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

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

Browse Source 
Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>
This commit is contained in:
rongfu.leng
2025-04-18 01:48:31 +08:00
committed by GitHub
parent a018e555fd
commit 5125d72f02
5 changed files with 216 additions and 193 deletions
Download Patch File Download Diff File Expand all files Collapse all files

74
vllm/model_executor/models/persimmon.py
View File

@@ -46,7 +46,7 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsPP
from .utils import (is_pp_missing_parameter,
from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
@@ -221,7 +221,7 @@ class PersimmonModel(nn.Module):
quant_config = vllm_config.quant_config
self.vocab_size = config.vocab_size
self.config = config
self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
config.hidden_size)
self.start_layer, self.end_layer, self.layers = make_layers(
@@ -260,6 +260,38 @@ class PersimmonModel(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(remove_duplicate=False))
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
if "query_key_value" in name:
# copy from vllm/model_executor/models/bloom.py
# NOTE: Persimmon's fused QKV's output_dim has the shape of
# (num_heads * 3 * head_size), while the
# required shape is (3 * num_heads * head_size).
# Thus, we need weight conversion.
output_dim = getattr(param, "output_dim", None)
num_heads = self.config.num_attention_heads
if output_dim is not None:
loaded_weight_shape = loaded_weight.shape
loaded_weight = loaded_weight.view(
loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
loaded_weight_shape[output_dim + 1:])
loaded_weight = loaded_weight.transpose(
output_dim, output_dim + 1)
loaded_weight = loaded_weight.reshape(loaded_weight_shape)
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params
class PersimmonForCausalLM(nn.Module, SupportsPP):
@@ -315,39 +347,5 @@ class PersimmonForCausalLM(nn.Module, SupportsPP):
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
params_dict = dict(self.named_parameters(remove_duplicate=False))
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
if ("rotary_emb.cos_cached" in name
or "rotary_emb.sin_cached" in name):
# Models trained using ColossalAI may include these tensors in
# the checkpoint. Skip them.
continue
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
if "query_key_value" in name:
# copy from vllm/model_executor/models/bloom.py
# NOTE: Persimmon's fused QKV's output_dim has the shape of
# (num_heads * 3 * head_size), while the
# required shape is (3 * num_heads * head_size).
# Thus, we need weight conversion.
output_dim = getattr(param, "output_dim", None)
num_heads = self.config.num_attention_heads
if output_dim is not None:
loaded_weight_shape = loaded_weight.shape
loaded_weight = loaded_weight.view(
loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
loaded_weight_shape[output_dim + 1:])
loaded_weight = loaded_weight.transpose(
output_dim, output_dim + 1)
loaded_weight = loaded_weight.reshape(loaded_weight_shape)
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)