load CLIP on training device
This commit is contained in:
parent
6b4ced2608
commit
3f4b3fa322
|
@ -234,7 +234,7 @@ def convert_models(ctx: ConversionContext, args, models: Models):
|
|||
inversion_name = inversion["name"]
|
||||
inversion_source = inversion["source"]
|
||||
inversion_source = fetch_model(ctx, f"{name}-inversion-{inversion_name}", inversion_source)
|
||||
convert_diffusion_textual_inversion(ctx, inversion_name, source, inversion_source)
|
||||
convert_diffusion_textual_inversion(ctx, inversion_name, model["source"], inversion_source)
|
||||
|
||||
except Exception as e:
|
||||
logger.error("error converting diffusion model %s: %s", name, e)
|
||||
|
|
|
@ -29,11 +29,11 @@ def convert_diffusion_textual_inversion(context: ConversionContext, name: str, b
|
|||
tokenizer = CLIPTokenizer.from_pretrained(
|
||||
base_model,
|
||||
subfolder="tokenizer",
|
||||
)
|
||||
).to(context.training_device)
|
||||
text_encoder = CLIPTextModel.from_pretrained(
|
||||
base_model,
|
||||
subfolder="text_encoder",
|
||||
)
|
||||
).to(context.training_device)
|
||||
|
||||
loaded_embeds = torch.load(embeds_file, map_location=context.map_location)
|
||||
|
||||
|
|
Loading…
Reference in New Issue