fixing conversation formatting
Browse files
app.py
CHANGED
@@ -10,54 +10,82 @@ import requests
|
|
10 |
model_path = "facebook/chameleon-7b"
|
11 |
# model = ChameleonForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, device_map="auto")
|
12 |
# processor = ChameleonProcessor.from_pretrained(model_path)
|
13 |
-
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, device_map="auto"
|
14 |
model.eval()
|
15 |
-
processor = ChameleonProcessor.from_pretrained(model_path
|
16 |
tokenizer = processor.tokenizer
|
17 |
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
|
24 |
@spaces.GPU(duration=90)
|
25 |
def respond(
|
26 |
message,
|
27 |
-
history:
|
28 |
system_message,
|
29 |
max_tokens,
|
30 |
temperature,
|
31 |
top_p,
|
32 |
):
|
33 |
-
# messages = [{"role": "system", "content": system_message}]
|
34 |
-
|
35 |
-
# for val in history:
|
36 |
-
# if val[0]:
|
37 |
-
# messages.append({"role": "user", "content": val[0]})
|
38 |
-
# if val[1]:
|
39 |
-
# messages.append({"role": "assistant", "content": val[1]})
|
40 |
-
|
41 |
-
# messages.append({"role": "user", "content": message})
|
42 |
-
|
43 |
response = ""
|
44 |
|
45 |
-
|
46 |
-
|
|
|
|
|
|
|
|
|
|
|
47 |
|
48 |
-
inputs = processor(prompt, images=
|
49 |
|
50 |
streamer = TextIteratorStreamer(tokenizer)
|
51 |
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=20)
|
52 |
|
53 |
-
|
54 |
-
|
55 |
-
|
|
|
56 |
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
|
|
|
|
61 |
|
62 |
|
63 |
"""
|
@@ -65,6 +93,7 @@ For information on how to customize the ChatInterface, peruse the gradio docs: h
|
|
65 |
"""
|
66 |
demo = gr.ChatInterface(
|
67 |
respond,
|
|
|
68 |
additional_inputs=[
|
69 |
gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
|
70 |
gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
|
@@ -81,4 +110,4 @@ demo = gr.ChatInterface(
|
|
81 |
|
82 |
|
83 |
if __name__ == "__main__":
|
84 |
-
demo.launch()
|
|
|
10 |
model_path = "facebook/chameleon-7b"
|
11 |
# model = ChameleonForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, device_map="auto")
|
12 |
# processor = ChameleonProcessor.from_pretrained(model_path)
|
13 |
+
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, device_map="auto")
|
14 |
model.eval()
|
15 |
+
processor = ChameleonProcessor.from_pretrained(model_path)
|
16 |
tokenizer = processor.tokenizer
|
17 |
|
18 |
+
multimodal_file = tuple[str, str]
|
19 |
+
multimodal_message = list[str | multimodal_file] | multimodal_file
|
20 |
+
# todo: verify this type with gr.ChatInterface
|
21 |
+
message_t = str | multimodal_message
|
22 |
+
history_t = list[tuple[str, str] | list[tuple[multimodal_message, multimodal_message]]]
|
23 |
+
|
24 |
+
def history_to_prompt(
|
25 |
+
message,
|
26 |
+
history: history_t,
|
27 |
+
eot_id = "<reserved08706>",
|
28 |
+
image_placeholder = "<image>"
|
29 |
+
):
|
30 |
+
|
31 |
+
prompt = ""
|
32 |
+
images = []
|
33 |
+
for turn in history + (message, None):
|
34 |
+
print("turn:", turn)
|
35 |
+
# turn should be a tuple of user message and assistant message
|
36 |
+
for message in turn:
|
37 |
+
if isinstance(message, str):
|
38 |
+
prompt += user_message
|
39 |
+
prompt += eot_id
|
40 |
+
if isinstance(message, list):
|
41 |
+
for item in message:
|
42 |
+
if isinstance(item, str):
|
43 |
+
prompt += item
|
44 |
+
elif isinstance(item, tuple):
|
45 |
+
image_path, alt = item
|
46 |
+
prompt += image_placeholder
|
47 |
+
image = Image.open(requests.get(image_path, stream=True).raw)
|
48 |
+
images.append(image)
|
49 |
+
else:
|
50 |
+
prompt += f"(unhandled message type: {message})"
|
51 |
+
prompt += eot_id
|
52 |
+
return prompt, images
|
53 |
|
54 |
@spaces.GPU(duration=90)
|
55 |
def respond(
|
56 |
message,
|
57 |
+
history: history_t,
|
58 |
system_message,
|
59 |
max_tokens,
|
60 |
temperature,
|
61 |
top_p,
|
62 |
):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
response = ""
|
64 |
|
65 |
+
print(f"message: {message}\nhistory:\n\n{history}\n")
|
66 |
+
prompt, images = history_to_prompt(message, history)
|
67 |
+
print(f"prompt:\n\n{prompt}\n")
|
68 |
+
|
69 |
+
# prompt = "I'm very intrigued by this work of art:<image>Please tell me about the artist."
|
70 |
+
# image = Image.open(requests.get("https://uploads4.wikiart.org/images/paul-klee/death-for-the-idea-1915.jpg!Large.jpg", stream=True).raw)
|
71 |
+
# images = [image]
|
72 |
|
73 |
+
inputs = processor(prompt, images=images, return_tensors="pt").to(model.device, dtype=torch.bfloat16)
|
74 |
|
75 |
streamer = TextIteratorStreamer(tokenizer)
|
76 |
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=20)
|
77 |
|
78 |
+
try:
|
79 |
+
# launch generation in the background
|
80 |
+
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
81 |
+
thread.start()
|
82 |
|
83 |
+
partial_message = ""
|
84 |
+
for new_token in streamer:
|
85 |
+
partial_message += new_token
|
86 |
+
yield partial_message
|
87 |
+
except e:
|
88 |
+
return f"Error: {e}"
|
89 |
|
90 |
|
91 |
"""
|
|
|
93 |
"""
|
94 |
demo = gr.ChatInterface(
|
95 |
respond,
|
96 |
+
multimodal=True,
|
97 |
additional_inputs=[
|
98 |
gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
|
99 |
gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
|
|
|
110 |
|
111 |
|
112 |
if __name__ == "__main__":
|
113 |
+
demo.launch(debug=True)
|