Skip to main content

Model Recipes

Model recipes are organized in three levels:

  1. Choose a task using the Hugging Face task taxonomy below.
  2. Open the task to see every model family with a declared recipe.
  3. Open a family to see its exact manifest recipes, applicable trtmc CLI commands, Hugging Face model_type and architecture or pipeline class, TRTMC task/head contract, and every family-owned --set configuration key.

Every recipe row is generated from a model-owned manifest under tests/e2e/models/<family>/manifests/. A manifest declares an executable test contract; it is not, by itself, a current hardware pass receipt. Family config tables are generated from registered Python or C++ config schemas rather than maintained by hand. Use Supported Models for the retained release-support snapshot.

Multimodal

Hugging Face taskModel familiesDeclared recipesWhat it covers
Any-to-Any11Models that accept or produce more than one modality through one task contract.
Image-Text-to-Text614Vision-language models that answer or generate text from an image and prompt.

Natural Language Processing

Hugging Face taskModel familiesDeclared recipesWhat it covers
Text Generation3488Decoder, encoder-decoder, recurrent, and diffusion-style text generation.
Translation34Text generation recipes whose declared user contract is translation.
Feature Extraction1335Encoder and embedding models that return vector representations.
Text Ranking12Models that score documents against a query.

Computer Vision

Hugging Face taskModel familiesDeclared recipesWhat it covers
Image Classification12Models that assign a class to an input image.
Image Segmentation12Models that produce segmentation masks for an image.
Mask Generation23Prompted segmentation models that generate masks from points or text.
Text-to-Image420Diffusion recipes that generate an image from text.
Image-to-Image11Image-conditioned generation and editing recipes.
Text-to-Video48Diffusion recipes that generate video frames from text.
Image-to-Video11Image-conditioned video and world-model generation recipes.

Audio

Hugging Face taskModel familiesDeclared recipesWhat it covers
Automatic Speech Recognition39Speech-to-text models, including offline and streaming contracts.
Text-to-Speech27Models that synthesize audio from text.
Audio-to-Audio13Speech-to-speech models that consume and generate audio.

Time Series

Hugging Face taskModel familiesDeclared recipesWhat it covers
Time Series Forecasting49Forecasting and neural-operator models that consume numerical sequences.