xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-11 19:46:42 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix qwen-moe packed_modules_mapping (#26634)

Browse Source 
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li 2025-10-11 23:21:33 +08:00 committed by GitHub
parent 9d6cff3ede
commit f0a30a067b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 23 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/model_executor/models/interfaces.py
View File

@ -325,7 +325,7 @@ class SupportsLoRA(Protocol):
# are empty by default. # are empty by default.
embedding_modules: ClassVar[dict[str, str]] = {} embedding_modules: ClassVar[dict[str, str]] = {}
embedding_padding_modules: ClassVar[list[str]] = [] embedding_padding_modules: ClassVar[list[str]] = []
packed_modules_mapping: ClassVar[dict[str, list[str]]] = {} packed_modules_mapping: dict[str, list[str]] = {}
# We can't use runtime_checkable with ClassVar for issubclass checks # We can't use runtime_checkable with ClassVar for issubclass checks

18
vllm/model_executor/models/qwen2_moe.py
View File

@ -534,11 +534,7 @@ class Qwen2MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
"q_proj", "q_proj",
"k_proj", "k_proj",
"v_proj", "v_proj",
], ]
"gate_up_proj": [
"gate_proj",
"up_proj",
],
} }
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
@ -547,6 +543,18 @@ class Qwen2MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
quant_config = vllm_config.quant_config quant_config = vllm_config.quant_config
self.config = config self.config = config
self.quant_config = quant_config self.quant_config = quant_config
# Only perform the following mapping when Qwen2MoeMLP exists
if (
getattr(config, "mlp_only_layers", [])
or config.shared_expert_intermediate_size > 0
):
self.packed_modules_mapping["gate_up_proj"] = (
[
"gate_proj",
"up_proj",
],
)
self.model = Qwen2MoeModel( self.model = Qwen2MoeModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
) )

14
vllm/model_executor/models/qwen3_moe.py
View File

@ -634,11 +634,7 @@ class Qwen3MoeForCausalLM(
"q_proj", "q_proj",
"k_proj", "k_proj",
"v_proj", "v_proj",
], ]
"gate_up_proj": [
"gate_proj",
"up_proj",
],
} }
fall_back_to_pt_during_load = False fall_back_to_pt_during_load = False
@ -649,6 +645,14 @@ class Qwen3MoeForCausalLM(
quant_config = vllm_config.quant_config quant_config = vllm_config.quant_config
self.config = config self.config = config
self.quant_config = quant_config self.quant_config = quant_config
# Only perform the following mapping when Qwen3MoeMLP exists
if getattr(config, "mlp_only_layers", []):
self.packed_modules_mapping["gate_up_proj"] = (
[
"gate_proj",
"up_proj",
],
)
self.model = Qwen3MoeModel( self.model = Qwen3MoeModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
) )