe_perf
wine99
committed
f24f3d5
Fix windows build
wine99
committed
8cf1835