openvino
[GPU] Fix fp16 intermediate overflow in fc_bf_tiled DQ scale path
#35228

Merged

[GPU] Fix fp16 intermediate overflow in fc_bf_tiled DQ scale path #35228

isanghao merged 8 commits into openvinotoolkit:master from ahnyoung-paul:fix_fp16_overflow_fc_bf_tiled

ahnyoung-paul added category: GPU

ahnyoung-paul added pr: needs tests

ahnyoung-paul added do_not_review

ahnyoung-paul added under_perf_check

ahnyoung-paul added do_not_merge

ahnyoung-paul requested a review 55 days ago

ahnyoung-paul force pushed from b9cd5b8e to 68a567c3 55 days ago

ahnyoung-paul force pushed from 68a567c3 to 25a1ee1f 55 days ago

ahnyoung-paul removed pr: needs tests

ahnyoung-paul removed do_not_review

ahnyoung-paul removed do_not_merge

ahnyoung-paul removed under_perf_check

jade-cho approved these changes on 2026-04-14

isanghao commented on 2026-04-14

[GPU] Fix fp16 overflow issue for FC bf tiled kernel

f6e5a950

fix unit test

17e2d34f

optimize fc bf titled kernel

bd208037

fix(GPU): reorder dq scale multiplication to match description

ff939f98

fix(GPU): reorder dq scale multiplication without float tmp

8bb137e8

fix(GPU): use float tmp for non-INT8 dq scale in fc_bf_tiled

b62fd64c

ahnyoung-paul force pushed from 0721ca5a to b62fd64c 49 days ago

remove redundant unit test

6b84f81b

isanghao commented on 2026-04-15

ahnyoung-paul changed the title ~~Fix fp16 overflow fc bf tiled~~ [GPU] Fix fp16 intermediate overflow in fc_bf_tiled DQ scale path 48 days ago

isanghao approved these changes on 2026-04-16

simplify non-INT8 DQ scale computation to single expression

3da775e6

ahnyoung-paul force pushed from 48300c2d to 3da775e6 48 days ago

isanghao enabled auto-merge 47 days ago

isanghao merged 0f352d09 into master 47 days ago

isanghao deleted the fix_fp16_overflow_fc_bf_tiled branch 47 days ago

Reviewers

isanghao

jade-cho

Assignees

No one assigned

Labels

category: GPU

Milestone

No milestone

openvino [GPU] Fix fp16 intermediate overflow in fc_bf_tiled DQ scale path #35228 Merged

[GPU] Fix fp16 intermediate overflow in fc_bf_tiled DQ scale path #35228

openvino
[GPU] Fix fp16 intermediate overflow in fc_bf_tiled DQ scale path
#35228

Merged