Skip to content

Fix the KeyError when loading bloom-based models #441

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

Merged
merged 6 commits into from
Jul 14, 2023
Merged

Fix the KeyError when loading bloom-based models #441

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
901f74a
fix: KeyError when loading bloom-based models
HermitSun Jul 12, 2023
8b40de8
fix: use lm_head_weight in existing weights
HermitSun Jul 12, 2023
fedda9a
refactor: remove unnecessary state_dict
HermitSun Jul 12, 2023
2f779f6
refactor: load weight in parallel instead
HermitSun Jul 12, 2023
f9843e8
fix: use provided lm_head weight
HermitSun Jul 12, 2023
5c4d345
style: remove trailing whitespace in comments
HermitSun Jul 14, 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
13 changes: 10 additions & 3 deletions vllm/model_executor/models/bloom.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -284,10 +284,17 @@ def load_weights(self,
state_dict = self.state_dict()
for name, loaded_weight in hf_model_weights_iterator(
model_name_or_path, cache_dir, use_np_cache):
if not name.startswith("transformer."):
name = "transformer." + name
if name == "lm_head.weight":
# Since hidden_states are parallelized, we need to
# load lm_head.weight in parallel.
self._column_parallel_weights.append(name)
# If lm_head is provided, use it instead.
param = self.lm_head_weight
else:
if not name.startswith("transformer."):
name = "transformer." + name
param = state_dict[name]

param = state_dict[name]
if "query_key_value" in name:
# NOTE(woosuk): BLOOM's fused QKV has the shape of
# [num_heads * 3 * head_size, hidden_size], while the
Expand Down