bettter handling of llama model import
Browse files- scripts/finetune.py +19 -9
scripts/finetune.py
CHANGED
@@ -19,7 +19,7 @@ from peft import (
|
|
19 |
get_peft_model_state_dict, PeftModel,
|
20 |
)
|
21 |
from torch import nn
|
22 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
23 |
|
24 |
# add src to the pythonpath so we don't need to pip install this
|
25 |
from transformers.trainer_pt_utils import get_parameter_names
|
@@ -53,16 +53,23 @@ def load_model(base_model, model_type, tokenizer_type, cfg, adapter="lora"):
|
|
53 |
raise NotImplementedError(f"{adapter} peft adapter not available")
|
54 |
if "llama" in base_model:
|
55 |
from axolotl.flash_attn import replace_llama_attn_with_flash_attn
|
56 |
-
|
57 |
replace_llama_attn_with_flash_attn()
|
58 |
|
59 |
try:
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
except:
|
67 |
model = AutoModelForCausalLM.from_pretrained(
|
68 |
base_model,
|
@@ -72,7 +79,10 @@ def load_model(base_model, model_type, tokenizer_type, cfg, adapter="lora"):
|
|
72 |
)
|
73 |
|
74 |
try:
|
75 |
-
|
|
|
|
|
|
|
76 |
except:
|
77 |
tokenizer = AutoTokenizer.from_pretrained(base_model)
|
78 |
|
|
|
19 |
get_peft_model_state_dict, PeftModel,
|
20 |
)
|
21 |
from torch import nn
|
22 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, LlamaForCausalLM, LlamaTokenizer
|
23 |
|
24 |
# add src to the pythonpath so we don't need to pip install this
|
25 |
from transformers.trainer_pt_utils import get_parameter_names
|
|
|
53 |
raise NotImplementedError(f"{adapter} peft adapter not available")
|
54 |
if "llama" in base_model:
|
55 |
from axolotl.flash_attn import replace_llama_attn_with_flash_attn
|
|
|
56 |
replace_llama_attn_with_flash_attn()
|
57 |
|
58 |
try:
|
59 |
+
if "llama" in base_model:
|
60 |
+
model = LlamaForCausalLM.from_pretrained(
|
61 |
+
base_model,
|
62 |
+
load_in_8bit=cfg.load_in_8bit,
|
63 |
+
torch_dtype=torch.float16 if cfg.load_in_8bit else torch.float32,
|
64 |
+
device_map=cfg.device_map,
|
65 |
+
)
|
66 |
+
else:
|
67 |
+
model = getattr(transformers, model_type).from_pretrained(
|
68 |
+
base_model,
|
69 |
+
load_in_8bit=cfg.load_in_8bit,
|
70 |
+
torch_dtype=torch.float16 if cfg.load_in_8bit else torch.float32,
|
71 |
+
device_map=cfg.device_map,
|
72 |
+
)
|
73 |
except:
|
74 |
model = AutoModelForCausalLM.from_pretrained(
|
75 |
base_model,
|
|
|
79 |
)
|
80 |
|
81 |
try:
|
82 |
+
if "llama" in base_model:
|
83 |
+
tokenizer = LlamaTokenizer.from_pretrained(model)
|
84 |
+
else:
|
85 |
+
tokenizer = getattr(transformers, tokenizer_type).from_pretrained(model)
|
86 |
except:
|
87 |
tokenizer = AutoTokenizer.from_pretrained(base_model)
|
88 |
|