Huggingface models

Text Generation Model

cat ./models/hf-text-gen.yaml

apiVersion: mlops.seldon.io/v1alpha1
kind: Model
metadata:
  name: text-gen
spec:
  storageUri: "gs://seldon-models/mlserver/huggingface/text-generation"
  requirements:
  - huggingface

Load the model

kubectl

kubectl apply -f ./models/hf-text-gen.yaml

model.mlops.seldon.io/text-gen created

seldon-cli

Wait for the model to be ready

kubectl

kubectl get model text-gen -n ${NAMESPACE} -o json | jq -r '.status.conditions[] | select(.message == "ModelAvailable") | .status'

True

seldon-cli

Do a REST inference call

kubectl

curl --location 'http://${MESH_IP}:9000/v2/models/text-gen/infer' \
	--header 'Content-Type: application/json'  \
    --data '{"inputs": [{"name": "args","shape": [1],"datatype": "BYTES","data": ["Once upon a time in a galaxy far away"]}]}'

{
	"model_name": "text-gen_1",
	"model_version": "1",
	"id": "121ff5f4-1d4a-46d0-9a5e-4cd3b11040df",
	"parameters": {},
	"outputs": [
		{
			"name": "output",
			"shape": [
				1,
				1
			],
			"datatype": "BYTES",
			"parameters": {
				"content_type": "hg_jsonlist"
			},
			"data": [
				"{\"generated_text\": \"Once upon a time in a galaxy far away, the planet is full of strange little creatures. A very strange combination of creatures in that universe, that is. A strange combination of creatures in that universe, that is. A kind of creature that is\"}"
			]
		}
	]
}

seldon-cli

res = !seldon model infer text-gen --inference-mode grpc \
   '{"inputs":[{"name":"args","contents":{"bytes_contents":["T25jZSB1cG9uIGEgdGltZSBpbiBhIGdhbGF4eSBmYXIgYXdheQo="]},"datatype":"BYTES","shape":[1]}]}'

import json
import base64
r = json.loads(res[0])
base64.b64decode(r["outputs"][0]["contents"]["bytesContents"][0])

b'{"generated_text": "Once upon a time in a galaxy far away\\n\\nThe Universe is a big and massive place. How can you feel any of this? Your body doesn\'t make sense if the Universe is in full swing \\u2014 you don\'t have to remember whether the"}'

Unload the model

kubectl

kubectl delete model text-gen

seldon-cli

Custom Text Generation Model

cat ./models/hf-text-gen-custom-tiny-stories.yaml

apiVersion: mlops.seldon.io/v1alpha1
kind: Model
metadata:
  name: custom-tiny-stories-text-gen
spec:
  storageUri: "gs://seldon-models/scv2/samples/mlserver_1.3.5/huggingface-text-gen-custom-tiny-stories"
  requirements:
    - huggingface

Load the model

kubectl

kubectl apply -f ./models/hf-text-gen-custom-tiny-stories.yaml

model.mlops.seldon.io/custom-tiny-stories-text-gen created

kubectl get model custom-tiny-stories-text-gen -n ${NAMESPACE} -o json | jq -r '.status.conditions[] | select(.message == "ModelAvailable") | .status'

True

curl --location 'http://${MESH_IP}:9000/v2/models/custom-tiny-stories-text-gen/infer' \
	--header 'Content-Type: application/json'  \
    --data '{"inputs": [{"name": "args","shape": [1],"datatype": "BYTES","data": ["Once upon a time in a galaxy far away"]}]}'

{
	"model_name": "custom-tiny-stories-text-gen_1",
	"model_version": "1",
	"id": "d0fce59c-76e2-4f81-9711-1c93d08bcbf9",
	"parameters": {},
	"outputs": [
		{
			"name": "output",
			"shape": [
				1,
				1
			],
			"datatype": "BYTES",
			"parameters": {
				"content_type": "hg_jsonlist"
			},
			"data": [
				"{\"generated_text\": \"Once upon a time in a galaxy far away. It was a very special place to live.\\n\"}"
			]
		}
	]
}

seldon-cli

res = !seldon model infer custom-tiny-stories-text-gen --inference-mode grpc \
   '{"inputs":[{"name":"args","contents":{"bytes_contents":["T25jZSB1cG9uIGEgdGltZSBpbiBhIGdhbGF4eSBmYXIgYXdheQo="]},"datatype":"BYTES","shape":[1]}]}'

import json
import base64
r = json.loads(res[0])
base64.b64decode(r["outputs"][0]["contents"]["bytesContents"][0])

b'{"generated_text": "Once upon a time in a galaxy far away\\nOne night, a little girl named Lily went to"}'

Unload the model

kubectl

kubectl delete custom-tiny-stories-text-gen

seldon-cli

As a next step, why not try running a larger-scale model? You can find a definition for one in ./models/hf-text-gen-custom-gpt2.yaml. However, you may need to request and allocate more memory!