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:
co-authored by
Jasmond Loh
Zhuohan Li
parent
9841d48a10
commit
ab019eea75
@@ -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
|
||||
|
||||
|
||||
@@ -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
|
||||
|
||||
|
||||
@@ -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.
|
||||
|
||||
@@ -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()
|
||||
|
||||
@@ -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.
|
||||
|
||||
@@ -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.
|
||||
|
||||
@@ -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
|
||||
|
||||
|
||||
@@ -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
|
||||
|
||||
@@ -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
|
||||
|
||||
|
||||
@@ -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
|
||||
|
||||
|
||||
@@ -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].
|
||||
|
||||
@@ -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
|
||||
|
||||
|
||||
@@ -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
|
||||
|
||||
|
||||
Reference in New Issue
Block a user