Skip to content

[vllm compat] Separate renaming from conversion ops #43621

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Cyrilvallez merged 6 commits into from
Jan 30, 2026
Merged

[vllm compat] Separate renaming from conversion ops #43621

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
f9b8130
switch mapping to test poc
Cyrilvallez Jan 30, 2026
802b3c1
fix
Cyrilvallez Jan 30, 2026
45aea14
more
Cyrilvallez Jan 30, 2026
78f418e
fix
Cyrilvallez Jan 30, 2026
210cd93
comment
Cyrilvallez Jan 30, 2026
4e3213a
oupsi
Cyrilvallez Jan 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 13 additions & 12 deletions src/transformers/conversion_mapping.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,13 +85,13 @@ def _build_checkpoint_conversion_mapping():
),
],
"mixtral": [
WeightRenaming(".block_sparse_moe.gate", ".mlp.gate"),
WeightRenaming(".block_sparse_moe.", ".mlp."),
WeightConverter(
source_patterns=[
"block_sparse_moe.experts.*.w1.weight",
"block_sparse_moe.experts.*.w3.weight",
".experts.*.w1.weight",
".experts.*.w3.weight",
], # you give me a list of 2 keys, I collect a list of a list of tensors
target_patterns="mlp.experts.gate_up_proj", # target key gets the list of two tensors
target_patterns=".experts.gate_up_proj", # target key gets the list of two tensors
operations=[
MergeModulelist(
dim=0
Expand All @@ -101,9 +101,9 @@ def _build_checkpoint_conversion_mapping():
),
WeightConverter(
source_patterns=[
"block_sparse_moe.experts.*.w2.weight",
".experts.*.w2.weight",
],
target_patterns="mlp.experts.down_proj", # target key gets the list of two tensors
target_patterns=".experts.down_proj", # target key gets the list of two tensors
operations=[
MergeModulelist(
dim=0
Expand Down Expand Up @@ -142,18 +142,19 @@ def _build_checkpoint_conversion_mapping():
),
],
"phimoe": [
WeightRenaming(".block_sparse_moe.gate", ".mlp.router"),
WeightRenaming(".block_sparse_moe.", ".mlp."),
WeightRenaming(".gate.weight", ".router.weight"),
WeightConverter(
source_patterns=[
"block_sparse_moe.experts.*.w1.weight",
"block_sparse_moe.experts.*.w3.weight",
".experts.*.w1.weight",
".experts.*.w3.weight",
],
target_patterns="mlp.experts.gate_up_proj",
target_patterns=".experts.gate_up_proj",
operations=[MergeModulelist(dim=0), Concatenate(dim=1)],
),
WeightConverter(
source_patterns="block_sparse_moe.experts.*.w2.weight",
target_patterns="mlp.experts.down_proj",
source_patterns=".experts.*.w2.weight",
target_patterns=".experts.down_proj",
operations=[MergeModulelist(dim=0)],
),
],
Expand Down