Skip to content
This repository has been archived by the owner on Jan 24, 2024. It is now read-only.

fix(model): reduce memory usage while loading in half-precision #79

Merged
merged 2 commits into from
Mar 23, 2023
Merged

fix(model): reduce memory usage while loading in half-precision #79

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7428abd
fix(model): reduce memory usage while loading in half-precision
orionji Mar 23, 2023
1b98035
fix(model): update comments
orionji Mar 23, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 2 additions & 6 deletions basaran/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -323,8 +323,8 @@ def load_model(
kwargs["device_map"] = "auto"
kwargs["load_in_8bit"] = load_in_8bit

# Override the dtype to float16 as required by bitsandbytes.
if load_in_8bit:
# Cast all parameters to float16 if quantization is enabled.
if half_precision or load_in_8bit:
kwargs["torch_dtype"] = torch.float16

# Support both decoder-only and encoder-decoder models.
Expand All @@ -333,10 +333,6 @@ def load_model(
except ValueError:
model = AutoModelForSeq2SeqLM.from_pretrained(name_or_path, **kwargs)

# Cast all parameters to half-precision if required.
if half_precision:
model = model.half()

# Check if the model has text generation capabilities.
if not model.can_generate():
raise TypeError(f"{name_or_path} is not a text generation model")
Expand Down