Spaces:
Runtime error
Runtime error
zetavg
commited on
Commit
·
03b5741
1
Parent(s):
37f2c31
support --trust_remote_code, resolves #6
Browse files- app.py +3 -0
- llama_lora/globals.py +2 -0
- llama_lora/models.py +24 -8
app.py
CHANGED
|
@@ -15,6 +15,7 @@ def main(
|
|
| 15 |
base_model: str = "",
|
| 16 |
data_dir: str = "",
|
| 17 |
base_model_choices: str = "",
|
|
|
|
| 18 |
# Allows to listen on all interfaces by providing '0.0.0.0'.
|
| 19 |
server_name: str = "127.0.0.1",
|
| 20 |
share: bool = False,
|
|
@@ -60,6 +61,8 @@ def main(
|
|
| 60 |
if base_model not in Global.base_model_choices:
|
| 61 |
Global.base_model_choices = [base_model] + Global.base_model_choices
|
| 62 |
|
|
|
|
|
|
|
| 63 |
Global.data_dir = os.path.abspath(data_dir)
|
| 64 |
Global.load_8bit = load_8bit
|
| 65 |
|
|
|
|
| 15 |
base_model: str = "",
|
| 16 |
data_dir: str = "",
|
| 17 |
base_model_choices: str = "",
|
| 18 |
+
trust_remote_code: bool = False,
|
| 19 |
# Allows to listen on all interfaces by providing '0.0.0.0'.
|
| 20 |
server_name: str = "127.0.0.1",
|
| 21 |
share: bool = False,
|
|
|
|
| 61 |
if base_model not in Global.base_model_choices:
|
| 62 |
Global.base_model_choices = [base_model] + Global.base_model_choices
|
| 63 |
|
| 64 |
+
Global.trust_remote_code = trust_remote_code
|
| 65 |
+
|
| 66 |
Global.data_dir = os.path.abspath(data_dir)
|
| 67 |
Global.load_8bit = load_8bit
|
| 68 |
|
llama_lora/globals.py
CHANGED
|
@@ -20,6 +20,8 @@ class Global:
|
|
| 20 |
base_model_name: str = ""
|
| 21 |
base_model_choices: List[str] = []
|
| 22 |
|
|
|
|
|
|
|
| 23 |
# Functions
|
| 24 |
train_fn: Any = train
|
| 25 |
|
|
|
|
| 20 |
base_model_name: str = ""
|
| 21 |
base_model_choices: List[str] = []
|
| 22 |
|
| 23 |
+
trust_remote_code = False
|
| 24 |
+
|
| 25 |
# Functions
|
| 26 |
train_fn: Any = train
|
| 27 |
|
llama_lora/models.py
CHANGED
|
@@ -37,16 +37,21 @@ def get_new_base_model(base_model_name):
|
|
| 37 |
# device_map="auto",
|
| 38 |
# ? https://github.com/tloen/alpaca-lora/issues/21
|
| 39 |
device_map={'': 0},
|
|
|
|
| 40 |
)
|
| 41 |
elif device == "mps":
|
| 42 |
model = AutoModelForCausalLM.from_pretrained(
|
| 43 |
base_model_name,
|
| 44 |
device_map={"": device},
|
| 45 |
torch_dtype=torch.float16,
|
|
|
|
| 46 |
)
|
| 47 |
else:
|
| 48 |
model = AutoModelForCausalLM.from_pretrained(
|
| 49 |
-
base_model_name,
|
|
|
|
|
|
|
|
|
|
| 50 |
)
|
| 51 |
|
| 52 |
tokenizer = get_tokenizer(base_model_name)
|
|
@@ -68,10 +73,16 @@ def get_tokenizer(base_model_name):
|
|
| 68 |
return loaded_tokenizer
|
| 69 |
|
| 70 |
try:
|
| 71 |
-
tokenizer = AutoTokenizer.from_pretrained(
|
|
|
|
|
|
|
|
|
|
| 72 |
except Exception as e:
|
| 73 |
if 'LLaMATokenizer' in str(e):
|
| 74 |
-
tokenizer = LlamaTokenizer.from_pretrained(
|
|
|
|
|
|
|
|
|
|
| 75 |
else:
|
| 76 |
raise e
|
| 77 |
|
|
@@ -100,13 +111,15 @@ def get_model(
|
|
| 100 |
peft_model_name_or_path = peft_model_name
|
| 101 |
|
| 102 |
if peft_model_name:
|
| 103 |
-
lora_models_directory_path = os.path.join(
|
|
|
|
| 104 |
possible_lora_model_path = os.path.join(
|
| 105 |
lora_models_directory_path, peft_model_name)
|
| 106 |
if os.path.isdir(possible_lora_model_path):
|
| 107 |
peft_model_name_or_path = possible_lora_model_path
|
| 108 |
|
| 109 |
-
possible_model_info_json_path = os.path.join(
|
|
|
|
| 110 |
if os.path.isfile(possible_model_info_json_path):
|
| 111 |
try:
|
| 112 |
with open(possible_model_info_json_path, "r") as file:
|
|
@@ -115,7 +128,8 @@ def get_model(
|
|
| 115 |
if possible_hf_model_name and json_data.get("load_from_hf"):
|
| 116 |
peft_model_name_or_path = possible_hf_model_name
|
| 117 |
except Exception as e:
|
| 118 |
-
raise ValueError(
|
|
|
|
| 119 |
|
| 120 |
Global.loaded_models.prepare_to_set()
|
| 121 |
clear_cache()
|
|
@@ -148,7 +162,8 @@ def get_model(
|
|
| 148 |
)
|
| 149 |
|
| 150 |
if re.match("[^/]+/llama", base_model_name):
|
| 151 |
-
model.config.pad_token_id = get_tokenizer(
|
|
|
|
| 152 |
model.config.bos_token_id = 1
|
| 153 |
model.config.eos_token_id = 2
|
| 154 |
|
|
@@ -166,7 +181,8 @@ def get_model(
|
|
| 166 |
|
| 167 |
|
| 168 |
def prepare_base_model(base_model_name=Global.default_base_model_name):
|
| 169 |
-
Global.new_base_model_that_is_ready_to_be_used = get_new_base_model(
|
|
|
|
| 170 |
Global.name_of_new_base_model_that_is_ready_to_be_used = base_model_name
|
| 171 |
|
| 172 |
|
|
|
|
| 37 |
# device_map="auto",
|
| 38 |
# ? https://github.com/tloen/alpaca-lora/issues/21
|
| 39 |
device_map={'': 0},
|
| 40 |
+
trust_remote_code=Global.trust_remote_code
|
| 41 |
)
|
| 42 |
elif device == "mps":
|
| 43 |
model = AutoModelForCausalLM.from_pretrained(
|
| 44 |
base_model_name,
|
| 45 |
device_map={"": device},
|
| 46 |
torch_dtype=torch.float16,
|
| 47 |
+
trust_remote_code=Global.trust_remote_code
|
| 48 |
)
|
| 49 |
else:
|
| 50 |
model = AutoModelForCausalLM.from_pretrained(
|
| 51 |
+
base_model_name,
|
| 52 |
+
device_map={"": device},
|
| 53 |
+
low_cpu_mem_usage=True,
|
| 54 |
+
trust_remote_code=Global.trust_remote_code
|
| 55 |
)
|
| 56 |
|
| 57 |
tokenizer = get_tokenizer(base_model_name)
|
|
|
|
| 73 |
return loaded_tokenizer
|
| 74 |
|
| 75 |
try:
|
| 76 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
| 77 |
+
base_model_name,
|
| 78 |
+
trust_remote_code=Global.trust_remote_code
|
| 79 |
+
)
|
| 80 |
except Exception as e:
|
| 81 |
if 'LLaMATokenizer' in str(e):
|
| 82 |
+
tokenizer = LlamaTokenizer.from_pretrained(
|
| 83 |
+
base_model_name,
|
| 84 |
+
trust_remote_code=Global.trust_remote_code
|
| 85 |
+
)
|
| 86 |
else:
|
| 87 |
raise e
|
| 88 |
|
|
|
|
| 111 |
peft_model_name_or_path = peft_model_name
|
| 112 |
|
| 113 |
if peft_model_name:
|
| 114 |
+
lora_models_directory_path = os.path.join(
|
| 115 |
+
Global.data_dir, "lora_models")
|
| 116 |
possible_lora_model_path = os.path.join(
|
| 117 |
lora_models_directory_path, peft_model_name)
|
| 118 |
if os.path.isdir(possible_lora_model_path):
|
| 119 |
peft_model_name_or_path = possible_lora_model_path
|
| 120 |
|
| 121 |
+
possible_model_info_json_path = os.path.join(
|
| 122 |
+
possible_lora_model_path, "info.json")
|
| 123 |
if os.path.isfile(possible_model_info_json_path):
|
| 124 |
try:
|
| 125 |
with open(possible_model_info_json_path, "r") as file:
|
|
|
|
| 128 |
if possible_hf_model_name and json_data.get("load_from_hf"):
|
| 129 |
peft_model_name_or_path = possible_hf_model_name
|
| 130 |
except Exception as e:
|
| 131 |
+
raise ValueError(
|
| 132 |
+
"Error reading model info from {possible_model_info_json_path}: {e}")
|
| 133 |
|
| 134 |
Global.loaded_models.prepare_to_set()
|
| 135 |
clear_cache()
|
|
|
|
| 162 |
)
|
| 163 |
|
| 164 |
if re.match("[^/]+/llama", base_model_name):
|
| 165 |
+
model.config.pad_token_id = get_tokenizer(
|
| 166 |
+
base_model_name).pad_token_id = 0
|
| 167 |
model.config.bos_token_id = 1
|
| 168 |
model.config.eos_token_id = 2
|
| 169 |
|
|
|
|
| 181 |
|
| 182 |
|
| 183 |
def prepare_base_model(base_model_name=Global.default_base_model_name):
|
| 184 |
+
Global.new_base_model_that_is_ready_to_be_used = get_new_base_model(
|
| 185 |
+
base_model_name)
|
| 186 |
Global.name_of_new_base_model_that_is_ready_to_be_used = base_model_name
|
| 187 |
|
| 188 |
|