Support decode token logprobs (#130)

This commit is contained in:
Cody Yu
2024-02-06 12:24:55 -08:00
committed by GitHub
parent ee1df26a77
commit a7334aeea1
10 changed files with 233 additions and 96 deletions

View File

@@ -9,10 +9,24 @@ The capital of France is Paris.\nThe capital of the United States is Washington,
"""
import argparse
import time
import requests
def test_decode(url, return_logprob):
response = requests.post(
url + "/generate",
json={
"text": "The capital of France is",
"sampling_params": {
"temperature": 0,
"max_new_tokens": 32,
},
"return_logprob": return_logprob,
"logprob_start_len": 0,
},
)
print(response.json())
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--host", type=str, default="http://127.0.0.1")
@@ -21,16 +35,5 @@ if __name__ == "__main__":
url = f"{args.host}:{args.port}"
response = requests.post(
url + "/generate",
json={
"text": "The capital of France is",
"sampling_params": {
"temperature": 0,
"max_new_tokens": 32,
},
# "return_logprob": True,
# "logprob_start_len": 0,
},
)
print(response.json())
test_decode(url, False)
test_decode(url, True)