DocsGPT/application/llm/sagemaker.py

140 lines
4.8 KiB
Python
Raw Normal View History

2023-09-29 00:09:01 +00:00
from application.llm.base import BaseLLM
from application.core.settings import settings
import json
2023-10-06 00:52:29 +00:00
import io
class LineIterator:
"""
2024-04-15 14:05:59 +00:00
A helper class for parsing the byte stream input.
2023-10-06 00:52:29 +00:00
The output of the model will be in the following format:
```
b'{"outputs": [" a"]}\n'
b'{"outputs": [" challenging"]}\n'
b'{"outputs": [" problem"]}\n'
...
```
2024-04-15 14:05:59 +00:00
While usually each PayloadPart event from the event stream will contain a byte array
2023-10-06 00:52:29 +00:00
with a full json, this is not guaranteed and some of the json objects may be split across
PayloadPart events. For example:
```
{'PayloadPart': {'Bytes': b'{"outputs": '}}
{'PayloadPart': {'Bytes': b'[" problem"]}\n'}}
```
2024-04-15 14:05:59 +00:00
2023-10-06 00:52:29 +00:00
This class accounts for this by concatenating bytes written via the 'write' function
and then exposing a method which will return lines (ending with a '\n' character) within
2024-04-15 14:05:59 +00:00
the buffer via the 'scan_lines' function. It maintains the position of the last read
position to ensure that previous bytes are not exposed again.
2023-10-06 00:52:29 +00:00
"""
2024-04-15 14:05:59 +00:00
2023-10-06 00:52:29 +00:00
def __init__(self, stream):
self.byte_iterator = iter(stream)
self.buffer = io.BytesIO()
self.read_pos = 0
def __iter__(self):
return self
def __next__(self):
while True:
self.buffer.seek(self.read_pos)
line = self.buffer.readline()
2024-04-15 14:05:59 +00:00
if line and line[-1] == ord("\n"):
2023-10-06 00:52:29 +00:00
self.read_pos += len(line)
return line[:-1]
try:
chunk = next(self.byte_iterator)
except StopIteration:
if self.read_pos < self.buffer.getbuffer().nbytes:
continue
raise
2024-04-15 14:05:59 +00:00
if "PayloadPart" not in chunk:
print("Unknown event type:" + chunk)
2023-10-06 00:52:29 +00:00
continue
self.buffer.seek(0, io.SEEK_END)
2024-04-15 14:05:59 +00:00
self.buffer.write(chunk["PayloadPart"]["Bytes"])
2023-09-29 00:09:01 +00:00
class SagemakerAPILLM(BaseLLM):
2024-04-15 17:02:24 +00:00
def __init__(self, api_key=None, *args, **kwargs):
2023-10-06 00:52:29 +00:00
import boto3
2024-04-15 14:05:59 +00:00
2023-10-06 00:52:29 +00:00
runtime = boto3.client(
2024-04-15 14:05:59 +00:00
"runtime.sagemaker",
aws_access_key_id="xxx",
aws_secret_access_key="xxx",
region_name="us-west-2",
2023-10-06 00:52:29 +00:00
)
super().__init__(*args, **kwargs)
self.api_key = api_key
2024-04-15 14:05:59 +00:00
self.endpoint = settings.SAGEMAKER_ENDPOINT
2023-10-06 00:52:29 +00:00
self.runtime = runtime
2024-04-15 17:02:24 +00:00
def _raw_gen(self, baseself, model, messages, stream=False, **kwargs):
2024-04-15 14:05:59 +00:00
context = messages[0]["content"]
user_question = messages[-1]["content"]
2023-09-29 00:09:01 +00:00
prompt = f"### Instruction \n {user_question} \n ### Context \n {context} \n ### Answer \n"
2023-10-06 00:52:29 +00:00
# Construct payload for endpoint
payload = {
"inputs": prompt,
"stream": False,
"parameters": {
"do_sample": True,
"temperature": 0.1,
"max_new_tokens": 30,
"repetition_penalty": 1.03,
2024-04-15 14:05:59 +00:00
"stop": ["</s>", "###"],
},
2023-10-06 00:52:29 +00:00
}
2024-04-15 14:05:59 +00:00
body_bytes = json.dumps(payload).encode("utf-8")
2023-09-29 00:09:01 +00:00
2023-10-06 00:52:29 +00:00
# Invoke the endpoint
2024-04-15 14:05:59 +00:00
response = self.runtime.invoke_endpoint(
EndpointName=self.endpoint, ContentType="application/json", Body=body_bytes
)
result = json.loads(response["Body"].read().decode())
2023-10-06 00:52:29 +00:00
import sys
2023-09-29 00:09:01 +00:00
2024-04-15 14:05:59 +00:00
print(result[0]["generated_text"], file=sys.stderr)
return result[0]["generated_text"][len(prompt) :]
2024-04-15 17:02:24 +00:00
def _raw_gen_stream(self, baseself, model, messages, stream=True, **kwargs):
2024-04-15 14:05:59 +00:00
context = messages[0]["content"]
user_question = messages[-1]["content"]
2023-10-06 00:52:29 +00:00
prompt = f"### Instruction \n {user_question} \n ### Context \n {context} \n ### Answer \n"
# Construct payload for endpoint
payload = {
"inputs": prompt,
"stream": True,
"parameters": {
"do_sample": True,
"temperature": 0.1,
"max_new_tokens": 512,
"repetition_penalty": 1.03,
2024-04-15 14:05:59 +00:00
"stop": ["</s>", "###"],
},
2023-10-06 00:52:29 +00:00
}
2024-04-15 14:05:59 +00:00
body_bytes = json.dumps(payload).encode("utf-8")
2023-10-06 00:52:29 +00:00
# Invoke the endpoint
2024-04-15 14:05:59 +00:00
response = self.runtime.invoke_endpoint_with_response_stream(
EndpointName=self.endpoint, ContentType="application/json", Body=body_bytes
)
# result = json.loads(response['Body'].read().decode())
event_stream = response["Body"]
start_json = b"{"
2023-10-06 00:52:29 +00:00
for line in LineIterator(event_stream):
2024-04-15 14:05:59 +00:00
if line != b"" and start_json in line:
# print(line)
data = json.loads(line[line.find(start_json) :].decode("utf-8"))
if data["token"]["text"] not in ["</s>", "###"]:
print(data["token"]["text"], end="")
yield data["token"]["text"]