Serialization & Deserialization for Requests

Hugging Face Inference Endpount comes with a default serving container which is used for all supported Transformers and Sentence-Transformers tasks and for custom inference handler. The serving container takes care of serialization and deserialization of the request and response payloads based on the content-type and accept headers of the request. That means that when you send a request with a JSON body and a content-type: application/json header, the serving container will deserialize the JSON payload into a Python dictionary and pass it to the inference handler and if you send a request with a accept: image/png header, the serving container will serialize the response from the task/custom handler into a image.

Below is a list of supported content-types and the deserialized payload that is passed to the inference handler.

Content-Type	Payload
application/json	dict
text/csv	raw
text/plain	raw
image/png	binary
image/jpeg	binary
image/jpg	binary
image/tiff	binary
image/bmp	binary
image/gif	binary
image/webp	binary
image/x-image	binary
audio/x-flac	{"inputs": bytes(body)}
audio/flac	{"inputs": bytes(body)}
audio/mpeg	{"inputs": bytes(body)}
audio/x-mpeg-3	{"inputs": bytes(body)}
audio/wave	{"inputs": bytes(body)}
audio/wav	{"inputs": bytes(body)}
audio/x-wav	{"inputs": bytes(body)}
audio/ogg	{"inputs": bytes(body)}
audio/x-audio	{"inputs": bytes(body)}
audio/webm	{"inputs": bytes(body)}
audio/webm;codecs=opus	{"inputs": bytes(body)}
audio/AMR	{"inputs": bytes(body)}
audio/amr	{"inputs": bytes(body)}
audio/AMR-WB	{"inputs": bytes(body)}
audio/AMR-WB+	{"inputs": bytes(body)}
audio/m4a	{"inputs": bytes(body)}
audio/x-m4a	{"inputs": bytes(body)}

Below is a list of supported accept headers and the serialized payload is returned.

Accept	Payload
application/json	JSON
text/csv	raw
text/plain	raw
image/png	binary
image/jpeg	binary
image/jpg	binary
image/tiff	binary
image/bmp	binary
image/gif	binary
image/webp	binary
image/x-image	binary

< > Update on GitHub