Add Model Revision Support (#1014)

Co-authored-by: Jasmond Loh <Jasmond.Loh@hotmail.com>
Co-authored-by: Zhuohan Li <zhuohan123@gmail.com>
This commit is contained in:
Jasmond L
2023-09-13 15:20:02 -07:00
committed by GitHub
co-authored by Jasmond Loh Zhuohan Li
parent 9841d48a10
commit ab019eea75
20 changed files with 75 additions and 35 deletions
+1 -1
View File
@@ -64,6 +64,6 @@ def get_model(model_config: ModelConfig) -> nn.Module:
else:
# Load the weights from the cached or downloaded files.
model.load_weights(model_config.model, model_config.download_dir,
model_config.load_format)
model_config.load_format, model_config.revision)
model = model.cuda()
return model.eval()
+3 -2
View File
@@ -288,7 +288,8 @@ class AquilaForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
tp_size = get_tensor_model_parallel_world_size()
tensor_model_parallel_rank = get_tensor_model_parallel_rank()
q_proj_shard_size = (self.config.hidden_size // tp_size)
@@ -305,7 +306,7 @@ class AquilaForCausalLM(nn.Module):
state_dict = self.state_dict()
for name, loaded_weight in hf_model_weights_iterator(
model_name_or_path, cache_dir, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "rotary_emb.inv_freq" in name:
continue
+3 -2
View File
@@ -303,13 +303,14 @@ class BaiChuanBaseForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
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, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "rotary_emb.inv_freq" in name:
continue
+3 -2
View File
@@ -279,11 +279,12 @@ class BloomForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
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, load_format):
model_name_or_path, cache_dir, load_format, revision):
if name == "lm_head.weight":
# Since hidden_states are parallelized, we need to
# load lm_head.weight in parallel.
+3 -2
View File
@@ -420,7 +420,8 @@ class FalconForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
tp_size = (get_tensor_model_parallel_world_size())
tp_rank = get_tensor_model_parallel_rank()
@@ -452,7 +453,7 @@ class FalconForCausalLM(nn.Module):
state_dict = self.state_dict()
for name, loaded_weight in hf_model_weights_iterator(
model_name_or_path, cache_dir, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "query_key_value" in name:
loaded_weight = convert_pyslice_to_tensor(loaded_weight)
loaded_weight_size = loaded_weight.size()
+3 -2
View File
@@ -231,14 +231,15 @@ class GPT2LMHeadModel(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
tensor_model_parallel_world_size = (
get_tensor_model_parallel_world_size())
tensor_model_parallel_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, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "lm_head.weight" in name:
# GPT-2 ties the weights of the embedding layer and the final
# linear layer.
+3 -2
View File
@@ -259,14 +259,15 @@ class GPTBigCodeForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
tensor_model_parallel_world_size = (
get_tensor_model_parallel_world_size())
tensor_model_parallel_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, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "lm_head.weight" in name:
# GPT-2 ties the weights of the embedding layer and the final
# linear layer.
+3 -2
View File
@@ -222,11 +222,12 @@ class GPTJForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
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, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "attn.bias" in name or "attn.masked_bias" in name:
continue
+3 -2
View File
@@ -231,11 +231,12 @@ class GPTNeoXForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
tensor_model_parallel_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, load_format):
model_name_or_path, cache_dir, load_format, revision):
if ("attention.bias" in name or "attention.masked_bias" in name
or "rotary_emb.inv_freq" in name):
continue
+3 -2
View File
@@ -233,12 +233,13 @@ class InternLMForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
tensor_model_parallel_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, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "rotary_emb.inv_freq" in name:
continue
+3 -2
View File
@@ -271,7 +271,8 @@ class LlamaForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
tp_size = get_tensor_model_parallel_world_size()
tensor_model_parallel_rank = get_tensor_model_parallel_rank()
q_proj_shard_size = (self.config.hidden_size // tp_size)
@@ -288,7 +289,7 @@ class LlamaForCausalLM(nn.Module):
state_dict = self.state_dict()
for name, loaded_weight in hf_model_weights_iterator(
model_name_or_path, cache_dir, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "rotary_emb.inv_freq" in name:
continue
+3 -2
View File
@@ -244,12 +244,13 @@ class MPTForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
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, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "Wqkv" in name:
# NOTE(woosuk): MPT's fused QKV has the shape of
# [3 * num_heads * head_size, hidden_size].
+3 -2
View File
@@ -297,12 +297,13 @@ class OPTForCausalLM(nn.Module):
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto"):
load_format: str = "auto",
revision: Optional[str] = None):
tensor_model_parallel_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, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "lm_head.weight" in name:
continue
+2 -1
View File
@@ -251,13 +251,14 @@ class QWenLMHeadModel(nn.Module):
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto",
revision: Optional[str] = None,
):
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, load_format):
model_name_or_path, cache_dir, load_format, revision):
if "rotary_emb.inv_freq" in name:
continue
+8 -3
View File
@@ -83,6 +83,7 @@ def prepare_hf_model_weights(
cache_dir: Optional[str] = None,
use_safetensors: bool = False,
fall_back_to_pt: bool = True,
revision: Optional[str] = None,
):
# Download model weights from huggingface.
is_local = os.path.isdir(model_name_or_path)
@@ -94,7 +95,8 @@ def prepare_hf_model_weights(
hf_folder = snapshot_download(model_name_or_path,
allow_patterns=allow_patterns,
cache_dir=cache_dir,
tqdm_class=Disabledtqdm)
tqdm_class=Disabledtqdm,
revision=revision)
else:
hf_folder = model_name_or_path
hf_weights_files = glob.glob(os.path.join(hf_folder, allow_patterns))
@@ -107,7 +109,8 @@ def prepare_hf_model_weights(
return prepare_hf_model_weights(model_name_or_path,
cache_dir=cache_dir,
use_safetensors=False,
fall_back_to_pt=False)
fall_back_to_pt=False,
revision=revision)
if len(hf_weights_files) == 0:
raise RuntimeError(
@@ -120,6 +123,7 @@ def hf_model_weights_iterator(
model_name_or_path: str,
cache_dir: Optional[str] = None,
load_format: str = "auto",
revision: Optional[str] = None,
) -> Iterator[Tuple[str, torch.Tensor]]:
use_safetensors = False
use_np_cache = False
@@ -140,7 +144,8 @@ def hf_model_weights_iterator(
model_name_or_path,
cache_dir=cache_dir,
use_safetensors=use_safetensors,
fall_back_to_pt=fall_back_to_pt)
fall_back_to_pt=fall_back_to_pt,
revision=revision)
if use_np_cache:
# Currently np_cache only support *.bin checkpoints