sgugger
Follow
Popular repositories
3,197 contributions in the last year
Activity overview
Contributed to
huggingface/transformers,
huggingface/accelerate,
huggingface/doc-builder
and 18 other
repositories
Contribution activity
July 2022
Created 61 commits in 4 repositories
Created a pull request in huggingface/transformers that received 22 comments
Custom pipeline
What does this PR do? This PR adds the ability to support custom pipelines on the Hub and share it with everyone else. Like the code in the Hub fea…
+433
−51
•
22
comments
Opened 26 other pull requests in 5 repositories
huggingface/transformers
15
merged
1
closed
- Replace false parameter by a buffer
- Fix OwlViT tests
- Make errors for loss-less models more user-friendly
- Fix TrainingArguments help section
- Add vision example to README
- Better default for offload_state_dict in from_pretrained
- Fix template for new models in README
- Make sharded checkpoints work in offline mode
- Make Nezha tests slow
- Fix image segmentation and object detection pipeline tests
- Add filename to info displayed when downloading things in from_pretrained
- Fix slow CI by pinning resampy
- Update localized READMES when template is filled.
- Place inputs on device when include_inputs_for_metrics is True
- Doc to dataset
- Sort doc toc
huggingface/accelerate
6
merged
huggingface/doc-builder
2
merged
huggingface/diffusers
1
merged
NielsRogge/transformers
1
merged
Reviewed 159 pull requests in 5 repositories
huggingface/transformers
25 pull requests
- Ignore small batches of examples for CLM training
- Mixed 8bit quantization
- Add Spanish translation of custom_models.mdx
- Troubleshooting.mdx Translation Italian
- Add Italian translation of sharing_custom_models.mdx
- Fix Sylvain's nits on the original KerasMetricCallback PR
- Raise a TF-specific error when importing Torch classes
- Add Italian translation of converting_tensorflow_models.mdx
- [ create_a_model.mdx ] translate to pt
- [XLA] Improve t5 model performance
- Update translation.mdx
- [EncoderDecoder] Improve docs
-
Allows
KerasMetricCallbackto use XLA generation - Add TFAutoModelForImageClassification to pipelines.py
- Generalize decay_mask_fn to apply mask to all LayerNorm params
-
Generate: validate
model_kwargs(and catch typos in generate arguments) - Fix dtype of input_features in docstring
-
Change how
take_along_axisis computed in DeBERTa to stop confusing XLA - Add Italian translation of create_model.mdx and serialization.mdx
-
Skip passes report for
--make-reports - Generate: validate arguments
-
Fix
no_trainerCI - Allow loading pretrained shared Pytorch checkpoints into flax models
- Update all no_trainer scripts
- Italian/accelerate
- Some pull request reviews not shown.
huggingface/accelerate
25 pull requests
- Complete revamp of the docs
- Rename actions to be a bit more accurate
- Fix clean
- set default num_cpu_threads_per_process to improve oob performance
- enhancements and fixes for FSDP and DeepSpeed
- ccl version check and import different module according to version
- Add a tqdm helper
- Update examples with grad accum
- fix: saving model weights
- Update new torch_ccl name in imports.py
- Speed up CI more
-
Create good defaults in
accelerate launch - Fix a few minor issues with example code in docs
- Create value errror in cli when accelerate config hasn't been run
- add on_main_process decorators
-
deepspeed version
0.6.7fix - Rename test extras to testing
- Add production testing + fix failing CI
- Allow for kwargs to be passed to trackers
- Add a gather_for_metrics capability
- Add 8 bit support - chapter II
- Add support for downcasting bf16 on TPUs
- Add balanced option for auto device map creation
- Add more documentation for device maps computations
- Restyle prepare_one
- Some pull request reviews not shown.
huggingface/blog
2 pull requests
huggingface/doc-builder
2 pull requests
huggingface/notebooks
1 pull request
Created an issue in bmcfee/resampy that received 19 comments
Resampy 0.3.1 slows down other libraries
Hi there! We just debugged a problem on the Transformers CI that was suddenly taking way longer than before (by a x20 factor) and it all came down …
19
comments