Enable q_scale calibration for deepseek v2 on Gaudi #1299
mengniwang95
merged
25 commits into
main
from
mengni/fp8_sdpa
Enable q_scale for deepseek and fix kv_cache_scheme
0065ee58
skip complex tensor on hpu
ada8a093
bug fix
81b2a5c7
fix bug
c3080998
Merge branch 'main' into mengni/fp8_sdpa
50ffe0de
![pre-commit-ci[bot]](https://avatars.githubusercontent.com/in/68672?v=4&s=60)
[pre-commit.ci] auto fixes from pre-commit.com hooks
1b67ad09
fix CI
ac1837d3
fix conditional statements
096d5b93
[pre-commit.ci] auto fixes from pre-commit.com hooks
5d926393
yiliu30
commented
on 2026-01-20
fix replace on cuda
30f77c54
use original kv
b9390db8
inherit replace module from the original module
2c6f76b5
[pre-commit.ci] auto fixes from pre-commit.com hooks
0f4c7990
fix CI
486054d5
refine code
9ce75f8b
add ut
fa9f8aa6
clean code
513a0196
[pre-commit.ci] auto fixes from pre-commit.com hooks
5fa153df
fix
903dbcc1
Merge branch 'main' into mengni/fp8_sdpa
9b61c5e3
add hpu ut req
5828ecb4
Revert "use original kv"
e64dd1b9
fix CI
136e26f6
yiliu30
approved these changes
on 2026-01-21
Update deepseek_v2.py
61755260
Update deepseek_v2.py
3c46d984
Login to write a write a comment.
Login via GitHub