Skip to content

[transformers_modeling_backend] Upgrade transformers from 4.57.1 to 5.0.0rc0 #2154

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

Open
3outeille wants to merge 14 commits into
base: main
Choose a base branch
from
Open

[transformers_modeling_backend] Upgrade transformers from 4.57.1 to 5.0.0rc0 #2154

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
4f36924
add tooling
3outeille Nov 24, 2025
5a63932
add check checkpoint correctness
3outeille Nov 24, 2025
819ea64
add compare throughput to tooling
3outeille Dec 15, 2025
a54016b
Enhance debug_local.sh to support compile flag
3outeille Dec 15, 2025
409483a
Enhance debug_local.sh to support dynamic flavor selection
3outeille Dec 15, 2025
b7e88ce
modify template to use reservation
3outeille Dec 15, 2025
33f7e57
this doesnt exist in v5 anymore (maybe for encoder tho but not for de…
3outeille Dec 15, 2025
f544b93
Fix padding index handling for DTensor in HFTransformerModel class
3outeille Dec 15, 2025
15e44c3
Enhance grid_search_slurm.sh to support dynamic flavor selection and …
3outeille Dec 15, 2025
f0eb064
Merge branch 'tooling' into move-to-v5
3outeille Dec 15, 2025
769a61a
upgrade to transformers v5
3outeille Dec 15, 2025
b6bad9a
avoid copy using _local_tensor()
3outeille Dec 15, 2025
da21ac8
remove tooling
3outeille Dec 15, 2025
ff0dbb9
linting
3outeille Dec 16, 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
2 changes: 1 addition & 1 deletion .ci/docker/requirements-transformers-modeling-backend.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1 +1 @@
transformers==4.57.1
transformers==5.0.0rc0
2 changes: 1 addition & 1 deletion torchtitan/experiments/transformers_modeling_backend/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@ This enables HF transformers models to be trained with `4D parallelism + torch.c

## Quick start

- Requirements `transformers==4.57.1`
- Requirements `transformers==5.0.0rc0`

- Config: `torchtitan/torchtitan/experiments/transformers_modeling_backend/configs/qwen3.toml`
```diff
Expand Down
10 changes: 5 additions & 5 deletions torchtitan/experiments/transformers_modeling_backend/model/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@

import torch
from torch import nn
from torch.distributed.tensor import DTensor
from torch.nn import init
from torchtitan.tools.logging import logger
from transformers.configuration_utils import PretrainedConfig
Expand Down Expand Up @@ -270,7 +271,10 @@ def _init_weights_patched(self, module):
module.weight.data.normal_(mean=0.0, std=std)

if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
if isinstance(module.weight.data, DTensor):
module.weight.data._local_tensor[module.padding_idx].zero_()
Copy link
Contributor

@tianyu-l tianyu-l Dec 16, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

sorry I probably didn't what you are doing here.
If the padding is on the "global tensor", we should just do the same thing module.weight.data[module.padding_idx].zero_()

The code here is doing local modification, which may or may not be correct depending on if the padding_idx is meant to be local or global.

else:
module.weight.data[module.padding_idx].zero_()

elif (
isinstance(
Expand Down Expand Up @@ -429,10 +433,6 @@ def init_weights(self, *args, **kwargs):
# This method replicates the behavior of the original PreTrainedModel.init_weights,
# but with a custom weight initialization function that skips nn.Identity modules (when PP is enabled)

if self.model.config.pruned_heads:
logger.info("Pruning heads as per model configuration.")
self.model.prune_heads(self.model.config.pruned_heads)

original_init_weights_fn = self.model._init_weights

def selective_init(module):
Expand Down
Loading