biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] use AutoWeightsLoader in model load_weights (#15770)

Browse Source 
Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>
This commit is contained in:
rongfu.leng
2025-04-02 22:47:31 +08:00
committed by GitHub
parent 550b2801ad
commit e86c414d6a
4 changed files with 189 additions and 165 deletions
Download Patch File Download Diff File Expand all files Collapse all files

105
vllm/model_executor/models/bamba.py
View File

@@ -34,7 +34,7 @@ from vllm.utils import LayerBlockType
from .interfaces import (HasInnerState, IsHybrid, SupportsLoRA, SupportsPP,
SupportsQuant, SupportsV0Only)
from .utils import (is_pp_missing_parameter,
from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
@@ -363,6 +363,58 @@ class BambaModel(nn.Module):
hidden_states, _ = self.final_layernorm(hidden_states, residual)
return hidden_states
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
stacked_params_mapping = [
# (param_name, shard_name, shard_id)
("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())
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
if "A_log" in name:
name = name.replace("A_log", "A")
if ".self_attn." in name:
name = name.replace(".self_attn", "")
for param_name, weight_name, shard_id in stacked_params_mapping:
if weight_name not in name:
continue
name = name.replace(weight_name, param_name)
# Skip loading extra bias for GPTQ models.
if name.endswith(".bias") and name not in params_dict:
continue
# Skip layers on other devices.
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
weight_loader = param.weight_loader
weight_loader(param, loaded_weight, shard_id)
break
else:
# Skip loading extra bias for GPTQ models.
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 BambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
IsHybrid, SupportsV0Only, SupportsQuant):
@@ -502,52 +554,5 @@ class BambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
stacked_params_mapping = [
# (param_name, shard_name, shard_id)
("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())
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
if "A_log" in name:
name = name.replace("A_log", "A")
if ".self_attn." in name:
name = name.replace(".self_attn", "")
for param_name, weight_name, shard_id in stacked_params_mapping:
if weight_name not in name:
continue
name = name.replace(weight_name, param_name)
# Skip loading extra bias for GPTQ models.
if name.endswith(".bias") and name not in params_dict:
continue
# Skip layers on other devices.
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
weight_loader = param.weight_loader
weight_loader(param, loaded_weight, shard_id)
break
else:
# Skip loading extra bias for GPTQ models.
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
loader = AutoWeightsLoader(self)
return loader.load_weights(weights)