transformers

transformers copied to clipboard

Add quantization_config in AutoModelForCausalLM.from_config()

10

### Feature request Add quantization_config feature to AutoModelForCausalLM from config . I am trying to pretrain a model from scratch and use bits and bytes so that It can be...

ishaansharma

# What does this PR do? This PR does the following: 1. Add a new field to `TrainerArguments.AcceleratorConfig`, `use_stateful_dataloader` which when set to true passes through `use_stateful_dataloader` to the DataLoaderConfiguration...

byi8220

ASR pipeline long-form audio processing requires `return_timestaps=True`

4

### System Info - `transformers` version: 4.45.2 - Platform: Linux-5.17.15-051715-generic-x86_64-with-glibc2.35 - Python version: 3.10.12 - Huggingface_hub version: 0.24.3 - Safetensors version: 0.4.3 - Accelerate version: 0.33.0 - Accelerate config: not...

as-suvorov

# What does this PR do?

gante

apply_chat_template return_assistant_tokens_mask not work for Qwen2.5

2

### System Info huggingface-hub-0.25.2 tokenizers-0.20.1 transformers-4.45.2 ### Who can help? @ArthurZucker @itazap ### Information - [ ] The official example scripts - [X] My own modified scripts ### Tasks -...

DogeWatch

Load a pretrainedfast tokenizer if fast=true and tokenizer.json exists

9

Current status for AutoTokenizer with fast=True: 1. checks tokenizer_config.json if tokenizer_class name ends with Fast 2. if not, load a slow tokenizer (This PR):
(unchanged) 1. checks tokenizer_config.json if tokenizer_class ...

itazap

GGUF support for BERT architecture

1

### Feature request I want to add the ability to use GGUF BERT models in transformers. Currently the library does not support this architecture. When I try to load it,...

Dimmension

Sync token-classification pipeline with Hub spec

1

One unusual change this time: The TokenClassificationPipeline put the docstrings in an odd place, with a decorator. As far as I can tell, this isn't necessary at all - I...

Rocketknight1

The chat template docs describe using generation prompts, but I realize it's a little unclear for model authors after reading #34209. This PR adds a section on writing generation prompts.

Rocketknight1

exclude fsdp from delay_optimizer_creation

5

# What does this PR do? It passes the model and the optimizer to accelerate.prepare in order to enable fp8 mixed precision, if any. Fixes #34024 ## Who can review?...

eljandoubi