Skip to content

Transformers backend tweaks #17365

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
vllm-bot merged 2 commits into from
Apr 29, 2025
Merged

Transformers backend tweaks #17365

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a11719a
Remove `meta_to_empty` as it should no longer be needed
hmellor Apr 29, 2025
425cbaf
Warn if buffer is a direct child of the model
hmellor Apr 29, 2025
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
22 changes: 9 additions & 13 deletions vllm/model_executor/models/transformers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,6 @@
# limitations under the License.
"""Wrapper around `transformers` models"""
import re
from itertools import chain
from typing import Iterable, Literal, Optional, Union

import torch
Expand Down Expand Up @@ -166,12 +165,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
# Initialize buffers (e.g. rotary embedding inverse frequency)
self.init_buffers(self.model)

# Initialize parameters
# Initialize any parameters that have not had their modules replaced
self.init_parameters(self.model)

# Move remaining meta tensors to device (should happen last)
self.meta_to_empty(self.model)

self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(["hidden_states"],
config.hidden_size))
Expand Down Expand Up @@ -296,6 +292,14 @@ def init_buffers(self, module: nn.Module):
"""
for name, buffer in module.named_buffers(recurse=False):
if buffer.device == torch.device("meta"):
if module == self.model:
logger.warning(
"To initialize buffers correctly, we instantiate the "
"parent module and and extract the value of the "
"buffer from it. In this case, the parent module is "
"the base model. Instantiating the entire model here "
"risks GPU OOM. Could this buffer be moved to a child "
"module?")
new_buffer = getattr(type(module)(self.config), name)
setattr(module, name, new_buffer)
for child in module.children():
Expand All @@ -320,14 +324,6 @@ def init_parameters(self, module: nn.Module):
for child in module.children():
self.init_parameters(child)

def meta_to_empty(self, module: nn.Module):
tensors = list(chain(module.buffers(), module.parameters()))
if tensors and all(t.device == torch.device("meta") for t in tensors):
module.to_empty(device=self.device_config.device)
return # We can stop recursing because to_empty is recursive
for child in module.children():
self.meta_to_empty(child)

def get_input_embeddings(self) -> nn.Module:
return self.model.get_input_embeddings()

Expand Down