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

Enable safetensors loading for all models (#974)

Browse Source
This commit is contained in:
Zhuohan Li
2023-09-07 15:49:52 -07:00
committed by GitHub
parent c07ece5ca4
commit c957c741d9
18 changed files with 143 additions and 83 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/model_executor/models/mpt.py
View File

@@ -10,7 +10,8 @@ from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention import PagedAttentionWithALiBi
from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.weight_utils import (hf_model_weights_iterator,
from vllm.model_executor.weight_utils import (convert_pyslice_to_tensor,
hf_model_weights_iterator,
load_tensor_parallel_weights)
from vllm.model_executor.parallel_utils.parallel_state import (
get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
@@ -243,12 +244,12 @@ class MPTForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
use_np_cache: bool = False):
load_format: str = "auto"):
tp_world_size = get_tensor_model_parallel_world_size()
tp_rank = get_tensor_model_parallel_rank()
state_dict = self.state_dict()
for name, loaded_weight in hf_model_weights_iterator(
model_name_or_path, cache_dir, use_np_cache):
model_name_or_path, cache_dir, load_format):
if "Wqkv" in name:
# NOTE(woosuk): MPT's fused QKV has the shape of
# [3 * num_heads * head_size, hidden_size].
@@ -260,7 +261,7 @@ class MPTForCausalLM(nn.Module):
num_heads = total_num_heads // tp_world_size
head_start = tp_rank * num_heads
head_end = (tp_rank + 1) * num_heads
loaded_weight = convert_pyslice_to_tensor(loaded_weight)
if name.endswith(".weight"):
loaded_weight = loaded_weight.view(3, total_num_heads,
head_size, hidden_size)