clika_runtime.modelverse.auto

Load a model by name, the way you would from a model hub.

:class:AutoModel picks the load door from the family's declared modalities; the task loaders (:class:AutoModelForCausalLM, :class:AutoModelForImageTextToText, :class:AutoModelForSpeechSeq2Seq, :class:AutoModelForTextToWaveform, :class:AutoModelForSeq2SeqLM, :class:AutoModelForObjectDetection, :class:AutoModelForZeroShotObjectDetection, :class:AutoModelForDepthEstimation) each name their door; :class:AutoTokenizer loads a checkpoint's tokenizer without its weights; :class:AutoProcessor loads its preprocessing (the tokenizer and the media processor its directory describes) without its weights; :class:AutoConfig reads a checkpoint's identity, a :class:PretrainedConfig, without its weights::

import clika_runtime.modelverse as mv

model = mv.AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-0.5B-Instruct", device="cuda:0", dtype="bfloat16")
tokenizer = mv.AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")

Name	Kind
AutoConfig	class
AutoModel	class
AutoModelForCausalLM	class
AutoModelForDepthEstimation	class
AutoModelForImageTextToText	class
AutoModelForObjectDetection	class
AutoModelForSeq2SeqLM	class
AutoModelForSpeechSeq2Seq	class
AutoModelForTextToWaveform	class
AutoModelForZeroShotObjectDetection	class
AutoProcessor	class
AutoTokenizer	class
PretrainedConfig	class
functions	module functions