added custom handler for sharded loading

README.md

@@ -1 +1,9 @@
-# Shareded fp16 copy of [EleutherAI/gpt-j-6B](https://huggingface.co/EleutherAI/gpt-j-6B)
+---
+tags:
+- endpoints-template
+library_name: generic
+---
+
+# Shareded fp16 copy of [EleutherAI/gpt-j-6B](https://huggingface.co/EleutherAI/gpt-j-6B)
+
+> This is fork of [EleutherAI/gpt-j-6B](https://huggingface.co/EleutherAI/gpt-j-6B) with shareded fp16 weights implementing a custom `handler.py` as an example for how to use `gpt-j` [inference-endpoints](https://hf.co/inference-endpoints)

handler.py

@@ -0,0 +1,27 @@
+import torch
+from typing import Dict, List, Any
+from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
+
+# check for GPU
+device = 0 if torch.cuda.is_available() else -1
+
+
+class EndpointHandler:
+    def __init__(self, path=""):
+        # load the model
+        tokenizer = AutoTokenizer.from_pretrained(path)
+        model = AutoModelForCausalLM.from_pretrained(path, low_cpu_mem_usage=True)
+        # create inference pipeline
+        self.pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer, device=device)
+
+    def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
+        inputs = data.pop("inputs", data)
+        parameters = data.pop("parameters", None)
+
+        # pass inputs with all kwargs in data
+        if parameters is not None:
+            prediction = self.pipeline(inputs, **parameters)
+        else:
+            prediction = self.pipeline(inputs)
+        # postprocess the prediction
+        return prediction

requirements.txt

@@ -0,0 +1 @@
+accelerate