Spaces:
Runtime error
Runtime error
File size: 5,470 Bytes
e4815ea 3e937fb e4815ea 5fead4a e4815ea 3e937fb e4815ea 3e937fb e4815ea 3e937fb e4815ea 3e937fb e4815ea 3e937fb e4815ea 3e937fb e4815ea f011203 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 |
from pydantic import BaseModel
from llama_cpp import Llama
import os
import gradio as gr # Not suitable for production
from dotenv import load_dotenv
from fastapi import FastAPI, Request
from fastapi.responses import StreamingResponse
import spaces
import asyncio
import random
from llama_cpp.tokenizers import LlamaTokenizer
from peft import PeftModel, LoraConfig, get_peft_model
import torch
from multiprocessing import Process, Queue
from google.cloud import storage
import json
app = FastAPI()
load_dotenv()
HUGGINGFACE_TOKEN = os.getenv("HUGGINGFACE_TOKEN")
GOOGLE_CLOUD_BUCKET = os.getenv("GOOGLE_CLOUD_BUCKET")
GOOGLE_CLOUD_CREDENTIALS = os.getenv("GOOGLE_CLOUD_CREDENTIALS")
gcp_credentials = json.loads(GOOGLE_CLOUD_CREDENTIALS)
storage_client = storage.Client.from_service_account_info(gcp_credentials)
bucket = storage_client.bucket(GOOGLE_CLOUD_BUCKET)
class ModelManager:
def __init__(self):
self.params = {"n_ctx": 2048, "n_batch": 512, "n_predict": 512, "repeat_penalty": 1.1, "n_threads": 1, "seed": -1, "stop": ["</s>"], "tokens": []}
self.tokenizer = LlamaTokenizer.from_pretrained("decapoda-research/llama-7b-hf") #Load tokenizer from GCS for production
self.request_queue = Queue()
self.response_queue = Queue()
self.model = self.load_model_from_bucket("llama-2-7b-chat/llama-2-7b-chat.Q4_K_M.gguf")
self.start_processing_processes()
def load_model_from_bucket(self, bucket_path):
blob = bucket.blob(bucket_path)
try:
model = Llama(model_path=blob.download_as_string(), **self.params)
return model
except Exception as e:
print(f"Error loading model: {e}")
return None
def save_model_to_bucket(self, model, bucket_path):
blob = bucket.blob(bucket_path)
try:
blob.upload_from_string(model.save_pretrained(), content_type='application/octet-stream')
except Exception as e:
print(f"Error saving model: {e}")
def train_model(self): #This function needs a complete overhaul for production use. This is a placeholder.
config = LoraConfig(r=8, lora_alpha=32, target_modules=["q_proj", "v_proj"], lora_dropout=0.05, bias="none", task_type="CAUSAL_LM")
base_model_path = "llama-2-7b-chat/llama-2-7b-chat.Q4_K_M.gguf"
try:
base_model = self.load_model_from_bucket(base_model_path)
if base_model:
model = get_peft_model(base_model, config)
# Placeholder training data - needs a robust data loading mechanism
for batch in [{"question": ["a"], "answer":["b"]}, {"question":["c"], "answer":["d"]}]:
inputs = self.tokenizer(batch["question"], return_tensors="pt", padding=True, truncation=True)
labels = self.tokenizer(batch["answer"], return_tensors="pt", padding=True, truncation=True)
outputs = model(**inputs, labels=labels.input_ids)
loss = outputs.loss
loss.backward()
self.save_model_to_bucket(model, "llama_finetuned/llama_finetuned.gguf")
del model
del base_model
except Exception as e:
print(f"Error during training: {e}")
def generate_text(self, prompt):
if self.model:
inputs = self.tokenizer(prompt, return_tensors="pt")
outputs = self.model.generate(**inputs, max_new_tokens=100)
generated_text = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
return generated_text
else:
return "Error loading model."
def start_processing_processes(self):
p = Process(target=self.process_requests)
p.start()
def process_requests(self):
while True:
request_data = self.request_queue.get()
if request_data is None:
break
inputs, top_p, top_k, temperature, max_tokens = request_data
try:
response = self.generate_text(inputs)
self.response_queue.put(response)
except Exception as e:
print(f"Error during inference: {e}")
self.response_queue.put("Error generating text.")
model_manager = ModelManager()
class ChatRequest(BaseModel):
message: str
@spaces.GPU()
async def generate_streaming_response(inputs):
top_p = 0.9
top_k = 50
temperature = 0.7
max_tokens = model_manager.params["n_ctx"] - len(model_manager.tokenizer.encode(inputs))
model_manager.request_queue.put((inputs, top_p, top_k, temperature, max_tokens))
full_text = model_manager.response_queue.get()
async def stream_response():
yield full_text
return StreamingResponse(stream_response())
async def process_message(message):
inputs = message.strip()
return await generate_streaming_response(inputs)
@app.post("/generate_multimodel")
async def api_generate_multimodel(request: Request):
data = await request.json()
message = data["message"]
return await process_message(message)
iface = gr.Interface(fn=process_message, inputs=gr.Textbox(lines=2, placeholder="Enter your message here..."), outputs=gr.Markdown(stream=True), title="Unified Multi-Model API", description="Enter a message to get responses from a unified model.") #gradio is not suitable for production
if __name__ == "__main__":
iface.launch() |