Skip to content

[Bugfix] LoRA: extend expert base_layer loading to Qwen3.5 and Step3.x #37114

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

Open
HollowMan6 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[Bugfix] LoRA: extend expert base_layer loading to Qwen3.5 and Step3.x #37114

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
7 changes: 5 additions & 2 deletions vllm/model_executor/models/qwen3_5.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -403,9 +403,12 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
loaded_params: set[str] = set()
expert_params_mapping = self.get_expert_mapping()
is_fused_expert = False
base_layer = (
"base_layer." if any(".base_layer." in name for name in params_dict) else ""
)
fused_expert_params_mapping = [
("experts.w13_weight", "experts.gate_up_proj", 0, "w1"),
("experts.w2_weight", "experts.down_proj", 0, "w2"),
(f"experts.{base_layer}w13_weight", "experts.gate_up_proj", 0, "w1"),
(f"experts.{base_layer}w2_weight", "experts.down_proj", 0, "w2"),
]
num_experts = (
self.config.num_experts if hasattr(self.config, "num_experts") else 0
Expand Down
7 changes: 5 additions & 2 deletions vllm/model_executor/models/qwen3_5_mtp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,9 +198,12 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
is_fused_expert = False
base_layer = (
"base_layer." if any(".base_layer." in name for name in params_dict) else ""
)
fused_expert_params_mapping = [
("experts.w13_weight", "experts.gate_up_proj", 0, "w1"),
("experts.w2_weight", "experts.down_proj", 0, "w2"),
(f"experts.{base_layer}w13_weight", "experts.gate_up_proj", 0, "w1"),
(f"experts.{base_layer}w2_weight", "experts.down_proj", 0, "w2"),
]
num_experts = (
self.config.num_experts if hasattr(self.config, "num_experts") else 0
Expand Down
7 changes: 5 additions & 2 deletions vllm/model_executor/models/qwen3_vl_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,9 +183,12 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
loaded_params: set[str] = set()
expert_params_mapping = self.get_expert_mapping()
is_fused_expert = False
base_layer = (
"base_layer." if any(".base_layer." in name for name in params_dict) else ""
)
fused_expert_params_mapping = [
("experts.w13_weight", "experts.gate_up_proj", 0, "w1"),
("experts.w2_weight", "experts.down_proj", 0, "w2"),
(f"experts.{base_layer}w13_weight", "experts.gate_up_proj", 0, "w1"),
(f"experts.{base_layer}w2_weight", "experts.down_proj", 0, "w2"),
]
num_experts = self.config.num_experts
for name, loaded_weight in weights:
Expand Down
9 changes: 6 additions & 3 deletions vllm/model_executor/models/step3_text.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -467,11 +467,14 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
]
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
base_layer = (
"base_layer." if any(".base_layer." in name for name in params_dict) else ""
)

expert_params_mapping = [
(".moe.experts.w13_weight", ".moe.gate_proj.weight", "w1"),
(".moe.experts.w13_weight", ".moe.up_proj.weight", "w3"),
(".moe.experts.w2_weight", ".moe.down_proj.weight", "w2"),
(f".moe.experts.{base_layer}w13_weight", ".moe.gate_proj.weight", "w1"),
(f".moe.experts.{base_layer}w13_weight", ".moe.up_proj.weight", "w3"),
(f".moe.experts.{base_layer}w2_weight", ".moe.down_proj.weight", "w2"),
Copy link
Copy Markdown
Collaborator

@jeejeelee jeejeelee Mar 16, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It looks like these changes are unrelated. If needed, please submit a separate PR. Thank you.

Copy link
Copy Markdown
Contributor Author

@HollowMan6 HollowMan6 Mar 16, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you @jeejeelee for letting me know #36976 and #37019! I just removed all the Qwen3.5 LoRA fix in this PR and focus on the changes you mentioned here, those are extensions to #31104, please let me know what you think, thanks!

]

disable_moe_stacked_params = [data[1] for data in expert_params_mapping]
Expand Down
9 changes: 6 additions & 3 deletions vllm/model_executor/models/step3p5.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -641,12 +641,15 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:

params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
base_layer = (
"base_layer." if any(".base_layer." in name for name in params_dict) else ""
)

# Old packed 3D format: .moe.gate_proj.weight [num_experts, out, in]
expert_params_mapping = [
(".moe.experts.w13_weight", ".moe.gate_proj.weight", "w1"),
(".moe.experts.w13_weight", ".moe.up_proj.weight", "w3"),
(".moe.experts.w2_weight", ".moe.down_proj.weight", "w2"),
(f".moe.experts.{base_layer}w13_weight", ".moe.gate_proj.weight", "w1"),
(f".moe.experts.{base_layer}w13_weight", ".moe.up_proj.weight", "w3"),
(f".moe.experts.{base_layer}w2_weight", ".moe.down_proj.weight", "w2"),
]

# New per-expert format: .moe.experts.E.gate_proj.weight_packed [out, in]
Expand Down
11 changes: 7 additions & 4 deletions vllm/model_executor/models/step3p5_mtp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -181,14 +181,17 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
("gate_up_proj", "gate_proj", 0),
("gate_up_proj", "up_proj", 1),
]
params_dict = dict(self.named_parameters())
base_layer = (
"base_layer." if any(".base_layer." in name for name in params_dict) else ""
)

expert_params_mapping = [
(".moe.experts.w13_weight", ".moe.gate_proj.weight", "w1"),
(".moe.experts.w13_weight", ".moe.up_proj.weight", "w3"),
(".moe.experts.w2_weight", ".moe.down_proj.weight", "w2"),
(f".moe.experts.{base_layer}w13_weight", ".moe.gate_proj.weight", "w1"),
(f".moe.experts.{base_layer}w13_weight", ".moe.up_proj.weight", "w3"),
(f".moe.experts.{base_layer}w2_weight", ".moe.down_proj.weight", "w2"),
]

params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
Expand Down
Loading