sgugger
Follow
Popular repositories
3,911 contributions in the last year
Activity overview
Contributed to
huggingface/transformers,
huggingface/accelerate,
huggingface/doc-builder
and 18 other
repositories
Contribution activity
March 2023
Created 37 commits in 2 repositories
Created a pull request in huggingface/transformers that received 7 comments
Add a progress bar for the total download of shards
What does this PR do? This PR adds a feature requested in #22047 and fixes a small bug I encountered while testing it. The feature: a new progress …
+9
−2
•
7
comments
Opened 19 other pull requests in 2 repositories
huggingface/transformers
1
open
12
merged
- Rework a bit the LLaMA conversion script
- LLaMA house-keeping
- Regression pipeline device
- Revert "Enforce same behavior as PyTorch 2.0 for older versions"
- Run all tests by default
- Load optimizer state on CPU to avoid CUDA OOM
- Remove backend check for torch.compile
- Fix doc link for MGP-STR
- Enforce same behavior as PyTorch 2.0 for older versions
- Fix big model inference for T5 models in float16
- Fix imports of TF MobileViT
- Fix test for torchneuroncore in Trainer
- Mark pipeline tests to skip them easily
huggingface/accelerate
6
merged
Reviewed 171 pull requests in 6 repositories
huggingface/transformers
25 pull requests
- Rework a bit the LLaMA conversion script
- Fix Unnecessary move of tensors from CPU to GPU in LlamaRotaryEmbedding
- [trainer] param count for deepspeed zero3
- Fix llama_tokenizer
- Add MaskedImageModelingOutput
- Fix natten
- fix AutoTP in deepspeed could not work for bloom
- [neptune] fix checkpoint bug with relative out_dir
- Hotfix for natten on CircleCI
-
🔥 py38 + torch 2🔥 🔥 🔥 🚀 - Temporarily fix ONNX model exporting error
- Modify electra loss calcualation part
- Update tiny model creation script
- t5 remove data dependency
- LLaMA Implementation
- to_pil - don't rescale if int and in range 0-255
-
[trainer] add
--optim adamw_torch_fusedfor pt-2.0+ - Prepare daily CI for torch 2.0.0
- Remove backend check for torch.compile
- [Safetensors] Add explicit flag to from pretrained
- Trainer: let generate pick its inputs
-
Move
is_pipeline_test_to_skipto specific model test classes - [trainer] fix bug in grad accum with multiple epochs
- Chunkable token classification pipeline
- Add a new script to check model testers' config
- Some pull request reviews not shown.
huggingface/accelerate
25 pull requests
- Fix typo in TPU config
- Better error message when using multi-GPU and Accelerate on torch <1.9.1
- Set drop last to ensure modulo16 restriction for fp8
- Make the Scheduler adjust the steps taken relative to the gradient accumulation steps
- Add documentation around FSDP state dict save behavior
-
add
use_orig_paramsto FullyShardedDataParallelPlugin - Run accelerate_test in cli
- Fix CPU error always being raised
-
[
Accelerator] We should not calltoon modules that wrapsaccelerateloaded models - Remove outdated command directions and use in tests
-
Fix what files get deleted through
total_limit - Document skip_first_batches in the checkpoint usage guides
-
Attempt to fix import error when PyTorch is build without
torch.distributedmodule -
[
Accelerator] Fix issue with 8bit models - Let GradientState know active dataloaders and reset the remainder
- Load custom state to cpu
- Include a note in the gradient synchronization docs on "what can go wrong" and show the timings
- [Safetensors] Relax missing metadata constraint
- update support for torch dynamo compile
- Fixed typos in notebook
- FSDP enhancements and fixes
- fix partial state
- Fix notebook_launcher
- Fp8 integration
- fix ds dist init kwargs issue
- Some pull request reviews not shown.
huggingface/safetensors
3 pull requests
huggingface/doc-builder
3 pull requests
huggingface/tokenizers
2 pull requests
huggingface/diffusers
1 pull request
Created an issue in pytorch/pytorch that received 4 comments
CUDA IMA on forward pass of BertForMaskedLM with PT2 RC
BertForMaskedLM with torch.compile, an illegal memory CUDA error appears. The sa…
4
comments