GptOss experts implementation #43227
ArthurZucker
merged
35 commits into
main
from
gpt-oss-experts-impl
experts impl gpt oss
2aff4a88
no need to transpose dequantized experts
9958efba
skip test_reverse_loading_mapping
b23e1ffa
fix custom gating
e28f1555
Merge branch 'main' into gpt-oss-experts-impl
e57d0a87
revert transposition and simply support transposed experts to avoid m…
be08fe48
style
e1dba4d3
don't rely on weight shapes as they can be square matrices
0261a467
IlyasMoutawwakil
commented
on 2026-01-13
no need to relaod
5bd25c75
fallback to eager
846adcad
IlyasMoutawwakil
commented
on 2026-01-14
Update src/transformers/models/gpt_oss/modeling_gpt_oss.py
b1a71a79
vasqu
commented
on 2026-01-14
fix
9dbed89b
force 16 bytes alignmenet during weight loading
2f3fd11c
simplify logic
dd377e19
quantization conversions should be applied first
52e07786
IlyasMoutawwakil
commented
on 2026-01-15
avoid baddbmm as it is less performant / less optimizable by max-auto…
1c491124
no need for logger
4b0323ce
Merge branch 'main' into gpt-oss-experts-impl
aa34996f
add comment explaining limitation
f094c319
standarize operations and only reshape when needed
221f9bda
Merge branch 'main' into gpt-oss-experts-impl
944afb5c
fixup conversion and test
1fc01dc3
Update src/transformers/conversion_mapping.py
d8207138
force alignment docstring
71fdb18c
move default apply gate
e852cbb0
offsets
d698dcb4
vasqu
approved these changes
on 2026-01-16
Merge branch 'main' into gpt-oss-experts-impl
5c2ca3cc
add docs and make kernel_config optional
d6631bba
use reshapes as they are equivalent to views when memory is contiguous
4f7226d8
fix and better notes
21173033
reshapes instead of views
944a0eca
Merge branch 'main' into gpt-oss-experts-impl
1a0ea125
keep model saving and reloading in grouped_mm test to catch misalignm…
16e65366
Merge branch 'main' into gpt-oss-experts-impl
75ab2759
Merge branch 'main' into gpt-oss-experts-impl
711a652b
ArthurZucker
approved these changes
on 2026-01-22
Reviewers
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub