Skip to content

v4.36: Mixtral, Llava/BakLlava, SeamlessM4T v2, AMD ROCm, F.sdpa wide-spread support

Compare
Choose a tag to compare
@LysandreJik LysandreJik released this 11 Dec 12:12
· 2942 commits to main since this release

New model additions

Mixtral

Mixtral is the new open-source model from Mistral AI announced by the blogpost Mixtral of Experts. The model has been proven to have comparable capabilities to Chat-GPT according to the benchmark results shared on the release blogpost.

The architecture is a sparse Mixture of Experts with Top-2 routing strategy, similar as NllbMoe architecture in transformers. You can use it through AutoModelForCausalLM interface:

>>> import torch
>>> from transformers import AutoModelForCausalLM, AutoTokenizer

>>> model = AutoModelForCausalLM.from_pretrained("mistralai/Mixtral-8x7B", torch_dtype=torch.float16, device_map="auto")
>>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-8x7B")

>>> prompt = "My favourite condiment is"

>>> model_inputs = tokenizer([prompt], return_tensors="pt").to(device)
>>> model.to(device)

>>> generated_ids = model.generate(**model_inputs, max_new_tokens=100, do_sample=True)
>>> tokenizer.batch_decode(generated_ids)[0]

The model is compatible with existing optimisation tools such Flash Attention 2, bitsandbytes and PEFT library. The checkpoints are release under mistralai organisation on the Hugging Face Hub.

Llava / BakLlava

Llava is an open-source chatbot trained by fine-tuning LlamA/Vicuna on GPT-generated multimodal instruction-following data. It is an auto-regressive language model, based on the transformer architecture. In other words, it is an multi-modal version of LLMs fine-tuned for chat / instructions.

The Llava model was proposed in Improved Baselines with Visual Instruction Tuning by Haotian Liu, Chunyuan Li, Yuheng Li and Yong Jae Lee.

The integration also includes BakLlava which is a Llava model trained with Mistral backbone.

The mode is compatible with "image-to-text" pipeline:

from transformers import pipeline
from PIL import Image    
import requests

model_id = "llava-hf/llava-1.5-7b-hf"
pipe = pipeline("image-to-text", model=model_id)
url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/ai2d-demo.jpg"

image = Image.open(requests.get(url, stream=True).raw)
prompt = "USER: <image>\nWhat does the label 15 represent? (1) lava (2) core (3) tunnel (4) ash cloud\nASSISTANT:"

outputs = pipe(image, prompt=prompt, generate_kwargs={"max_new_tokens": 200})
print(outputs)

And you can find all Llava weights under llava-hf organisation on the Hub.

SeamlessM4T v2

SeamlessM4T-v2 is a collection of models designed to provide high quality translation, allowing people from different linguistic communities to communicate effortlessly through speech and text. It is an improvement on the previous version and was proposed in Seamless: Multilingual Expressive and Streaming Speech Translation by the Seamless Communication team from Meta AI.

For more details on the differences between v1 and v2, refer to section Difference with SeamlessM4T-v1.

SeamlessM4T enables multiple tasks without relying on separate models:

  • Speech-to-speech translation (S2ST)
  • Speech-to-text translation (S2TT)
  • Text-to-speech translation (T2ST)
  • Text-to-text translation (T2TT)
  • Automatic speech recognition (ASR)

PatchTST

The PatchTST model was proposed in A Time Series is Worth 64 Words: Long-term Forecasting with Transformers by Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong and Jayant Kalagnanam.

At a high level, the model vectorizes time series into patches of a given size and encodes the resulting sequence of vectors via a Transformer that then outputs the prediction length forecast via an appropriate head. The model is illustrated in the following figure:

patchtst

PatchTSMixer

The PatchTSMixer model was proposed in TSMixer: Lightweight MLP-Mixer Model for Multivariate Time Series Forecasting by Vijay Ekambaram, Arindam Jati, Nam Nguyen, Phanwadee Sinthong and Jayant Kalagnanam.

PatchTSMixer is a lightweight time-series modeling approach based on the MLP-Mixer architecture. In this HuggingFace implementation, we provide PatchTSMixer’s capabilities to effortlessly facilitate lightweight mixing across patches, channels, and hidden features for effective multivariate time-series modeling. It also supports various attention mechanisms starting from simple gated attention to more complex self-attention blocks that can be customized accordingly. The model can be pretrained and subsequently used for various downstream tasks such as forecasting, classification and regression.

CLVP

The CLVP (Contrastive Language-Voice Pretrained Transformer) model was proposed in Better speech synthesis through scaling by James Betker.

Phi-1/1.5

The Phi-1 model was proposed in Textbooks Are All You Need by Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tauman Kalai, Yin Tat Lee and Yuanzhi Li.

The Phi-1.5 model was proposed in Textbooks Are All You Need II: phi-1.5 technical report by Yuanzhi Li, Sébastien Bubeck, Ronen Eldan, Allie Del Giorno, Suriya Gunasekar and Yin Tat Lee.

TVP

The text-visual prompting (TVP) framework was proposed in the paper Text-Visual Prompting for Efficient 2D Temporal Video Grounding by Yimeng Zhang, Xin Chen, Jinghan Jia, Sijia Liu, Ke Ding.

This research addresses temporal video grounding (TVG), which is the process of pinpointing the start and end times of specific events in a long video, as described by a text sentence. Text-visual prompting (TVP), is proposed to enhance TVG. TVP involves integrating specially designed patterns, known as ‘prompts’, into both the visual (image-based) and textual (word-based) input components of a TVG model. These prompts provide additional spatial-temporal context, improving the model’s ability to accurately determine event timings in the video. The approach employs 2D visual inputs in place of 3D ones. Although 3D inputs offer more spatial-temporal detail, they are also more time-consuming to process. The use of 2D inputs with the prompting method aims to provide similar levels of context and accuracy more efficiently.

DINOv2 depth estimation

Depth estimation is added to the DINO v2 implementation.

ROCm support for AMD GPUs

AMD's ROCm GPU architecture is now supported across the board and fully tested in our CI with MI210/MI250 GPUs. We further enable specific hardware acceleration for ROCm in Transformers, such as Flash Attention 2, GPTQ quantization and DeepSpeed.

PyTorch scaled_dot_product_attention native support

PyTorch's torch.nn.functional.scaled_dot_product_attention operator is now supported in the most-used Transformers models and used by default when using torch>=2.1.1, allowing to dispatch on memory-efficient attention and Flash Attention backend implementations with no other package than torch required. This should significantly speed up attention computation on hardware that that supports these fastpath.

While Transformers automatically handles the dispatch to use SDPA when available, it is possible to force the usage of a given attention implementation ("eager" being the manual implementation, where each operation is implemented step by step):

# or `attn_implementation="sdpa", or `attn_implementation="flash_attention_2"`
model = AutoModelForSpeechSeq2Seq.from_pretrained("openai/whisper-tiny", attn_implementation="eager")

Training benchmark, run on A100-SXM4-80GB.

Model Batch size Sequence length Time per batch ("eager", s) Time per batch ("sdpa", s) Speedup Peak memory ("eager", MB) Peak memory ("sdpa", MB) Memory savings
llama2 7b 4 1024 1.065 0.90 19.4% 73878.28 45977.81 60.7%
llama2 7b 4 2048 OOM 1.87 / OOM 78394.58 SDPA does not OOM
llama2 7b 1 2048 0.64 0.48 32.0% 55557.01 29795.63 86.4%
llama2 7b 1 3072 OOM 0.75 / OOM 37916.08 SDPA does not OOM
llama2 7b 1 4096 OOM 1.03 / OOM 46028.14 SDPA does not OOM
llama2 7b 2 4096 OOM 2.05 / OOM 78428.14 SDPA does not OOM

Inference benchmark, run on A100-SXM4-80GB.

Model Batch size Prompt length Num new tokens Per token latency "eager" (ms) Per token latency "sdpa" (ms) Speedup
llama2 13b 1 1024 1 (prefill) 178.66 159.36 12.11%
llama2 13b 1 100 100 40.35 37.62 7.28%
llama2 13b 8 100 100 40.55 38.06 6.53%
Whisper v3 large 1 / 62 20.05 18.90 6.10%
Whisper v3 large 8 / 77 25.42 24.77 2.59%
Whisper v3 large 16 / 77 28.51 26.32 8.34%

New Cache abstraction & Attention Sinks support

We are rolling out a new abstraction for the past_key_values cache, which enables the use of different types of caches. For now, only llama and llama-inspired architectures (mistral, persimmon, phi) support it, with other architectures scheduled to have support in the next release. By default, a growing cache (DynamicCache) is used, which preserves the existing behavior.

This release also includes a new SinkCache cache, which implements the Attention Sinks paper. With SinkCache, the model is able to continue generating high-quality text well beyond its training sequence length! Note that it does not expand the context window, so it can’t digest very long inputs — it is suited for streaming applications such as multi-round dialogues. Check this colab for an example.

image

  • Generate: New Cache abstraction and Attention Sinks support by @tomaarsen in #26681
  • Generate: SinkCache can handle iterative prompts by @gante in #27907

Safetensors as a default

We continue toggling features enabling safetensors as a default across the board, in PyTorch, Flax, and TensorFlow.
When using PyTorch model and forcing the load of safetensors file with use_safetensors=True, if the repository does not contain the safetensors files, they will now be converted on-the-fly server-side.

Breaking changes

pickle files

We now disallow the use of pickle.load internally for security purposes. To circumvent this, you can use the TRUST_REMOTE_CODE=True command to indicate that you would still like to load it.

  • 🚨🚨🚨 Disallow pickle.load unless TRUST_REMOTE_CODE=True by @ydshieh in #27776

Beam score calculation for decoder-only models

In the previous implementation of beam search, when length_penalty is active, the beam score for decoder-only models was penalized by the total length of both prompt and generated sequence. However, the length of prompt should not be included in the penalization step -- this release fixes it.

  • 🚨🚨 Fix beam score calculation issue for decoder-only models by @VsonicV in #27351
  • Fix remaining issues in beam score calculation by @VsonicV in #27808
  • Fix beam score calculation issue for Tensorflow version by @VsonicV in #27814
  • Fix beam score calculation issue for JAX version by @VsonicV in #27816

Slight API changes/corrections

  • ⚠️ [VitDet] Fix test by @NielsRogge in #27832
  • [⚠️ removed a default argument] Make AttentionMaskConverter compatible with torch.compile(..., fullgraph=True) by @fxmarty in #27868

Bugfixes and improvements

Significant community contributions

The following contributors have made significant changes to the library over the last release:

  • @jiaqiw09
    • translate peft.md to chinese (#27215)
    • translate autoclass_tutorial to chinese (#27269)
    • translate run_scripts.md to chinese (#27246)
    • translate model_sharing.md and llm_tutorial.md to chinese (#27283)
    • translate big_models.md and performance.md to chinese (#27334)
    • translate debugging.md to chinese (#27374)
    • Perf torch compile (#27422)
    • translate hpo_train.md and perf_hardware.md to chinese (#27431)
    • translate Trainer.md to chinese (#27527)
    • translate deepspeed.md to chinese (#27495)
    • translation main-class files to chinese (#27588)
    • translate internal folder files to chinese (#27638)
  • @susnato
    • [FA2] Add flash attention for for DistilBert (#26489)
    • [FA2] Add flash attention for GPT-Neo (#26486)
    • Remove padding_masks from gpt_bigcode. (#27348)
    • Add CLVP (#24745)
    • Add Phi-1 and Phi-1_5 (#26170)
    • [FA2] Add flash attention for opt (#26414)
    • CLVP Fixes (#27547)
    • [FA-2] Add Flash Attention to Phi (#27661)
  • @jiqing-feng
    • add attention_mask and position_ids in assisted model (#26892)
    • TVP model (#25856)
    • fix assisted decoding assistant model inputs (#27503)
    • Fix bug of _prepare_4d_attention_mask (#27847)
  • @psinthong
    • [time series] Add PatchTST (#25927)
  • @Yuki-Imajuku
    • Translating en/model_doc docs to Japanese. (#27401)
  • @dg845
    • Add UnivNet Vocoder Model for Tortoise TTS Diffusers Integration (#24799)
  • @rajveer43
    • Translate en/model_doc to JP (#27264)
    • Translate en/tasks folder docs to Japanese 🇯🇵 (#27098)
    • Translating en/model_doc folder docs to Japanese(from blip to clap) 🇯🇵 (#27673)
    • Translate model_doc files from clip to cpm to JP (#27774)
  • @NoB0
    • [i18n-fr] Translate installation to French (#27657)
    • [i18n-fr] Translate autoclass tutorial to French (#27659)
  • @ajati
    • [Time series] Add PatchTSMixer (#26247)
  • @vvvm23
    • Add Llama Flax Implementation (#24587)