youngkingdom/vllm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[Feature][Kernel]FusedMoE LoRA (#21229)

Browse Source 
Signed-off-by: wuchen <cntryroa@gmail.com>
Signed-off-by: banjuede <lmklhc@163.com>
Signed-off-by: Chen Wu <cntryroa@gmail.com>
Signed-off-by: Danielle Robinson <dmmaddix@amazon.com>
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Signed-off-by: bk-201 <joy25810@foxmail.com>
Co-authored-by: wuchen <wuchen@zetyun.com>
Co-authored-by: Nathan Van Gheem <vangheem@gmail.com>
Co-authored-by: banjuede <lmklhc@163.com>
Co-authored-by: Danielle Robinson <dmmaddix@amazon.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: bk-201 <joy25810@foxmail.com>
This commit is contained in:
Chen Wu
2025-10-21 11:01:37 +08:00
committed by GitHub
parent 3ada34f9cb
commit 5f6cbf60d6
28 changed files with 2084 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/model_executor/models/deepseek_v2.py
View File

@ -1313,6 +1313,17 @@ class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts, SupportsLoR
logits = self.logits_processor(self.lm_head, hidden_states)
return logits
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
# Params for weights, fp8 weight scales, fp8 activation scales
# (param_name, weight_name, expert_id, shard_id)
return SharedFusedMoE.make_expert_params_mapping(
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.n_routed_experts,
num_redundant_experts=0,
)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
stacked_params_mapping = [
# (param_name, shard_name, shard_id)

15
vllm/model_executor/models/gpt_oss.py
View File

@ -32,7 +32,7 @@ from vllm.model_executor.models.utils import sequence_parallel_chunk
from vllm.sequence import IntermediateTensors
from vllm.utils import cdiv
from .interfaces import SupportsEagle3, SupportsPP
from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
from .utils import (
AutoWeightsLoader,
WeightsMapper,
@ -627,7 +627,7 @@ class GptOssModel(nn.Module):
)
class GptOssForCausalLM(nn.Module, SupportsPP, SupportsEagle3):
class GptOssForCausalLM(nn.Module, SupportsPP, SupportsEagle3, SupportsLoRA):
packed_modules_mapping = {"qkv": ["q_proj", "k_proj", "v_proj"]}
hf_to_vllm_mapper = WeightsMapper(
@ -696,6 +696,17 @@ class GptOssForCausalLM(nn.Module, SupportsPP, SupportsEagle3):
logits = self.logits_processor(self.lm_head, hidden_states)
return logits
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
# Params for weights, weight scales, activation scales
# (param_name, weight_name, expert_id, shard_id)
return FusedMoE.make_expert_params_mapping(
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.num_local_experts,
num_redundant_experts=0,
)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
loader = AutoWeightsLoader(
self,

12
vllm/model_executor/models/olmoe.py
View File

@ -49,7 +49,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsPP
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (
AutoWeightsLoader,
is_pp_missing_parameter,
@ -349,8 +349,6 @@ class OlmoeModel(nn.Module):
("qkv_proj", "q_proj", "q"),
("qkv_proj", "k_proj", "k"),
("qkv_proj", "v_proj", "v"),
("gate_up_proj", "gate_proj", 0),
("gate_up_proj", "up_proj", 1),
]
params_dict = dict(self.named_parameters())
@ -433,17 +431,13 @@ class OlmoeModel(nn.Module):
return loaded_params
class OlmoeForCausalLM(nn.Module, SupportsPP):
class OlmoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
"k_proj",
"v_proj",
],
"gate_up_proj": [
"gate_proj",
"up_proj",
],
]
}
def __init__(