Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

fix: convert bfloat16 to float because of numpy incompatibility #53

Merged
merged 2 commits into from
Oct 3, 2024
Merged

fix: convert bfloat16 to float because of numpy incompatibility #53

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1453c3a
convert bfloat16 to float because of numpy incompatibility
stephantul Oct 3, 2024
48da995
Also add bfloat to vocab inference
stephantul Oct 3, 2024
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
12 changes: 11 additions & 1 deletion model2vec/distill/inference.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,7 +67,12 @@
"""
encodings = tokenizer(tokens, return_tensors="pt", padding=True, truncation=True).to(model.device)
encoded: BaseModelOutputWithPoolingAndCrossAttentions = model(**encodings)
out = encoded.last_hidden_state.cpu()
out: torch.Tensor = encoded.last_hidden_state.cpu()
# NOTE: If the dtype is bfloat 16, we convert to float32,
# because numpy does not suport bfloat16
# See here: https://github.com/numpy/numpy/issues/19808
if out.dtype == torch.bfloat16:
out = out.float()

Check warning on line 75 in model2vec/distill/inference.py

View check run for this annotation

Codecov / codecov/patch

model2vec/distill/inference.py#L75 

Added line #L75 was not covered by tests

mask = encodings["attention_mask"].cpu()
# NOTE: evil hack. For any batch, there will be a mask vector
Expand Down Expand Up @@ -124,6 +129,11 @@
with torch.no_grad():
encoded: BaseModelOutputWithPoolingAndCrossAttentions = model(input_ids=batch.to(device))
out: torch.Tensor = encoded.last_hidden_state
# NOTE: If the dtype is bfloat 16, we convert to float32,
# because numpy does not suport bfloat16
# See here: https://github.com/numpy/numpy/issues/19808
if out.dtype == torch.bfloat16:
out = out.float()

Check warning on line 136 in model2vec/distill/inference.py

View check run for this annotation

Codecov / codecov/patch

model2vec/distill/inference.py#L136 

Added line #L136 was not covered by tests
intermediate_weights.append(out[:, 1].cpu().numpy())
out_weights = np.concatenate(intermediate_weights)

Expand Down