Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import annotations
|
2 |
+
|
3 |
+
import spaces
|
4 |
+
|
5 |
+
import gradio as gr
|
6 |
+
from threading import Thread
|
7 |
+
from transformers import TextIteratorStreamer
|
8 |
+
import hashlib
|
9 |
+
import os
|
10 |
+
|
11 |
+
from transformers import AutoModel, AutoProcessor
|
12 |
+
import torch
|
13 |
+
import sys
|
14 |
+
import subprocess
|
15 |
+
from PIL import Image
|
16 |
+
|
17 |
+
from cobra import load
|
18 |
+
import time
|
19 |
+
|
20 |
+
subprocess.check_call([sys.executable, '-m', 'pip', 'install', 'mamba-ssm'])
|
21 |
+
subprocess.check_call([sys.executable, '-m', 'pip', 'install', 'causal-conv1d'])
|
22 |
+
|
23 |
+
vlm = load("cobra+3b")
|
24 |
+
|
25 |
+
if torch.cuda.is_available():
|
26 |
+
DEVICE = "cuda"
|
27 |
+
DTYPE = torch.bfloat16
|
28 |
+
else:
|
29 |
+
DEVICE = "cpu"
|
30 |
+
DTYPE = torch.float32
|
31 |
+
vlm.to(DEVICE, dtype=DTYPE)
|
32 |
+
|
33 |
+
prompt_builder = vlm.get_prompt_builder()
|
34 |
+
system_prompt = prompt_builder.system_prompt
|
35 |
+
|
36 |
+
@spaces.GPU
|
37 |
+
def bot_streaming(message, history):
|
38 |
+
print(message)
|
39 |
+
if message["files"]:
|
40 |
+
image = message["files"][-1]["path"]
|
41 |
+
else:
|
42 |
+
# if there's no image uploaded for this turn, look for images in the past turns
|
43 |
+
# kept inside tuples, take the last one
|
44 |
+
for hist in history:
|
45 |
+
if type(hist[0])==tuple:
|
46 |
+
image = hist[0][0]
|
47 |
+
|
48 |
+
image = Image.open(image).convert("RGB")
|
49 |
+
|
50 |
+
prompt_builder.add_turn(role="human", message=message)
|
51 |
+
prompt_text = prompt_builder.get_prompt()
|
52 |
+
|
53 |
+
# Generate from the VLM
|
54 |
+
generated_text = vlm.generate(
|
55 |
+
image,
|
56 |
+
prompt_text,
|
57 |
+
cg=True,
|
58 |
+
do_sample=False,
|
59 |
+
temperature=1.0,
|
60 |
+
max_new_tokens=2048,
|
61 |
+
# do_sample=cfg.do_sample,
|
62 |
+
# temperature=cfg.temperature,
|
63 |
+
# max_new_tokens=cfg.max_new_tokens,
|
64 |
+
)
|
65 |
+
prompt_builder.add_turn(role="gpt", message=generated_text)
|
66 |
+
|
67 |
+
# streamer = TextIteratorStreamer(processor, **{"skip_special_tokens": True})
|
68 |
+
# generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=100)
|
69 |
+
# generation_kwargs = dict(image, prompt_text, cg=True, do_sample=cfg.do_sample, temperature=cfg.temperature, max_new_tokens=cfg.max_new_tokens)
|
70 |
+
generation_kwargs = dict(image, prompt_text, cg=True, do_sample=True, temperature=1.0, max_new_tokens=2048)
|
71 |
+
|
72 |
+
thread = Thread(target=vlm.generate, kwargs=generation_kwargs)
|
73 |
+
thread.start()
|
74 |
+
|
75 |
+
text_prompt =f"[INST] \n{message['text']} [/INST]"
|
76 |
+
print(generated_text)
|
77 |
+
|
78 |
+
|
79 |
+
buffer = ""
|
80 |
+
yield generated_text
|
81 |
+
# for new_text in streamer:
|
82 |
+
|
83 |
+
# buffer += new_text
|
84 |
+
|
85 |
+
# generated_text_without_prompt = buffer[len(text_prompt):]
|
86 |
+
# time.sleep(0.04)
|
87 |
+
# yield generated_text_without_prompt
|
88 |
+
|
89 |
+
|
90 |
+
demo = gr.ChatInterface(fn=bot_streaming, title="LLaVA Next", examples=[{"text": "What is on the flower?", "files":["./bee.jpg"]},
|
91 |
+
{"text": "How to make this pastry?", "files":["./baklava.png"]}],
|
92 |
+
description="Try [LLaVA Next](https://huggingface.co/papers/2310.03744) in this demo. Upload an image and start chatting about it, or simply try one of the examples below.",
|
93 |
+
stop_btn="Stop Generation", multimodal=True)
|
94 |
+
demo.launch(debug=True)
|