423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-29 18:20:35 +08:00
Code Issues Packages Projects Releases Wiki Activity

[model] Update ernie_vl to adapt new version (#9665)

Browse Source
This commit is contained in:
Xunpeng Xiao
2025-12-26 19:57:49 +08:00
committed by GitHub
parent a882e2d5fc
commit 3c17f2722c
5 changed files with 24 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llamafactory/model/loader.py
View File

@@ -205,10 +205,6 @@ def load_model(
if not is_trainable:
model.requires_grad_(False)
for param in model.parameters():
if param.data.dtype == torch.float32 and model_args.compute_dtype != torch.float32:
param.data = param.data.to(model_args.compute_dtype)
model.eval()
else:
model.train()

2
src/llamafactory/model/patcher.py
View File

@@ -158,7 +158,7 @@ def patch_config(
# do not cast data type of the model deepspeed zero3 without qlora
if not (is_deepspeed_zero3_enabled() and model_args.quantization_bit is None):
init_kwargs["torch_dtype"] = model_args.compute_dtype
init_kwargs["torch_dtype"] = "auto"
if init_kwargs["low_cpu_mem_usage"] and not is_fsdp_enabled(): # fsdp does not need device map
if "device_map" not in init_kwargs and model_args.device_map: