Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -30,6 +30,7 @@ import subprocess
|
|
30 |
import pytesseract
|
31 |
from pdf2image import convert_from_path
|
32 |
import queue # ์ถ๊ฐ: queue.Empty ์์ธ ์ฒ๋ฆฌ๋ฅผ ์ํด
|
|
|
33 |
|
34 |
# -------------------- ์ถ๊ฐ: PDF to Markdown ๋ณํ ๊ด๋ จ import --------------------
|
35 |
try:
|
@@ -545,10 +546,15 @@ def clear_cuda_memory():
|
|
545 |
@spaces.GPU
|
546 |
def load_model():
|
547 |
try:
|
|
|
|
|
|
|
548 |
loaded_model = AutoModelForCausalLM.from_pretrained(
|
549 |
MODEL_ID,
|
550 |
torch_dtype=torch.bfloat16,
|
551 |
device_map="auto",
|
|
|
|
|
552 |
)
|
553 |
return loaded_model
|
554 |
except Exception as e:
|
@@ -628,19 +634,22 @@ def stream_chat(
|
|
628 |
if len(history) > max_history_length:
|
629 |
history = history[-max_history_length:]
|
630 |
|
|
|
|
|
631 |
try:
|
632 |
relevant_contexts = find_relevant_context(message)
|
633 |
-
|
634 |
-
|
635 |
-
|
636 |
-
|
637 |
-
|
638 |
-
|
639 |
-
|
|
|
640 |
except Exception as e:
|
641 |
print(f"์ปจํ
์คํธ ๊ฒ์ ์ค๋ฅ: {str(e)}")
|
642 |
-
wiki_context = ""
|
643 |
|
|
|
644 |
conversation = []
|
645 |
for prompt, answer in history:
|
646 |
conversation.extend([
|
@@ -648,43 +657,61 @@ def stream_chat(
|
|
648 |
{"role": "assistant", "content": answer}
|
649 |
])
|
650 |
|
651 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
652 |
conversation.append({"role": "user", "content": final_message})
|
653 |
|
|
|
654 |
input_ids_str = build_prompt(conversation)
|
655 |
-
|
656 |
-
|
657 |
-
|
658 |
-
inputs = tokenizer(input_ids_str, return_tensors="pt").to("cuda")
|
659 |
max_context = 8192
|
660 |
-
|
661 |
-
|
662 |
-
|
663 |
-
|
664 |
-
|
665 |
-
|
|
|
|
|
666 |
new_desired_input_length = max_context - min_generation
|
667 |
-
|
668 |
-
|
669 |
-
|
670 |
-
|
671 |
-
|
672 |
-
|
673 |
-
|
674 |
-
|
675 |
-
|
676 |
-
|
677 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
678 |
if remaining < max_new_tokens:
|
679 |
-
print(f"
|
680 |
max_new_tokens = remaining
|
681 |
|
682 |
print(f"์
๋ ฅ ํ
์ ์์ฑ ํ CUDA ๋ฉ๋ชจ๋ฆฌ: {torch.cuda.memory_allocated() / 1024**2:.2f} MB")
|
683 |
|
|
|
684 |
streamer = TextIteratorStreamer(
|
685 |
-
tokenizer, timeout=
|
686 |
)
|
687 |
|
|
|
688 |
generate_kwargs = dict(
|
689 |
**inputs,
|
690 |
streamer=streamer,
|
@@ -694,23 +721,51 @@ def stream_chat(
|
|
694 |
max_new_tokens=max_new_tokens,
|
695 |
do_sample=True,
|
696 |
temperature=temperature,
|
697 |
-
eos_token_id=
|
698 |
)
|
699 |
|
|
|
700 |
clear_cuda_memory()
|
701 |
|
|
|
702 |
thread = Thread(target=model.generate, kwargs=generate_kwargs)
|
703 |
thread.start()
|
704 |
|
|
|
705 |
buffer = ""
|
|
|
|
|
|
|
706 |
try:
|
707 |
for new_text in streamer:
|
708 |
buffer += new_text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
709 |
yield "", history + [[message, buffer]]
|
710 |
-
|
711 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
712 |
yield "", history + [[message, buffer]]
|
713 |
-
|
|
|
|
|
|
|
|
|
|
|
714 |
clear_cuda_memory()
|
715 |
|
716 |
except Exception as e:
|
@@ -825,6 +880,10 @@ def create_demo():
|
|
825 |
)
|
826 |
|
827 |
file_upload.change(
|
|
|
|
|
|
|
|
|
828 |
fn=init_msg,
|
829 |
outputs=msg,
|
830 |
queue=False
|
@@ -846,4 +905,4 @@ def create_demo():
|
|
846 |
|
847 |
if __name__ == "__main__":
|
848 |
demo = create_demo()
|
849 |
-
demo.launch()
|
|
|
30 |
import pytesseract
|
31 |
from pdf2image import convert_from_path
|
32 |
import queue # ์ถ๊ฐ: queue.Empty ์์ธ ์ฒ๋ฆฌ๋ฅผ ์ํด
|
33 |
+
import time # ์ถ๊ฐ: ์คํธ๋ฆฌ๋ฐ ํ์ด๋ฐ์ ์ํด
|
34 |
|
35 |
# -------------------- ์ถ๊ฐ: PDF to Markdown ๋ณํ ๊ด๋ จ import --------------------
|
36 |
try:
|
|
|
546 |
@spaces.GPU
|
547 |
def load_model():
|
548 |
try:
|
549 |
+
# ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ ๋จผ์ ์ํ
|
550 |
+
clear_cuda_memory()
|
551 |
+
|
552 |
loaded_model = AutoModelForCausalLM.from_pretrained(
|
553 |
MODEL_ID,
|
554 |
torch_dtype=torch.bfloat16,
|
555 |
device_map="auto",
|
556 |
+
# ๋ฎ์ ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ์ ์ํ ์ค์ ์ถ๊ฐ
|
557 |
+
low_cpu_mem_usage=True,
|
558 |
)
|
559 |
return loaded_model
|
560 |
except Exception as e:
|
|
|
634 |
if len(history) > max_history_length:
|
635 |
history = history[-max_history_length:]
|
636 |
|
637 |
+
# ์ํคํผ๋์ ์ปจํ
์คํธ ๊ฒ์
|
638 |
+
wiki_context = ""
|
639 |
try:
|
640 |
relevant_contexts = find_relevant_context(message)
|
641 |
+
if relevant_contexts: # ๊ฒฐ๊ณผ๊ฐ ์์ ๊ฒฝ์ฐ๋ง ์ถ๊ฐ
|
642 |
+
wiki_context = "\n\n๊ด๋ จ ์ํคํผ๋์ ์ ๋ณด:\n"
|
643 |
+
for ctx in relevant_contexts:
|
644 |
+
wiki_context += (
|
645 |
+
f"Q: {ctx['question']}\n"
|
646 |
+
f"A: {ctx['answer']}\n"
|
647 |
+
f"์ ์ฌ๋: {ctx['similarity']:.3f}\n\n"
|
648 |
+
)
|
649 |
except Exception as e:
|
650 |
print(f"์ปจํ
์คํธ ๊ฒ์ ์ค๋ฅ: {str(e)}")
|
|
|
651 |
|
652 |
+
# ๋ํ ๋ด์ญ ๊ตฌ์ฑ
|
653 |
conversation = []
|
654 |
for prompt, answer in history:
|
655 |
conversation.extend([
|
|
|
657 |
{"role": "assistant", "content": answer}
|
658 |
])
|
659 |
|
660 |
+
# ์ต์ข
๋ฉ์์ง ๊ตฌ์ฑ
|
661 |
+
final_message = message
|
662 |
+
if file_context:
|
663 |
+
final_message = file_context + "\nํ์ฌ ์ง๋ฌธ: " + message
|
664 |
+
if wiki_context:
|
665 |
+
final_message = wiki_context + "\nํ์ฌ ์ง๋ฌธ: " + message
|
666 |
+
if file_context and wiki_context:
|
667 |
+
final_message = file_context + wiki_context + "\nํ์ฌ ์ง๋ฌธ: " + message
|
668 |
+
|
669 |
conversation.append({"role": "user", "content": final_message})
|
670 |
|
671 |
+
# ํ๋กฌํํธ ๊ตฌ์ฑ ๋ฐ ํ ํฐํ
|
672 |
input_ids_str = build_prompt(conversation)
|
673 |
+
|
674 |
+
# ๋จผ์ ์ปจํ
์คํธ ๊ธธ์ด ํ์ธ ๋ฐ ์ ํ
|
|
|
|
|
675 |
max_context = 8192
|
676 |
+
tokenized_input = tokenizer(input_ids_str, return_tensors="pt")
|
677 |
+
input_length = tokenized_input["input_ids"].shape[1]
|
678 |
+
|
679 |
+
# ์ปจํ
์คํธ๊ฐ ๋๋ฌด ๊ธธ๋ฉด ์๋ฅด๊ธฐ
|
680 |
+
if input_length > max_context - max_new_tokens:
|
681 |
+
print(f"์
๋ ฅ์ด ๋๋ฌด ๊น๋๋ค: {input_length} ํ ํฐ. ์๋ฅด๋ ์ค...")
|
682 |
+
# ์ต์ ์์ฑ ํ ํฐ ์ ํ๋ณด
|
683 |
+
min_generation = min(256, max_new_tokens)
|
684 |
new_desired_input_length = max_context - min_generation
|
685 |
+
|
686 |
+
# ์
๋ ฅ ํ
์คํธ๋ฅผ ํ ํฐ ๋จ์๋ก ์๋ฅด๊ธฐ
|
687 |
+
tokens = tokenizer.encode(input_ids_str)
|
688 |
+
if len(tokens) > new_desired_input_length:
|
689 |
+
tokens = tokens[-new_desired_input_length:]
|
690 |
+
input_ids_str = tokenizer.decode(tokens)
|
691 |
+
|
692 |
+
# ๋ค์ ํ ํฐํ
|
693 |
+
tokenized_input = tokenizer(input_ids_str, return_tensors="pt")
|
694 |
+
input_length = tokenized_input["input_ids"].shape[1]
|
695 |
+
|
696 |
+
print(f"์ต์ข
์
๋ ฅ ๊ธธ์ด: {input_length} ํ ํฐ")
|
697 |
+
|
698 |
+
# CUDA๋ก ์
๋ ฅ ์ด๋
|
699 |
+
inputs = tokenized_input.to("cuda")
|
700 |
+
|
701 |
+
# ๋จ์ ํ ํฐ ์ ๊ณ์ฐ ๋ฐ max_new_tokens ์กฐ์
|
702 |
+
remaining = max_context - input_length
|
703 |
if remaining < max_new_tokens:
|
704 |
+
print(f"max_new_tokens ์กฐ์ : {max_new_tokens} -> {remaining}")
|
705 |
max_new_tokens = remaining
|
706 |
|
707 |
print(f"์
๋ ฅ ํ
์ ์์ฑ ํ CUDA ๋ฉ๋ชจ๋ฆฌ: {torch.cuda.memory_allocated() / 1024**2:.2f} MB")
|
708 |
|
709 |
+
# ์คํธ๋ฆฌ๋จธ ์ค์
|
710 |
streamer = TextIteratorStreamer(
|
711 |
+
tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True
|
712 |
)
|
713 |
|
714 |
+
# ์์ฑ ๋งค๊ฐ๋ณ์ ์ค์
|
715 |
generate_kwargs = dict(
|
716 |
**inputs,
|
717 |
streamer=streamer,
|
|
|
721 |
max_new_tokens=max_new_tokens,
|
722 |
do_sample=True,
|
723 |
temperature=temperature,
|
724 |
+
eos_token_id=tokenizer.eos_token_id, # ๋ช
์์ EOS ํ ํฐ ์ง์
|
725 |
)
|
726 |
|
727 |
+
# ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ
|
728 |
clear_cuda_memory()
|
729 |
|
730 |
+
# ๋ณ๋ ์ค๋ ๋์์ ์์ฑ ์คํ
|
731 |
thread = Thread(target=model.generate, kwargs=generate_kwargs)
|
732 |
thread.start()
|
733 |
|
734 |
+
# ์๋ต ์คํธ๋ฆฌ๋ฐ
|
735 |
buffer = ""
|
736 |
+
partial_message = ""
|
737 |
+
last_yield_time = time.time()
|
738 |
+
|
739 |
try:
|
740 |
for new_text in streamer:
|
741 |
buffer += new_text
|
742 |
+
partial_message += new_text
|
743 |
+
|
744 |
+
# ์ผ์ ์๊ฐ๋ง๋ค ๋๋ ํ
์คํธ๊ฐ ์์ผ ๋๋ง๋ค ๊ฒฐ๊ณผ ์
๋ฐ์ดํธ
|
745 |
+
current_time = time.time()
|
746 |
+
if current_time - last_yield_time > 0.1 or len(partial_message) > 20:
|
747 |
+
yield "", history + [[message, buffer]]
|
748 |
+
partial_message = ""
|
749 |
+
last_yield_time = current_time
|
750 |
+
|
751 |
+
# ๋ง์ง๋ง ์๋ต ํ์ธ
|
752 |
+
if buffer:
|
753 |
yield "", history + [[message, buffer]]
|
754 |
+
|
755 |
+
# ๋ํ ๊ธฐ๋ก์ ์ ์ฅ
|
756 |
+
chat_history.add_conversation(message, buffer)
|
757 |
+
|
758 |
+
except Exception as e:
|
759 |
+
print(f"์คํธ๋ฆฌ๋ฐ ์ค ์ค๋ฅ ๋ฐ์: {str(e)}")
|
760 |
+
if not buffer: # ๋ฒํผ๊ฐ ๋น์ด์์ผ๋ฉด ์ค๋ฅ ๋ฉ์์ง ํ์
|
761 |
+
buffer = f"์๋ต ์์ฑ ์ค ์ค๋ฅ๊ฐ ๋ฐ์ํ์ต๋๋ค: {str(e)}"
|
762 |
yield "", history + [[message, buffer]]
|
763 |
+
|
764 |
+
# ์ค๋ ๋๊ฐ ์ฌ์ ํ ์คํ ์ค์ด๋ฉด ์ข
๋ฃ ๋๊ธฐ
|
765 |
+
if thread.is_alive():
|
766 |
+
thread.join(timeout=5.0)
|
767 |
+
|
768 |
+
# ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ
|
769 |
clear_cuda_memory()
|
770 |
|
771 |
except Exception as e:
|
|
|
880 |
)
|
881 |
|
882 |
file_upload.change(
|
883 |
+
fn=lambda: ("์ฒ๋ฆฌ ์ค...", [["์์คํ
", "ํ์ผ์ ๋ถ์ ์ค์
๋๋ค. ์ ์๋ง ๊ธฐ๋ค๋ ค์ฃผ์ธ์..."]]),
|
884 |
+
outputs=[msg, chatbot],
|
885 |
+
queue=False
|
886 |
+
).then(
|
887 |
fn=init_msg,
|
888 |
outputs=msg,
|
889 |
queue=False
|
|
|
905 |
|
906 |
if __name__ == "__main__":
|
907 |
demo = create_demo()
|
908 |
+
demo.launch()
|