Update handler.py
Browse files- handler.py +55 -18
handler.py
CHANGED
@@ -1,23 +1,60 @@
|
|
1 |
-
|
2 |
-
|
|
|
3 |
import torch
|
|
|
4 |
|
5 |
-
|
|
|
6 |
|
7 |
-
|
8 |
-
|
9 |
-
|
|
|
|
|
|
|
|
|
10 |
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
-
|
23 |
-
app.run()
|
|
|
1 |
+
import json
|
2 |
+
import os
|
3 |
+
from typing import Dict, List, Any
|
4 |
import torch
|
5 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
6 |
|
7 |
+
MAX_TOKENS=8192
|
8 |
+
GPU_LAYERS=99 if torch.cuda.is_available() else 0
|
9 |
|
10 |
+
class EndpointHandler():
|
11 |
+
def __init__(self, data):
|
12 |
+
cfg = {
|
13 |
+
"repo": "MrOvkill/Phi-3-Instruct-Bloated",
|
14 |
+
}
|
15 |
+
self.model = AutoModelForCausalLM.from_pretrained(cfg['repo'])
|
16 |
+
self.tokenizer = AutoTokenizer.from_pretrained(cfg['repo'])
|
17 |
|
18 |
+
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
19 |
+
inputs = data.pop("inputs", "")
|
20 |
+
temperature = data.pop("temperature", None)
|
21 |
+
if not temperature:
|
22 |
+
temperature = data.pop("temp", 0.33)
|
23 |
+
if temperature > 3 or temperature < 0:
|
24 |
+
return json.dumps({
|
25 |
+
"status": "error",
|
26 |
+
"reason": "invalid temperature ( 0.01 - 1.00 )"
|
27 |
+
})
|
28 |
+
top_p = data.pop("top-p", 0.85)
|
29 |
+
if top_p > 3 or top_p < 0:
|
30 |
+
return json.dumps({
|
31 |
+
"status": "error",
|
32 |
+
"reason": "invalid top percentage ( 0.01 - 1.00 )"
|
33 |
+
})
|
34 |
+
top_k = data.pop("top-k", 42)
|
35 |
+
if top_k > 100 or top_k < 0:
|
36 |
+
return json.dumps({
|
37 |
+
"status": "error",
|
38 |
+
"reason": "invalid top k ( 1 - 99 )"
|
39 |
+
})
|
40 |
+
system_prompt = data.pop("system-prompt", "You are a helpful assistant.")
|
41 |
+
fmat = data.pop("format", f"<|system|>\n{system_prompt} <|end|>\n<|user|>\n{inputs} <|end|>\n<|assistant|>")
|
42 |
+
try:
|
43 |
+
fmat = fmat.format(system_prompt = system_prompt, prompt = inputs)
|
44 |
+
except Exception as e:
|
45 |
+
return json.dumps({
|
46 |
+
"status": "error",
|
47 |
+
"reason": "invalid format"
|
48 |
+
})
|
49 |
+
max_length = data.pop("max_length", 1024)
|
50 |
+
try:
|
51 |
+
max_length = int(max_length)
|
52 |
+
except Exception as e:
|
53 |
+
return json.dumps({
|
54 |
+
"status": "error",
|
55 |
+
"reason": "max_length was passed as something that was absolutely not a plain old int"
|
56 |
+
})
|
57 |
+
|
58 |
+
res = self.model(fmat, temperature=temperature, top_p=top_p, top_k=top_k, max_tokens=max_length)
|
59 |
|
60 |
+
return res
|
|