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

rename var and reformat

Browse Source
This commit is contained in:
Aman Karmani
2023-08-27 19:55:11 +00:00
parent 7fd662dd89
commit f319b0bc67
1 changed files with 3 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/axolotl/utils/models.py
View File

@@ -355,7 +355,7 @@ def load_model(
if hasattr(module, "weight"): if hasattr(module, "weight"):
module.to(torch.float32) module.to(torch.float32)
fix_dtype = not cfg.adapter needs_fa2_dtype = not cfg.adapter
if not cfg.gptq and ( if not cfg.gptq and (
(cfg.adapter == "lora" and load_in_8bit) (cfg.adapter == "lora" and load_in_8bit)
or (cfg.adapter == "qlora" and cfg.load_in_4bit) or (cfg.adapter == "qlora" and cfg.load_in_4bit)
@@ -364,13 +364,11 @@ def load_model(
model = prepare_model_for_kbit_training( model = prepare_model_for_kbit_training(
model, use_gradient_checkpointing=cfg.gradient_checkpointing model, use_gradient_checkpointing=cfg.gradient_checkpointing
) )
fix_dtype = True needs_fa2_dtype = True
# LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
# convert them back to fp16/bf16 for flash-attn compatibility. # convert them back to fp16/bf16 for flash-attn compatibility.
if fix_dtype and ( if needs_fa2_dtype and (cfg.flash_attention and cfg.is_llama_derived_model):
cfg.flash_attention and cfg.is_llama_derived_model
):
for name, module in model.named_modules(): for name, module in model.named_modules():
if "norm" in name: if "norm" in name:
module.to(cfg.torch_dtype) module.to(cfg.torch_dtype)