[Model] use AutoWeightsLoader for olmoe,opt,orion,persimmon,phi3_small (#16548)

Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>
This commit is contained in:
rongfu.leng
2025-04-17 17:48:31 +00:00
committed by GitHub
parent a018e555fd
commit 5125d72f02
5 changed files with 216 additions and 193 deletions
+25 -19
View File
@@ -26,7 +26,7 @@ from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsPP
from .utils import (is_pp_missing_parameter,
from .utils import (AutoWeightsLoader, WeightsMapper, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
@@ -353,10 +353,29 @@ class Phi3SmallModel(nn.Module):
hidden_states = self.final_layernorm(hidden_states)
return hidden_states
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
params_dict = dict(self.named_parameters())
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if name.endswith(".bias") and name not in params_dict:
continue
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params
class Phi3SmallForCausalLM(nn.Module, SupportsPP):
_tied_weights_keys = ["lm_head.weight"]
hf_to_vllm_mapper = WeightsMapper(
orig_to_new_suffix={"rotary_emb.inv_freq": None})
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
@@ -448,21 +467,8 @@ class Phi3SmallForCausalLM(nn.Module, SupportsPP):
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
params_dict = dict(self.named_parameters())
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
if name.endswith(".bias") and name not in params_dict:
continue
if is_pp_missing_parameter(name, self):
continue
if "lm_head.weight" in name and self.config.tie_word_embeddings:
continue
param = params_dict[name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params
loader = AutoWeightsLoader(
self,
skip_prefixes=(["lm_head.weight"]
if self.config.tie_word_embeddings else None))
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)