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

prepare does all this already for qlora?

Browse Source
This commit is contained in:
Wing Lian
2023-05-23 11:33:41 -04:00
parent 3b4d055edd
commit b9d07aa95a
1 changed files with 12 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
src/axolotl/utils/models.py
View File

@@ -204,17 +204,17 @@ def load_model(
"""### Post-processing on the model """### Post-processing on the model
Finally, we need to apply some post-processing on the 8-bit model to enable training, let's freeze all our layers, and cast the layer-norm in `float32` for stability. We also cast the output of the last layer in `float32` for the same reasons. Finally, we need to apply some post-processing on the 8-bit model to enable training, let's freeze all our layers, and cast the layer-norm in `float32` for stability. We also cast the output of the last layer in `float32` for the same reasons.
""" """
if cfg.adapter == "qlora": # if cfg.adapter == "qlora":
for param in model.parameters(): # for param in model.parameters():
param.requires_grad = False # freeze the model - train adapters later # param.requires_grad = False # freeze the model - train adapters later
if param.ndim == 1: # if param.ndim == 1:
# cast the small parameters (e.g. layernorm) to fp32 for stability # # cast the small parameters (e.g. layernorm) to fp32 for stability
param.data = param.data.to(torch.float32) # param.data = param.data.to(torch.float32)
class CastOutputToFloat(nn.Sequential): # class CastOutputToFloat(nn.Linear):
def forward(self, x): # def forward(self, x):
return super().forward(x).to(torch.float32) # return super().forward(x).to(torch.float32)
#
model.lm_head = CastOutputToFloat(model.lm_head) # model.lm_head = CastOutputToFloat(model.lm_head.in_features, model.lm_head.out_features, model.lm_head.bias)
if not tokenizer: if not tokenizer:
try: try:
@@ -255,7 +255,7 @@ def load_model(
embeddings_len = math.ceil(len(tokenizer) / 32) * 32 embeddings_len = math.ceil(len(tokenizer) / 32) * 32
model.resize_token_embeddings(embeddings_len) model.resize_token_embeddings(embeddings_len)
if cfg.adapter and load_in_8bit and not cfg.load_4bit: if ((cfg.adapter == "lora" and load_in_8bit) or cfg.adapter == "qlora") and not cfg.load_4bit:
logging.info("converting PEFT model w/ prepare_model_for_int8_training") logging.info("converting PEFT model w/ prepare_model_for_int8_training")
model = prepare_model_for_int8_training(model) model = prepare_model_for_int8_training(model)