Spaces:
Sleeping
Sleeping
Commit
·
b6fa3b6
1
Parent(s):
f4ce971
update app
Browse files
app.py
CHANGED
|
@@ -1,82 +1,91 @@
|
|
| 1 |
-
import
|
|
|
|
| 2 |
|
|
|
|
|
|
|
|
|
|
| 3 |
from transformers import AutoProcessor, LlavaForConditionalGeneration
|
| 4 |
-
from transformers import
|
| 5 |
|
| 6 |
-
from threading import Thread
|
| 7 |
-
import re
|
| 8 |
-
import time
|
| 9 |
-
from PIL import Image
|
| 10 |
-
import torch
|
| 11 |
import spaces
|
| 12 |
-
import requests
|
| 13 |
-
|
| 14 |
-
CSS ="""
|
| 15 |
-
#component-3 {
|
| 16 |
-
height: 500px !important;
|
| 17 |
-
}"""
|
| 18 |
|
| 19 |
model_id = "xtuner/llava-llama-3-8b-v1_1-transformers"
|
| 20 |
|
| 21 |
processor = AutoProcessor.from_pretrained(model_id)
|
| 22 |
|
| 23 |
model = LlavaForConditionalGeneration.from_pretrained(
|
| 24 |
-
model_id,
|
| 25 |
-
torch_dtype=torch.float16,
|
| 26 |
-
low_cpu_mem_usage=True,
|
| 27 |
)
|
| 28 |
|
| 29 |
model.to("cuda:0")
|
| 30 |
model.generation_config.eos_token_id = 128009
|
| 31 |
|
|
|
|
| 32 |
@spaces.GPU
|
| 33 |
def bot_streaming(message, history):
|
| 34 |
print(message)
|
| 35 |
if message["files"]:
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
else:
|
| 38 |
# if there's no image uploaded for this turn, look for images in the past turns
|
| 39 |
# kept inside tuples, take the last one
|
| 40 |
for hist in history:
|
| 41 |
-
|
| 42 |
-
|
| 43 |
try:
|
| 44 |
if image is None:
|
| 45 |
# Handle the case where image is None
|
| 46 |
-
gr.Error("You need to upload an image for LLaVA to work.")
|
| 47 |
except NameError:
|
| 48 |
# Handle the case where 'image' is not defined at all
|
| 49 |
-
gr.Error("You need to upload an image for LLaVA to work.")
|
| 50 |
-
|
| 51 |
-
prompt=f"<|start_header_id|>user<|end_header_id|>\n\n<image>\n{message['text']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
|
| 52 |
-
print(f"prompt: {prompt}")
|
| 53 |
image = Image.open(image)
|
| 54 |
inputs = processor(prompt, image, return_tensors='pt').to(0, torch.float16)
|
| 55 |
-
|
| 56 |
-
streamer = TextIteratorStreamer(processor, **{"skip_special_tokens": True})
|
| 57 |
-
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
|
| 58 |
-
generated_text = ""
|
| 59 |
|
| 60 |
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
| 61 |
thread.start()
|
| 62 |
-
|
| 63 |
-
text_prompt =f"<|start_header_id|>user<|end_header_id|>\n\n{message['text']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
|
| 64 |
-
print(f"text_prompt: {text_prompt}")
|
| 65 |
|
| 66 |
buffer = ""
|
|
|
|
| 67 |
for new_text in streamer:
|
| 68 |
-
|
|
|
|
|
|
|
| 69 |
buffer += new_text
|
| 70 |
-
|
| 71 |
-
generated_text_without_prompt = buffer[len(text_prompt):]
|
| 72 |
-
|
|
|
|
|
|
|
|
|
|
| 73 |
yield generated_text_without_prompt
|
| 74 |
|
| 75 |
|
| 76 |
-
demo = gr.ChatInterface(
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80 |
|
| 81 |
demo.queue(api_open=False)
|
| 82 |
-
demo.launch(show_api=False, share=False)
|
|
|
|
| 1 |
+
import time
|
| 2 |
+
from threading import Thread
|
| 3 |
|
| 4 |
+
import gradio as gr
|
| 5 |
+
import torch
|
| 6 |
+
from PIL import Image
|
| 7 |
from transformers import AutoProcessor, LlavaForConditionalGeneration
|
| 8 |
+
from transformers import TextIteratorStreamer
|
| 9 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
import spaces
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
model_id = "xtuner/llava-llama-3-8b-v1_1-transformers"
|
| 13 |
|
| 14 |
processor = AutoProcessor.from_pretrained(model_id)
|
| 15 |
|
| 16 |
model = LlavaForConditionalGeneration.from_pretrained(
|
| 17 |
+
model_id,
|
| 18 |
+
torch_dtype=torch.float16,
|
| 19 |
+
low_cpu_mem_usage=True,
|
| 20 |
)
|
| 21 |
|
| 22 |
model.to("cuda:0")
|
| 23 |
model.generation_config.eos_token_id = 128009
|
| 24 |
|
| 25 |
+
|
| 26 |
@spaces.GPU
|
| 27 |
def bot_streaming(message, history):
|
| 28 |
print(message)
|
| 29 |
if message["files"]:
|
| 30 |
+
# message["files"][-1] is a Dict or just a string
|
| 31 |
+
if type(message["files"][-1]) == dict:
|
| 32 |
+
image = message["files"][-1]["path"]
|
| 33 |
+
else:
|
| 34 |
+
image = message["files"][-1]
|
| 35 |
else:
|
| 36 |
# if there's no image uploaded for this turn, look for images in the past turns
|
| 37 |
# kept inside tuples, take the last one
|
| 38 |
for hist in history:
|
| 39 |
+
if type(hist[0]) == tuple:
|
| 40 |
+
image = hist[0][0]
|
| 41 |
try:
|
| 42 |
if image is None:
|
| 43 |
# Handle the case where image is None
|
| 44 |
+
gr.Error("You need to upload an image for LLaVA to work.")
|
| 45 |
except NameError:
|
| 46 |
# Handle the case where 'image' is not defined at all
|
| 47 |
+
gr.Error("You need to upload an image for LLaVA to work.")
|
| 48 |
+
|
| 49 |
+
prompt = f"<|start_header_id|>user<|end_header_id|>\n\n<image>\n{message['text']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
|
| 50 |
+
# print(f"prompt: {prompt}")
|
| 51 |
image = Image.open(image)
|
| 52 |
inputs = processor(prompt, image, return_tensors='pt').to(0, torch.float16)
|
| 53 |
+
|
| 54 |
+
streamer = TextIteratorStreamer(processor, **{"skip_special_tokens": False, "skip_prompt": True})
|
| 55 |
+
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024, do_sample=False)
|
|
|
|
| 56 |
|
| 57 |
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
| 58 |
thread.start()
|
| 59 |
+
|
| 60 |
+
text_prompt = f"<|start_header_id|>user<|end_header_id|>\n\n{message['text']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
|
| 61 |
+
# print(f"text_prompt: {text_prompt}")
|
| 62 |
|
| 63 |
buffer = ""
|
| 64 |
+
time.sleep(0.5)
|
| 65 |
for new_text in streamer:
|
| 66 |
+
# find <|eot_id|> and remove it from the new_text
|
| 67 |
+
if "<|eot_id|>" in new_text:
|
| 68 |
+
new_text = new_text.split("<|eot_id|>")[0]
|
| 69 |
buffer += new_text
|
| 70 |
+
|
| 71 |
+
# generated_text_without_prompt = buffer[len(text_prompt):]
|
| 72 |
+
generated_text_without_prompt = buffer
|
| 73 |
+
# print(generated_text_without_prompt)
|
| 74 |
+
time.sleep(0.06)
|
| 75 |
+
# print(f"new_text: {generated_text_without_prompt}")
|
| 76 |
yield generated_text_without_prompt
|
| 77 |
|
| 78 |
|
| 79 |
+
demo = gr.ChatInterface(
|
| 80 |
+
fn=bot_streaming,
|
| 81 |
+
fill_height=False,
|
| 82 |
+
title="LLaVA Llama-3-8B",
|
| 83 |
+
examples=[{"text": "What is on the flower?", "files": ["./bee.jpg"]},
|
| 84 |
+
{"text": "How to make this pastry?", "files": ["./baklava.png"]}],
|
| 85 |
+
description="Try [LLaVA Llama-3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). Upload an image and start chatting about it, or simply try one of the examples below. If you don't upload an image, you will receive an error.",
|
| 86 |
+
stop_btn="Stop Generation",
|
| 87 |
+
multimodal=True
|
| 88 |
+
)
|
| 89 |
|
| 90 |
demo.queue(api_open=False)
|
| 91 |
+
demo.launch(show_api=False, share=False)
|