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

HF Model loading fixes #59

Merged
merged 3 commits into from
Jul 26, 2024
Merged

HF Model loading fixes #59

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
631273f
fix unexpected keys warning
farzadab Jul 26, 2024
47adebf
fix slow model loading due to random init
farzadab Jul 26, 2024
35f7c86
comments for unexpected and missing keys_to_ignore
farzadab Jul 26, 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
24 changes: 20 additions & 4 deletions ultravox/model/ultravox_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,14 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
config_class = UltravoxConfig
config: UltravoxConfig # for type hinting
_no_split_modules = ["Wav2Vec2Model", "WhisperEncoder", "LlamaDecoderLayer"]
# We minimize the weights in state_dict in order to reduce the size of the checkpoint
# The issue is that load_pretrained() uses state_dict() keys to know what keys are expected
# As such we have to tell is to ignore some keys that are not always in the model
_keys_to_ignore_on_load_unexpected = ["audio_tower.*", "language_model.*"]
# Usually we load encoder weights from a pretrained model, so we don't want to load the decoder weights
# Technically we never hit this issue because these keys are already removed from state_dict() however,
# but there's no harm in keeping it here for when we change that behavior.
_keys_to_ignore_on_load_missing = ["audio_tower.*"]

def __init__(self, config: UltravoxConfig):
super().__init__(config)
Expand Down Expand Up @@ -201,7 +209,12 @@ def _create_audio_tower(
if "whisper" in config.audio_config._name_or_path:
audio_tower = ModifiedWhisperEncoder(config.audio_config)
else:
audio_tower = transformers.AutoModel.from_config(config.audio_config)
with transformers.modeling_utils.no_init_weights():
# we only ever use from_config if the weights are retrained, hence initializing is not
# required. This makes the model quite creation faster since init on CPU is quite slow.
audio_tower = transformers.AutoModel.from_config(
config.audio_config
)

if isinstance(
audio_tower,
Expand All @@ -224,9 +237,12 @@ def _create_language_model(
config.text_model_id, attn_implementation=config._attn_implementation
)
else:
language_model = transformers.AutoModelForCausalLM.from_config(
config.text_config, attn_implementation=config._attn_implementation
)
with transformers.modeling_utils.no_init_weights():
# we only ever use from_config if the weights are retrained, hence initializing is not
# required. This makes the model quite creation faster since init on CPU is quite slow.
language_model = transformers.AutoModelForCausalLM.from_config(
config.text_config, attn_implementation=config._attn_implementation
)

language_model = apply_lora(language_model, config.text_model_lora_config)
return language_model
Expand Down
Loading