more fixes
Browse files
src/axolotl/utils/models.py
CHANGED
@@ -184,7 +184,8 @@ def load_model(
|
|
184 |
for k, v in cfg.tokens.items():
|
185 |
tokenizer.add_special_tokens({k: v})
|
186 |
|
187 |
-
|
|
|
188 |
|
189 |
if cfg.adapter and load_in_8bit and not cfg.load_4bit:
|
190 |
logging.info("converting PEFT model w/ prepare_model_for_int8_training")
|
@@ -207,7 +208,10 @@ def load_model(
|
|
207 |
m.scales = m.scales.half()
|
208 |
m.bias = m.bias.half()
|
209 |
|
210 |
-
if torch.cuda.device_count() > 1 and int(os.getenv("WORLD_SIZE", "1")) > 1:
|
|
|
|
|
|
|
211 |
model.is_parallelizable = True
|
212 |
model.model_parallel = True
|
213 |
|
|
|
184 |
for k, v in cfg.tokens.items():
|
185 |
tokenizer.add_special_tokens({k: v})
|
186 |
|
187 |
+
# this should only be needed if you are messing with new tokens in the vocab
|
188 |
+
# model.resize_token_embeddings(len(tokenizer))
|
189 |
|
190 |
if cfg.adapter and load_in_8bit and not cfg.load_4bit:
|
191 |
logging.info("converting PEFT model w/ prepare_model_for_int8_training")
|
|
|
208 |
m.scales = m.scales.half()
|
209 |
m.bias = m.bias.half()
|
210 |
|
211 |
+
if torch.cuda.device_count() > 1 and int(os.getenv("WORLD_SIZE", "1")) > 1 and cfg.load_4bit:
|
212 |
+
# llama is PROBABLY model parallelizable, but the default isn't that it is
|
213 |
+
# so let's only set it for the 4bit, see
|
214 |
+
# https://github.com/johnsmith0031/alpaca_lora_4bit/blob/08b3fca4a4a9e0d3945be1bab4529f100a428636/finetune.py#L130-L133
|
215 |
model.is_parallelizable = True
|
216 |
model.model_parallel = True
|
217 |
|