tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

fix sharegpt tokenization, refactor tokenization debugging

Browse Source
This commit is contained in:
Wing Lian
2023-04-30 00:23:53 -04:00
parent c0f50d9c61
commit 5159d00a86
5 changed files with 63 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/axolotl/utils/models.py
View File

@@ -53,7 +53,7 @@ def load_model(
logging.info("patching with xformers attention")
hijack_llama_attention()
torch_dtype = (torch.float16 if cfg.load_in_8bit or cfg.fp16 else torch.float32,)
torch_dtype = torch.float16 if cfg.load_in_8bit or cfg.fp16 or cfg.bf16 else torch.float32
try:
if cfg.load_4bit:
from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
@@ -161,11 +161,11 @@ def load_model(
tokenizer.add_special_tokens({"pad_token": "[PAD]"})
os.environ["TOKENIZERS_PARALLELISM"] = "false"
if cfg.special_tokens:
for k, v in cfg.special_tokens.items():
setattr(tokenizer, k, v)
if cfg.tokens:
for k, v in cfg.tokens.items():
tokenizer.add_special_tokens({k: v})
if load_in_8bit and not cfg.load_4bit:
if load_in_8bit and cfg.load_4bit:
logging.info("converting model w/ prepare_model_for_int8_training")
model = prepare_model_for_int8_training(model)